Research brief · AI agents · October 2026 · 10 dimensions · 54 questions · 4 AI models + 30 checked sources

Are CEOs ready
for AI agents?

Are the CEOs of mid-sized and large companies informed about AI agents, and ready to put them into daily operations? We gave the same 54 questions to four AI models, then checked their answers against 30 surveys and official statistics. The short answer: most CEOs know the word and say they are adopting, but only about one large European company in five, and one mid-sized company in fifteen, runs an agent in daily work.

By Christos Papadimitriou · House of 47 Research
Last updated:

Read the key findings
18–22%Large European firms with an agent in daily use, the four models' estimate
6–8%Mid-sized European firms (50–249 staff), same estimate
56%CEOs who see no significant financial benefit from AI yet (PwC, Jan 2026)
4 modelsSame questionnaire to each. Three are Grok-family, so diversity is limited
A modern office corridor lined with glass-walled rooms
Key findings · four AI models compared with the published data

Key findings

Are CEOs informed and ready? Informed about the word, yes. Ready to run agents in daily operations, mostly not yet. Here is what the four models said, and where the surveys confirm or contradict them.

  • Aware
    CEOs know the word, not the work

    The models think 25–45% of large-company CEOs could explain how an agent differs from a chatbot or RPA (Grok 4.7 is strictest at 28% in Europe, Composer 2.5 most generous at 35–45%). The surveys fit: Capgemini found only half of executives say they understand what agents can do, and only 22% of Greek businesses had even heard of agentic AI (AWS/Strand, Oct 2026).

  • Say ≠ do
    A large gap between saying and doing

    61% of CEOs told IBM they are actively adopting agents, and 79% of US executives told PwC that agents are being adopted. Yet only 2% of $1B+ firms had fully scaled them (Capgemini, 2025), and 23% use agentic AI at least moderately (Deloitte, 2026). All four models describe the same gap, putting it at 15 to 75 points depending on the model.

  • Daily use
    About one large firm in five, one mid-sized in fifteen

    Three Grok models give 18% for large European firms and 6–7% for mid-sized; Composer 2.5 gives 22% and 8%. No official series counts agents. Bitkom's 11% of AI-engaged German firms using agents fits these numbers. McKinsey's 40% of $1B+ firms 'scaling' agents and KPMG's 62% 'building, deploying or developing' show the very largest firms are further ahead than the models think.

  • ROI
    Returns are real for a minority

    The models say 15–20% of CEOs can show a measurable return. PwC's 2026 CEO survey falls on either side of that: 12% saw both cost and revenue gains from AI, 33% saw either, 56% saw none. BCG adds that 94% will keep investing anyway.

  • Barrier
    Skills stop firms starting; integration stops them finishing

    Grok 4.7, Grok 4.6 and Composer 2.5 rank legacy integration as the top barrier, and Grok 4.5 ranks skills first. Every official survey puts skills first: 70.9% of EU firms that considered AI cite lack of expertise (Eurostat), and so do 85% of German non-users (Bitkom). The two sides answer different questions: why firms do not start, and why projects stall.

  • Ownership
    CEOs now own the decision, which the models miss

    All four models say the CIO owns agent decisions in large firms, with the CEO as owner in only 10–18%. BCG found that 72% of CEOs now call themselves the main AI decision maker, twice last year's share. In France, the dirigeant drives AI projects in 73% of PME/ETI (Bpifrance).

  • Rules
    The AI Act clock moved; two models and the guides did not notice

    Grok 4.7 and Grok 4.6 put the Annex III high-risk duties on 2 August 2026. The AI Omnibus (in force 27 July 2026) moved them to 2 December 2027. Transparency duties did start on 2 August 2026. Our sister company theagency47's four country guides also miss the shift.

  • theagency47
    Models back the method, not every claim

    All four models agree with theagency47 on starting at the Operational or Task level (R1) and on its country framings (R8). On five other positions all four only partly agree. Grok 4.5 disagrees that bank assistants scale down to a 40–300 person firm (R5); the other three only partly agree.

The arc · 2025 → 2033 · from saying to doing

From talk to daily use

The surveys and the four models trace the same curve: CEOs said yes in 2025, pilots spread in 2026, agents arrive inside the software companies already rent, and most firms have one in daily use around the turn of the decade.

2025 · The talk

CEOs say yes

61%

IBM's 2025 CEO study: 61% say they are actively adopting AI agents. The same summer, Capgemini found 2% of $1B+ firms had fully scaled them.

Oct 2026 · Now

Many pilots, few in daily use

18–22%

The four models' estimate for large European firms with one agent in daily use (6–8% of mid-sized). Bitkom: 11% of AI-engaged German firms already use agents.

2026 → 2028 · Embedded

Agents arrive in existing software

40%

Gartner: 40% of enterprise apps will include task-specific agents by end-2026, up from under 5%. All four models expect most new use to come this way.

2028 → 2033 · Majority

Most firms run one agent

2029 / 2031

Grok models: a majority of large European firms in 2029 and of mid-sized firms in 2031. Composer 2.5: 2030 and 2033. Large US firms: 2028 in all four.

Cross-check · Grok 4.7, Grok 4.6, Grok 4.5, Composer 2.5, published evidence

Where they agree

  • All 4
    Broad but shallow awareness

    Every model says most CEOs know the phrase and few can say what an agent may and may not do. None knew of a survey that tests the definition, so all tag their numbers as estimates.

  • All 4
    Most 'adoption' is embedded

    50–80% of what companies report as agent adoption is a feature switched on in Microsoft, Salesforce, SAP or ServiceNow, not an agent the company designed for a process.

  • All 4
    Start narrow, keep a human on the button

    One frequent process at the Task or Operational level. The agent prepares the work, and a person approves anything that sends, pays or changes a record. Keep it or kill it on the numbers after about 90 days.

  • All 4
    No multi-agent systems for mid-sized firms yet

    One or two narrow agents is the ceiling for 2026–27. Orchestrating several agents adds failure modes these firms cannot staff.

  • All 4
    Same country order

    US, Germany, France, Italy, Greece, on both awareness and adoption. Germany moves fastest of the four European countries to 2028. Regulation explains only 20–35% of Europe's gap with the US.

  • All 4
    Budgets keep growing anyway

    Spending on agents rises in 2027 even if returns stay unclear, though more slowly. All four put the chance of an incident-led European slowdown before end-2027 at exactly 35%.

Where they disagree

Large US firms with an agent in daily use

T3 · % of large US companies, October 2026

0%70%

Cost of a first production agent

T7 · mid-sized firm, build plus first year, EUR thousands

€0€200k

Mid-sized firms that are agent-ready

T14 · % of mid-sized European firms meeting most criteria

0%25%

When most mid-sized European firms use agents daily

T9 · year when more than 50% have one agent in daily use

20272036

Point estimates as written in each model's summary table, 17 Oct 2026. Boxed pins are checked references, not model answers: McKinsey (Aug 2026, $1B+ firms 'scaling' agents, global), KPMG (Q3 2026, US $1B+ firms 'building, deploying or developing'), Capgemini (2025, $1B+ firms with high data readiness) and theagency47's published entry price (a group company).

Probability of outcome · House of 47 synthesis

The odds

A majority of large European firms run at least one agent daily~55%by 2029
A majority of mid-sized European firms do~45%by 2031
Agent budgets at large firms are still growing~85%by end-2027
A visible agent incident slows European adoption~35%by end-2027

Our scenario estimates, informed by the four models' answers and the checked sources below (for example BCG: 94% of CEOs will keep investing even without payoff). Not market forecasts and not any model's answer.

Dimension by dimension · 10 dimensions · 54 questions · 4 models

Dimension by dimension

For each dimension: the key numbers each model gave, where they agree, where they split (named), how that compares with the real survey data, and our synthesis. Full answers are in the raw files listed in the methodology.

DIMENSION 1

Awareness vs hype

ModelLarge-company CEOs who can explain an agent (EU / US)Mid-sized (EU / US)Say–do gap
Grok 4.728% / 40%12% / 18%30–45 points (large firms)
Grok 4.632% / 40%16% / 22%2.5–3× in Europe, 2× in US
Grok 4.525–35% / 35–45%12–22% / 20–30%60–75 points
Composer 2.535–45% / 45–55%20–30% / 25–35%15–25 points (large EU)
Common ground

All four say awareness is broad but shallow. Nearly every CEO has heard 'agent', because vendors have spent two years putting the label on every copilot. Far fewer can say why posting an invoice is a different control problem from a chat window. Enthusiasm comes mostly from vendors, boards and fear of missing out. Evidence from the company's own operations accounts for only 15–45% of it. The typical error is to overrate autonomy and underrate the plumbing: integration, permissions and the person who still has to approve.

Disagreements

Composer 2.5 is the most generous on understanding. Grok 4.7 is the strictest, at 12% of European mid-sized CEOs, and it adds that a peer who has run an agent for six months is the most reliable source of information and the rarest. Their estimates of the say–do gap differ most: Grok 4.5 thinks only 25–40% of CEOs who say they use agents have one in production, while Composer 2.5 puts the gap at 15–25 points.

Reality check · the data

No survey tests the definition, but self-reports fit the models. Only half of executives at $1B+ firms say they understand what agents can do (Capgemini, Jul 2025), and only 22% of Greek businesses had heard of agentic AI (AWS/Strand, Oct 2026). The models' FOMO share is confirmed: 64% of CEOs admit fear of falling behind drives investment before value is clear (IBM, 2025), and 33% of German AI users adopt mainly for that reason (Bitkom, 2026). Gartner adds that only about 130 of the thousands of 'agentic AI vendors' are real; the rest is 'agent washing'.

Synthesis

CEOs are well informed about the word and poorly informed about the work. A quick test: can the CEO say what the agent is not allowed to do, and who approves its actions? If not, the company is at the awareness stage, whatever the slide says.

Further reading: Definitions: What is an AI agent? and AI agent vs chatbot (theagency47). On agent washing and who actually delivers: Who sells AI agents that stick.

DIMENSION 2

Actual adoption vs pilots

ModelLarge EU in daily useMid-sized EULarge USPilots that reach production
Grok 4.718%6%28%30%
Grok 4.618%7%28%30%
Grok 4.518%7%28%25%
Composer 2.522%8%35%25%
Common ground

The three Grok models give almost the same numbers, and Composer 2.5 is a few points higher. All four say that 'nothing with agents' does not mean nothing with AI: staff paste text into ChatGPT beside the work, and that is not a programme. A quarter to a third of pilots reach daily production. They stall for ordinary reasons: no owner, no access to the system of record, no baseline, and exceptions that take as long as the old job. Model quality is rarely the cause. Most growth to end-2027 comes from agents built into software firms already use.

Disagreements

Composer 2.5 is the most bullish on levels (22% / 8% / 35%) and on 2027 (large US 45–55%). Grok 4.7 is the most pessimistic on pilots, saying about half are abandoned. On European examples they split. Grok 4.7 named Klarna, Vodafone's TOBi and Siemens. Grok 4.6 named Siemens and Carrefour's Hopla and would not vouch for the bank assistants. Grok 4.5 and Composer 2.5 put Commerzbank's Ava forward with medium confidence. We did not check any of these examples ourselves.

Reality check · the data

No official series counts agents in daily use. Eurostat's 2025 figure for any AI use (55.0% of large and 30.4% of medium EU firms) is an upper limit. Agent-specific surveys bracket the models. At the low end, 1% of very large Italian firms had agents in their processes in 2025 (Capgemini via Il Sole 24 Ore), and 11% of AI-engaged German firms use agents (Bitkom, Sep 2026). At the high end, 23% of companies use agentic AI at least moderately (Deloitte, 2026), and 40% of $1B+ firms say they are scaling agents, up from 27% (McKinsey, Aug 2026). On pilots, only 25% of firms have moved 40% or more of their AI pilots into production (Deloitte), and Gartner expects over 40% of agentic projects to be cancelled by end-2027.

Synthesis

For Europe the models' numbers sit inside the range of the surveys. For the biggest US firms they look low, because consultancy surveys count 'scaling' and 'building', a looser bar than daily use by ordinary staff. Any survey that reports more than 60% adoption is counting intent.

Further reading: Why pilots stall: Why AI agent projects fail (theagency47).

DIMENSION 3

Budgets and ROI

ModelFirst agent, build + year 1PaybackCEOs with a measurable return
Grok 4.7€150k (60–350k)~15 months (9–24)15% of large-company CEOs
Grok 4.6€120k (60–220k)12–24 months20% of those that deployed
Grok 4.5€80k (40–120k)6–18 months20% of those that tried; ~8% of all
Composer 2.5€120k (80–200k)18–36 months20% (15–25%)
Common ground

All four say a first production agent costs a mid-sized firm high five figures to low six figures in its first year, once integration and internal time are counted. The lines companies under-budget are integration and identity, data and written procedures, oversight staff and evaluation. For one narrow agent, tokens are rarely the main cost. Budgets keep growing in 2027 even if returns stay muddy, because much of the money sits inside Microsoft, SAP and Salesforce renewals and boards read a retreat as failure.

Disagreements

Grok 4.5 is the cheapest and fastest (€80k, 6–18 months). Grok 4.7 is the dearest because it counts €25–70k of internal staff time, and it guesses that only about half of first agents pay back within two years. Composer 2.5 has the slowest payback and adds a CFO test: programmes shrink where an agent cannot beat €3–8 per handled case. For large firms, agent spend ranges from 0.01–0.05% of revenue (Grok 4.5) to 0.05–0.2% (Composer 2.5).

Reality check · the data

Returns check out. In PwC's 2026 CEO survey, 56% saw no significant financial benefit from AI, 33% saw gains in cost or revenue, and 12% saw both. IBM found that 25% of AI initiatives delivered the expected ROI. Those figures cover AI in general, so agent-specific returns are probably lower. Spending looks underestimated. BCG says AI spend doubles to about 1.7% of revenue in 2026, with over 30% of it committed to agentic AI. By our arithmetic that is about 0.5% of revenue on agents at the large firms BCG surveyed, above every model's range. The hidden costs are confirmed: German firms name infrastructure (51%), data preparation (50%) and integration (41%) as major cost items, and licences only 21% (Bitkom). About 20% say AI operating costs, including tokens, constrain their use (McKinsey 2026).

Synthesis

Budget for integration, data and oversight, not for the licence. Set the keep-or-kill number before you start. A caveat from our own side: theagency47's fixed-scope packages start at €2,500 for one agent in 14 days, far below every model's estimate. The models priced custom integrator projects plus internal staff time. A narrow agent on existing systems can cost much less, but the internal time is real either way.

Further reading: Tools: ROI calculator and How much does an AI agent cost? (theagency47). Why model costs matter less each year: The collapsing cost of intelligence.

DIMENSION 4

Barriers: skills, data, governance, regulation

ModelTop barrier (weight)Mid-sized firms with agent-ready dataMost overrated / most underrated
Grok 4.7Integration (20)10%Legal uncertainty / integration plus ownership
Grok 4.6Integration (22)10%The AI Act as a reason to wait / integration plus no written process
Grok 4.5Skills, translators (22)8–15%'AI isn't good enough' / undocumented processes
Composer 2.5Integration (20)15–25%Employee resistance / integration debt, accountability
Common ground

The scarcest person is the 'translator', someone who can turn a process into steps, rules and a test set, usually a strong operations manager rather than a data scientist. Integration and messy data are two sides of one problem. Legal uncertainty is overrated for an ordinary back-office agent. GDPR, works councils and the CSE add weeks or months, not years, if they are brought in early. Only about one mid-sized firm in ten has data good enough for an agent to act on.

Disagreements

Grok 4.5 alone ranks skills above integration. Composer 2.5 is twice as optimistic as the others on data readiness. Estimates of the works-council delay range from 1–4 months (Grok 4.5) to 2–6 months (Grok 4.7, Composer 2.5); Grok 4.6 gives 4–10 weeks for an invoice agent and 3–9 months for anything touching HR. On the AI Act, Grok 4.7 and Grok 4.6 date the Annex III high-risk rules to 2 August 2026, Grok 4.5 would not give dates, and Composer 2.5 stayed vague ('2026–2027').

Reality check · the data

Every survey of firms that considered AI puts skills first. Eurostat 2025: lack of expertise 70.9%, legal uncertainty 52.5%, data protection 48.8%. Bitkom: 85% of German non-users lack know-how. Istat: 58.6% of Italian firms cite skills, and Insee says half of French firms do. Data is the next gap: fewer than one in five $1B+ firms report high data readiness (Capgemini), and 80% of US banks name data readiness or access as their top challenge for agents (KPMG). On rules, the AI Omnibus (in force 27 Jul 2026) moved Annex III high-risk duties to 2 December 2027 and Annex I to 2 August 2028. The transparency rules (Article 50) have applied and been enforced since 2 August 2026.

Synthesis

Both camps are right about different moments: skills stop firms starting, and integration stops them finishing. For an ordinary agent the legal homework is to tell people they are dealing with AI and to do the GDPR work. Hiring, worker scoring and credit are high-risk from December 2027.

Further reading: Security and controls: AI agent security and Trust (theagency47).

DIMENSION 5

Mid-sized vs large companies

ModelLarge vs mid-sized gapMulti-agent for mid-sized firmsThe gap by 2028
Grok 4.7~3× (18% vs 6%)Not before end-2027Adoption narrows, value gap stays
Grok 4.6~2.5× (18% vs 7%)Under 5% benefit by 2027Narrows on access, widens on sophistication
Grok 4.510–15 pointsMostly unrealisticAccess narrows, execution widens
Composer 2.52–3×Usually prematureAccess narrows, custom gap widens
Common ground

Large firms are two to three times further along. Mid-sized firms have real advantages: an owner who can decide in a week, fewer systems, and a problem the CEO can see. Their handicaps are no data team, a budget in which a miss is visible, little vendor attention, and know-how held by a few people. Building in-house rarely makes sense below 250 staff. Agents arriving inside existing software will narrow the gap in access by 2028, but not the gap in results.

Disagreements

On how to get the first agent, the three Grok models default to an agent embedded in existing software, adding an integrator only when the work crosses systems. Composer 2.5 calls an agency or integrator 'the best default' for a mid-sized firm with messy integrations, provided it hands over runbooks and logging. Composer 2.5 alone says private-equity-owned mid-caps can move faster than listed large firms, and that an owner-CEO may kill the programme after one visible mistake.

Reality check · the data

Official data confirm the size gap. Eurostat 2025: 55.0% of large, 30.4% of medium and 17.0% of small EU firms use AI. In Italy, 71% of large firms have started an AI project against 8% of SMEs (Politecnico di Milano). The gap is growing at the top: McKinsey's share of $1B+ firms scaling agents rose from 27% to 40% in a year, while smaller organisations stayed at 22%. Multi-agent systems are arriving faster at the top than the models expect: 25% of US $1B+ firms now use them, up from 6% (KPMG, Q3 2026). That says nothing about mid-sized firms.

Synthesis

Mid-sized firms should start with one narrow agent, embedded where the work already lives, and bring in an integrator only when it crosses two or three systems. Large firms should expect multi-agent pilots, which the models undercount.

Further reading: Why multi-agent systems fail in smaller firms: theagency47 analysis.

DIMENSION 6

Country differences: Greece, Italy, France, Germany vs US

ModelGermany large / midFranceItalyGreece
Grok 4.722% / 8%17% / 6%13% / 5%9% / 3%
Grok 4.622% / 9%18% / 7%15% / 5%10% / 3%
Grok 4.522% / 9%18% / 7%15% / 5%10% / 3%
Composer 2.514% blended12% blended9% blended7% blended
Common ground

All four rank the US first, then Germany, France, Italy and Greece, and all say Germany will move fastest of the four European countries to 2028, pulled by its labour shortage and ERP discipline. Greece is last because of its size structure: micro firms dominate and the 50–249 band is thin, with the systemic banks as the exception. Regulation explains a fifth to a third of Europe's gap with the US. The rest comes from where the software vendors sit, capital markets, language and implementation talent.

Disagreements

Grok 4.6 and Grok 4.5 gave identical country tables, a sign of shared training rather than independent judgement. Grok 4.7 alone warns that Germany and France 'are the pair I would soonest swap'. Composer 2.5 gave only blended figures, and notes that Spain, the Nordics and Benelux are missing from the guides. Estimates of regulation's share of the US gap range from about 20% (Grok 4.7) to 25–35% (Grok 4.6).

Reality check · the data

Eurostat's 2025 figures for any AI use confirm the order for all firms: Germany 26.0%, France 18.2%, Italy 16.4%, Greece 8.9%. Among large firms, France (58.0%) is just ahead of Germany (57.0%), so Grok 4.7's warning was right. No model knew that Greece went backwards, from 9.8% to 8.9% overall and from 15.9% to 15.2% among mid-sized firms, even as its large firms rose from 24.3% to 37.6%. Nor did they know that European CEOs are more confident than American ones that AI will pay off (61% vs 52%, BCG).

Synthesis

The order holds, but the gaps within countries are wider than the gaps between them. See the country section below.

DIMENSION 7

Which functions and levels agents enter first

ModelFirst functionFirst levelBack-office work agents handle, 2026 → 2028
Grok 4.7Customer service (mid); IT and customer service (large)Task, or a narrow Operational slice15% → 30%
Grok 4.6Customer service (L1)Task or Operational20% → 38%
Grok 4.5Customer serviceTask or Operational15–25% → 25–40%
Composer 2.5Customer service / ITTask or Operational15–25% → 25–40%
Common ground

Agents start in customer service and the IT service desk, followed by accounts payable and sales operations, because those functions have volume, a digital trail and a known KPI. Hiring and legal work look impressive and make poor first projects. The first agent belongs at the Task or Operational level. An executive briefing agent fails as a first project because the questions are ambiguous and the data is not yet trusted. Agents can reliably handle 15–25% of a back-office employee's tasks today, rising to 25–40% by 2028. That means tasks, not jobs.

Disagreements

Grok 4.6 is the most optimistic on task share (20% now, 38% by 2028). Grok 4.7 ranks IT first for large firms and the shared inbox and payables first for mid-sized firms. Composer 2.5 places sales enablement ahead of accounts payable. Grok 4.5 makes one exception to 'no executive agent first': a meeting-prep copilot that cites approved sources.

Reality check · the data

Capgemini's first functions are customer service, IT and sales, which matches the models. McKinsey 2026 finds agents scaled most in IT, knowledge management and software engineering. Among EU firms that use AI, the most common uses are marketing and sales (34.7%) and business administration (31.1%) (Eurostat). Only 7% of German firms use AI in the management function (Bitkom), which supports the view that executive agents come later.

Synthesis

This confirms theagency47's three-level model: Executive, Operational and Task are layers, not stages, and the first agent belongs in the bottom two.

Further reading: Frameworks: AI organisational levels, the three-tier AI workforce model and payback by function (theagency47). Customer-facing agents and the web: The End of the Page.

DIMENSION 8

Ownership and governance: CEO vs CIO

ModelOwner in large firmsOwner in mid-sized firmsFirms with governance in place (mid / large)
Grok 4.7CIO/CTO 34%; CEO 14%CEO/owner 48%35% / 55% of firms in production
Grok 4.6CIO/CTO 40%; CEO 18%Owner-CEO 45%8% / 25%
Grok 4.5CIO/CTO 35%; CEO 15%CEO 35%10–20% / 25–40% of firms in production
Composer 2.5CIO/CTO 35%; CEO 10%IT lead 40%; CEO 25%10–15% / 25–35%
Common ground

The CEO should own the agenda: which process, what the agent must never do, how the savings are used, and what day 90 must show. The CEO should not own vendor selection or prompts. If the topic is handed entirely to IT, it becomes a pilot that never touches the ledger. If the CEO buys straight from a vendor, it becomes shadow access with no log. The deploying company is accountable for an agent's mistakes. Insurers are not ready, and boards discuss AI more often than they oversee agents.

Disagreements

Grok 4.7 and Grok 4.6 put the owner-CEO in charge in 45–48% of mid-sized firms. Composer 2.5 says IT leads more often (40% vs 25%). Grok 4.7 is the harshest on governance: about 2% of mid-sized and 10% of large European firms overall have most of it in place. Composer 2.5 alone warns that managed-service providers sometimes host the agent while the CEO believes 'we have AI'.

Reality check · the data

Here the models are behind the data. BCG (Jan 2026) found that 72% of CEOs now say they are the main decision maker on AI, twice last year's share, against the models' 10–18% for large firms. The dirigeant drives the AI project in 73% of French PME/ETI (Bpifrance). 'Main AI decision maker' is a wider role than owning an agent's go-live, but the direction is clear. On governance the models are right: only 21% of firms have a mature model for agent governance (Deloitte). Controls are rising fast in the US. 75% of $1B+ firms no longer let agents near sensitive data without human oversight, up from 36% a quarter earlier, and 43% set token budgets (KPMG Q3 2026).

Synthesis

The CEO already holds the decision. What is missing is the governance under it: agent identities, approval rules, logs, spend caps and a same-day off switch.

DIMENSION 9

Timeline to daily use

ModelMajority, large USLarge EUMid-sized EUMost office staff use agents daily (large EU)
Grok 4.72028202920312031
Grok 4.62028202920312032
Grok 4.52028202920312031–33
Composer 2.52028203020332033
Common ground

All four give 2028 for a majority of large US firms, and all are more confident in the order (large US, large Europe, mid-sized US, mid-sized Europe) than in the years. Counting agents built into existing software pulls every date forward by about a year. The triggers are vendors building agents into the software firms already run, more reliable tool use, labour shortages, and success stories a CFO will believe. On the hype cycle, the grand claims are entering the trough while narrow agents climb the slope.

Disagreements

Composer 2.5 is one to two years later than the Grok models for Europe, and alone suggests that open-weight models could pull mid-sized timelines forward by two years. Grok 4.7 says the CEO conversation is 'at or just past the peak' while practitioners are already in the trough. Grok 4.6 describes a 'split hype cycle'.

Reality check · the data

Executives expect it sooner. 74% expect to use agentic AI at least moderately within two years (Deloitte, 2026). 81% of leaders expected agents to be moderately or extensively integrated into their AI strategy within 12–18 months (Microsoft, Apr 2025). Gartner is more cautious, expecting over 40% of agentic projects to be cancelled by end-2027, but also forecasts agentic AI in a third of enterprise apps by 2028. Bitkom offers a useful analogy: agents 'stand where AI overall stood three years ago', and German AI use went from 20% to 57% in two years.

Synthesis

If agents follow AI's own curve in Germany, the models' 2029 date for most large European firms looks reasonable and may be a year late. The date for mid-sized firms depends on vendors shipping usable agents inside the software those firms already rent.

Further reading: The wider economy: The AI Economy in 2030.

DIMENSION 10

What a ready company looks like, and the risks

ModelMid-sized EU meeting most criteriaLarge EUTop operational risk
Grok 4.74%15%Wrong writes into real systems
Grok 4.68%20%Wrong action in a system of record
Grok 4.55–10%15–25%Wrong actions from brittle integrations
Composer 2.510–18%30–40%Wrong action / hallucinated tool calls
Common ground

All four produced checklists of nine or ten items that overlap almost completely (see the checklist below). All four rank the top risks the same way: a wrong action in a real system, then data leakage, then prompt injection from an email or PDF. Ordinary errors will cost more in total than rare security incidents. Through 2028 the main effect on jobs is changed roles and skipped hires, not layoffs, especially in mid-sized firms that are already short of staff.

Disagreements

Composer 2.5 thinks two to three times as many firms are ready as Grok 4.7 does. In the 90-day plan, Composer 2.5 asks for a kill-switch demo and 90 days of replayable logs. Grok 4.7 rules out a second agent within the window, Grok 4.6 bans the executive briefing agent and the autonomous sales agent, and Grok 4.5 adds stopping unmonitored personal ChatGPT use on sensitive data.

Reality check · the data

Readiness is low in every survey. Fewer than one in five firms report high data readiness and 80% lack mature AI infrastructure (Capgemini). 21% have mature agent governance (Deloitte). A third (33.4%) of Italian firms using AI say it has no defined business purpose (Istat). 43% of French PME/ETI do not analyse their data (Bpifrance). On jobs, 39% of executives expect AI-driven headcount declines next year, but only 14% saw one last year, against the 32% who expected one (McKinsey 2026).

Synthesis

Readiness is a checklist, not a culture, and most firms could pass it for one process within a quarter.

Further reading: Knowledge readiness: The agent-readable second brain. Check your company: AI readiness scorecard (theagency47, 10 questions).

Which model said what · summary table T1–T15 · 17 October 2026

Which model said what

Every model filled in the same fixed table at the end of its answers. Here are the fifteen rows side by side, with the closest checked figure from our sources. A checked figure often measures something wider or narrower than the row, so read it as a bracket, not a verdict.

IdItem Grok 4.7 Grok 4.6 Grok 4.5 Composer 2.5 Closest checked data
T1Large EU firms with an agent in daily use18%18%18%22%No official series. Bitkom: 11% of AI-engaged German firms use agents
T2Mid-sized EU firms, same6%7%7%8%Eurostat: 30.4% of medium firms use any AI (upper limit)
T3Large US firms, same28%28%28%35%KPMG ($1B+): 62% building, deploying or developing
T4Large-company CEOs who can explain an agent35%36%40%40%Capgemini: half of leaders say they understand agents
T5Pilots that reach daily production30%30%25%25%Deloitte: 25% of firms moved ≥40% of pilots to production
T6CEOs with a measurable financial return15%20%20%*20%PwC (AI overall): 12% both, 33% either, 56% none
T7First agent, build + year 1 (EUR)150k120k80k120kNo survey. theagency47 fixed packages from €2.5k
T8Majority of large EU firms2029202920292030Deloitte: 74% expect agentic use within 2 years
T9Majority of mid-sized EU firms2031203120312033No survey
T10Adoption rankUS DE FR IT GRUS DE FR IT GRUS DE FR IT GRUS DE FR IT GREurostat any-AI: DE FR IT GR (large firms: FR ≥ DE)
T11Function agents enter firstCustomer service; IT (large)Customer serviceCustomer serviceCustomer service / ITCapgemini: customer service, IT, sales
T12Top barrier, mid-sized EUIntegrationIntegrationSkillsIntegrationEurostat: lack of expertise 70.9%
T13Who owns agent decisionsCEO mid; CIO largeCEO mid; CIO largeCEO mid; CIO largeCEO+IT mid; CIO largeBCG: 72% of CEOs say they are the main AI decision maker
T14Mid-sized EU firms agent-ready4%8%8%14%Capgemini: fewer than 1 in 5 report high data readiness
T15Chance of an incident-led slowdown by end-202735%35%35%35%No data. The identical figure suggests a shared anchor

* Grok 4.5: 20% of CEOs whose firms tried agents, about 8% of all CEOs. Grok 4.7: 15% of large-company CEOs. Model values are point estimates, all tagged [ESTIMATE] except Grok 4.5's T15 [GUESS] and Composer 2.5's T13 [GUESS].

Consistent across the four

  • All 4
    Same order everywhere

    Large US ahead of large Europe, then mid-sized US and mid-sized Europe. US, Germany, France, Italy, Greece. Every model is surer of the order than of the levels.

  • All 4
    Tagged as estimates

    All four separated what they knew (Gartner's 40% cancellation forecast, AI Act dates, Eurostat 2023–24) from guesses, and none treated the bank examples in the prompt as verified.

  • All 4
    35% for an incident-led slowdown

    The same number from four models is a warning sign: it looks like a shared anchor, not four independent judgements.

Where they consistently split

  • Speed
    Composer 2.5 highest today, latest tomorrow

    Composer 2.5 gives the highest adoption levels for 2026, yet the latest dates for a majority (2030 large, 2033 mid-sized Europe).

  • Rigour
    Grok 4.7 the strictest

    The lowest readiness (4%), the highest first-agent cost (€150k, including internal time) and the harshest governance count.

  • Cost
    Grok 4.5 the cheapest and the fastest payback

    €80k all-in and 6–18 months. It is also the only model to put skills ahead of integration.

Claims only one model made

Did the models favour their maker?

No. Three of the four are Grok models from xAI (SpaceXAI), and Composer 2.5 is Cursor's own model. None named its maker's products as an agent platform or example. All four named Microsoft, Salesforce, SAP and ServiceNow as the embedded path most firms will take. We also checked whether they favoured our sister company, whose positions were in the prompt. They did not: all four only partly agreed with six of the eight positions, and one disagreed outright.

Section R · the models react to theagency47's country guides

Verdicts on theagency47's eight positions

On 17 October 2026 theagency47 published four guides for CEOs: for Greek, Italian, French and German businesses. A summary was given to every model, which was asked to agree, partly agree or disagree with eight positions and to say what the guides miss.

Disclosure. theagency47 is a House of 47 company, founded by the author of this research. We report the models' criticisms in full, and we add one of our own.

Idtheagency47 position Grok 4.7 Grok 4.6 Grok 4.5 Composer 2.5 What the checked data says
R1Start with one Operational or Task agent, not an executive briefing agentAgreeAgreeAgreeAgreeSupported: only 7% of German firms use AI in management (Bitkom); agents scale first in IT and service (McKinsey, Capgemini).
R2Smaller firms lag mainly for lack of skills and a translator, not interest or fitPartlyPartlyPartlyPartlyThe surveys side with theagency47: skills are the top barrier (Eurostat 70.9%, Bitkom 85%, Istat 58.6%). The models add integration, volume and budget.
R3Human approval + logging + least privilege is enough for daily usePartlyPartlyPartlyPartlyAll four want a test set, prompt-injection defences and a kill switch too. KPMG: 75% of $1B+ US firms keep humans over sensitive-data access.
R4Typical back-office agents are not high-risk; the main duty is to disclose AIPartlyPartlyPartlyPartlyStill true. Disclosure has been enforced since 2 Aug 2026; Annex III high-risk moved to 2 Dec 2027, which the guides omit. GDPR duties remain.
R5Bank assistants (Sofia, Ellis, Ava) show a pattern that scales to 40–300 staffPartlyPartlyDisagreePartlyNot checked in our dossier; the bank figures are as cited by the guides. All four: copy the narrow scope, not the bank's stack.
R6A shared knowledge base ('Second Brain') is the main precondition, and the main value for family firmsPartlyPartlyPartlyPartly†Partly supported: under 1 in 5 firms have high data readiness (Capgemini); 33.4% of Italian AI users have no defined purpose (Istat).
R7Owning the agent environment (own cloud, no lock-in) matters to European CEOsPartlyPartlyPartlyPartlyNo survey in our dossier measures this. Three models: CEOs say it matters, then buy convenience.
R8Country framings: GR time & scattered info; IT know-how & handover; FR compliance; DE labour shortage & works councilAgreeAgreeAgreeAgreeMostly holds. But Greece went backwards on AI use (Eurostat), and French CEOs are the most bullish on agents (EY: 44% vs 37% globally).

† Composer 2.5: 'Agree' that a knowledge base is needed for sustained usefulness; 'Partly agree' that it is always the main precondition. "Partly" = partly agree. Tally: 2 positions with full agreement from all four (R1, R8), 5 where all four partly agree, 1 split (R5: Grok 4.5 disagrees).

Where the models push back

The main criticism is that the method is right but the guides make it look simpler than it is. Human approval, logs and least privilege are the right minimum (R3). They are not enough once an agent takes more of the queue, and an approval button that people click through 200 times becomes "theatre" (Grok 4.7). The AI Act summary is "a good CEO heuristic and a bit too tidy" (Grok 4.6): disclosure is not the only duty, and HR, worker scoring and credit can make an agent high-risk. The bank examples are customer-service assistants, while a mid-sized firm should usually start in the back office.

Where the data backs the guides

On R2 the surveys side with theagency47 against the models: skills are the barrier firms cite most often (Eurostat 70.9%, Bitkom 85%, Istat 58.6%). The guides' Istat and Insee figures match the official releases. Their AI Act reading still holds for ordinary agents. But the guides do not mention the AI Omnibus: high-risk Annex III duties now apply from 2 December 2027, not August 2026, and Annex I from 2 August 2028. That should be added.

What the guides miss, according to the models

Countries · Greece, Italy, France, Germany, United States

Country by country

For each country: what theagency47's guide says, what the four models estimate, what the official statistics and surveys show, and our read. Model figures are shares of firms with an agent in daily use; official figures are shares using any AI, a much wider measure.

COUNTRY · Read the theagency47 guide ↗

Greece

The guide says

Lost time, repetitive work and scattered information. Many firms have 'tried AI' next to the work, not inside it. The flagship example is the National Bank of Greece's phone agent 'Sofia' (July 2026). The guide cites no national statistics.

The models say

Last of the five in all four models: 9–10% of large and 3% of mid-sized firms with an agent in daily use (Composer 2.5: 7% blended). Micro firms dominate, the systemic banks are the exception, and a €150,000 experiment is a board-level risk for an 80-person firm (Grok 4.7).

The data says

Eurostat: only 8.9% of Greek firms used AI in 2025, down from 9.8%, the only one of the four to fall, and mid-sized firms also slipped (15.2%). Large firms rose to 37.6%. AWS/Strand: 47% use 'AI tools', but only 22% have heard of agentic AI, and of those 2% have fully deployed it. PwC Greece: 75% of CEOs say they are ready, more than 60% have seen no meaningful financial benefit, and 10% saw both revenue and cost gains (n=62).

Our readThe widest gap between confidence and use in this study. Start where the information is scattered: a shared inbox or order emails.

COUNTRY · Read the theagency47 guide ↗

Italy

The guide says

In the family PMI, the real 'manual' is in a few people's heads, which becomes fragile during growth or a generational handover. Agents force that know-how to be written down. Cites Istat (16.4% AI use in 2025) and Intesa Sanpaolo's 'Ellis'.

The models say

Fourth in all four: 13–15% of large and 5% of mid-sized firms. Northern districts (Lombardy, Emilia-Romagna, Veneto) have the most latent fit because succession is urgent, but the national average lags. Composer 2.5: PNRR e-invoicing is a 'feeder' for invoice agents.

The data says

The Istat figures in the guide check out: 16.4% in 2025 (8.2% in 2024), large firms 53.1% and SMEs 15.7%, skills a barrier for 58.6%. Istat also finds that 33.4% of AI-using firms have no defined business purpose. Agents are rare: 1% of very large Italian firms had them in their processes, while 43% plan to within 2–3 years (Capgemini, 2025). 71% of large firms have started an AI project against 8% of SMEs (Politecnico di Milano).

Our readThe handover argument is the strongest of the four. Writing the know-how down pays off even without an agent.

COUNTRY · Read the theagency47 guide ↗

France

The guide says

Written for the dirigeant of a PME/ETI, with compliance first: GDPR, CNIL guidance, and the DPO and CSE involved early. Cites Insee (18% in 2025, 58% of 250+ firms vs 15% of 10–49) and Carrefour's 'Hopla', which works because it is wired to the catalogue and the basket.

The models say

Third in all four: 17–18% of large and 6–7% of mid-sized firms. 'The French delay is "we will do it properly"' (Grok 4.6). Grok 4.7 alone says France and Germany could swap places.

The data says

Insee confirms the guide: 18% of firms, 31% of 50–249 firms (EU 30%) and 58% of 250+ firms (EU 55%). Among large firms France (58.0%) edges Germany (57.0%) (Eurostat). At the top, French CEOs are the most bullish: 44% cite autonomous agents (37% globally) and 66% make AI their first investment priority (EY, 2026). Bpifrance: the dirigeant drives AI in 73% of PME/ETI, but 27% of leaders are sceptics and 19% advanced.

Our read'Compliance-first' describes the PME, not the CEO at the top. France's large firms already match Germany's.

COUNTRY · Read the theagency47 guide ↗

Germany

The guide says

'The work is there, the people aren't.' Agents absorb back-office load in the Mittelstand. Cites Destatis (26% AI use; barriers: know-how 72%, legal 62%, data protection 60%) and Commerzbank's 'Ava'. Brief the works council before the pilot and agree how logs may be used.

The models say

Second in all four and the fastest of the four European countries to 2028: 22% of large and 8–9% of mid-sized firms (Composer 2.5: 14% blended). The labour shortage provides the pull, and the Betriebsrat slows the first agent and then makes it stick.

The data says

Eurostat: 26.0% (medium 35.6%, large 57.0%). Bitkom (Sep 2026): for the first time most firms (57%) use AI; 11% of AI-engaged firms already use agents, 29% plan to and 31% are discussing it. 'Agents stand where AI overall stood three years ago.' Data protection is the top barrier for users (66%). We could not re-read the Destatis barrier figures, so they appear here as cited by theagency47.

Our readBest soil for narrow agents that save hours, and the only country with an agent-specific survey showing double-digit use.

COUNTRY

United States

Benchmark

No theagency47 guide. The US is the benchmark in the questionnaire.

The models say

First in all four: 28% (Grok models) to 35% (Composer 2.5) of large and 11–14% of mid-sized firms. Majority of large firms in 2028. Regulation explains only part of Europe's lag; vendor proximity, capital and an English-language toolchain explain more.

The data says

Large firms are further ahead than the models think. KPMG (Q3 2026, $1B+): 62% are building, deploying or developing agents, 25% use multi-agent systems, and 44% report significant workforce adoption. PwC US (2025): 79% say agents are being adopted, but in 68% of firms half or fewer of employees use them. Mid-market (RSM): 91% use generative AI, 25% fully integrated into core operations. US CEOs are less confident than Europeans that AI will pay off (52% vs 61%, BCG).

Our readThe lead is real at the very top. The mid-market looks more like Europe than the headlines suggest.

What a ready company looks like · synthesis of the four models' checklists

What a ready company looks like

Each model wrote its own checklist of nine or ten points without seeing the others. They overlap almost completely. Grok 4.6: "Readiness is a checklist, not a culture." The models think 4–14% of European mid-sized firms pass it today.

The mark shows how many of the four models listed the criterion. How the knowledge layer is built (plain-text procedures that staff and agents both update) is the subject of our research The agent-readable second brain. To score your own company, theagency47 offers a free 2-minute AI readiness scorecard.

The first 90 days

The four 90-day plans in one. It is also the path theagency47's guides recommend: one process, a baseline, one agent, a decision.

Days 1–15

Pick and measure. CEO, operations and IT pick one frequent, document-heavy process where a mistake can be undone. Write it down in two pages. Start the baseline. Rule out the executive briefing agent and the autonomous sales agent.

Days 16–30

Check and choose. Map the systems and the personal data. Check for high-risk uses (hiring, credit, worker scoring). Brief the works council or CSE. Choose between an agent built into your software, an off-the-shelf product or an integrator. The rule from day one: the agent prepares, a person approves.

Days 31–60

Test in the shadow. Run 30–50 past cases under a low-privilege account with a spend cap, first read-only, then with approved writes. Train the five people who will use it.

Days 61–90

Run and decide. Limited production for one team. Hold a weekly error review. Keep it, fix it or kill it on the numbers. Do not start a second agent in this window. If you stop, keep the written process. Budget for the quarter: €20,000–80,000 plus internal time (Grok 4.6).

House of 47 read · How to think about it

Informed, not yet ready

Count daily use, not announcements

Ask how many people use the agent every working day, in which process, and with what error rate. Surveys reporting more than 60% adoption mostly measure intent.

The CEO already owns this

72% of CEOs say they are the main AI decision maker. The job is to set what the agent may never do and what day 90 must prove, then let IT build the controls.

Skills to start, integration to finish

Hire or rent the translator first, then budget for the plumbing. Neither is solved by a bigger model.

One agent, 90 days, a number

Every model and theagency47's guides end in the same place: one painful process, a baseline, a human on the button, and a keep-or-kill decision.

Key divergences
  • !
    Ownership

    All four models say the CIO decides in large firms. BCG says 72% of CEOs now do. The models learned from a world that has already changed.

  • !
    Barrier

    Three models say integration and one says skills. The surveys say skills, because they ask why firms do not start.

  • !
    Speed at the top

    The models put large US firms at 28–35%, while KPMG finds 62% of $1B+ firms building or deploying and 25% running multi-agent systems.

  • !
    Diversity

    Three of four answers come from the Grok family, plus Cursor's Composer. Two models gave identical country tables, and all four gave 35% for T15. Agreement partly reflects shared training.

Open questions
  • Q

    How many firms that say they are 'scaling agents' have one used daily by ordinary staff? No survey measures this yet.

  • Q

    Will Greece's fall in AI use (Eurostat 2025) reverse in the 2026 data?

  • Q

    How quickly will CEOs' ownership of the AI decision turn into governance: agent identities, approval rules and an off switch?

  • Q

    Does the December 2027 high-risk date change anything for HR and credit agents in mid-sized firms?

From the House of 47 group · related work

Where the group comes in

AI agents

This research ends where theagency47 starts: one agent in one process, running in the client's own cloud account, with logs, spending limits and a person approving anything consequential. Mid-sized firms usually begin with Spark (one agent) or Workforce Starter. Large companies look at Enterprise and the OS47. A 30-minute discovery call is free.

Agents at work

search47 is run by 25+ named AI agents supervised by senior strategists, an example of agents in daily operations inside the group. As buyers' own agents start doing supplier research, being the answer that ChatGPT, Gemini and Perplexity give is what its generative engine optimisation work covers (see also AI Search vs Google).

The front door

Customer-facing agents such as Carrefour's Hopla need a fast, modern site to live on. That is the work of pixel47, the group's web design and development studio.

House of 47 group: House of 47 · theagency47 · search47 · pixel47 · 47HNWI. theagency47 is a group company; see the disclosure in the theagency47 section.

CEOs know the word. Few have run the work. Start with one process.

Methodology · models, questionnaire, sources, raw data

Methodology

Models used · 17 October 2026 · Cursor cloud agents
ModelRequested idSelf-reported idMakerRaw file
Grok 4.7grok-4.7grok-4.7xAI / SpaceXAIraw/grok-4.7.md
Grok 4.6grok-4.6cursor-grok-4.6-high-fastxAI / SpaceXAIraw/grok-4.6.md
Grok 4.5grok-4.5cursor-grok-4.5xAI / SpaceXAIraw/grok-4.5.md
Composer 2.5composer-2.5composer-2.5Cursorraw/composer-2.5.md

Four models answered: three from the Grok family (xAI) and Cursor's Composer 2.5. Each Cursor cloud agent received the identical prompt (cloud-agent-prompt.md, with questions.json). Each was told to answer every question itself, from its own knowledge, in English, with no web browsing, no other models and no code, and to tag every number [KNOWN], [ESTIMATE] or [GUESS]. One run per model, default parameters. It is the same multi-model method as our Home robots and Humanoid robots in everyday life studies. For how these models compare on benchmarks and price, see Best AI models 2026, compared.

Reality check. Separately, we collected 30 published sources dated 2024–2026 (CEO and executive surveys, official statistics from Eurostat, ELSTAT, Istat, Insee and Destatis, country studies, and the EU AI Act timeline), each read on the source page on 17 October 2026. The models did not see them. Only figures recorded in that dossier are used here. Caveats: consultancy and vendor surveys (PwC, KPMG, BCG, Capgemini, IBM, Deloitte, Microsoft, AWS) sample firms already engaged with AI, mostly with revenue above $500M or $1B, and are not representative. Official statistics are representative but measure any AI use, not agents. PwC's and BCG's figures were read on their press releases, because the reports could not be downloaded. The Destatis page timed out, so German figures come from Eurostat's harmonised dataset. The AWS/Strand Greek study is vendor-commissioned. Model ranges are opinion, not data; where they conflict with checked figures, the checked figures stand.

The preamble (identical for every model)

Research question: are CEOs of mid-sized and large companies informed about, and ready to integrate, AI agents into their company's daily operations? Definitions used throughout: an 'AI agent' is software built on a foundation model that carries out multi-step work inside a business process (reads inputs, decides within limits, takes or prepares actions in company systems), as opposed to a chatbot that only answers or a fixed rule-based automation. 'Mid-sized' = 50-249 employees (EU definition); 'large' = 250+ employees; 'very large' = more than EUR/USD 1 billion annual revenue. 'Daily operations' = agents used every working day in a real process by ordinary staff, not a pilot or a demo. Time frame: today is October 2026. Answer from your own knowledge. Give concrete numbers (percentages, years, euro amounts) wherever you can, each as a point estimate with a plausible range, and say how confident you are. Separate clearly what you know from published data (name the source and year if you can) from what you are estimating or guessing. If you do not know, say so rather than invent a statistic.

The 54 questions, plus section R (8 positions) and the fixed table T1–T15

D1 · Awareness vs hype. D1.1. What share of CEOs of mid-sized and of large companies (separately) can correctly explain what an AI agent is and how it differs from a chatbot or RPA? Give % estimates for Europe and the US. D1.2. How much of current CEO enthusiasm for agents is driven by vendor marketing, board/investor pressure and fear of missing out, versus evidence from their own operations? D1.3. What are the most common CEO misconceptions about AI agents in 2026 (both over- and under-estimation)? D1.4. How large is the gap between what CEOs say in surveys about agents and what is actually running in their companies? Estimate it. D1.5. Where do CEOs of mid-sized firms actually get their information about AI agents (peers, consultants, vendors, media, own staff), and how reliable is it?

D2 · Actual adoption vs pilots. D2.1. Today (Oct 2026), what share of mid-sized and of large companies in Europe have at least one AI agent in daily production use (not a pilot)? And in the US? D2.2. What share are only piloting or experimenting, and what share have done nothing with agents yet? D2.3. What fraction of agent pilots reach production, and what fraction are abandoned? Why do most stall? D2.4. How much of reported 'agent adoption' is really agent features embedded in existing software (CRM, ERP, office suites) rather than agents the company deliberately deployed? D2.5. Give one or two concrete, verifiable examples of large European companies running agents in daily operations, and say how sure you are that each example is real. D2.6. How do you expect these adoption shares to change by end-2027?

D3 · Budgets and ROI. D3.1. What do mid-sized and large companies typically spend per year on AI agents (euro ranges, and as % of revenue or of IT budget)? D3.2. What does a first production agent cost a mid-sized company (build, integration, running costs) and how long is a realistic payback? D3.3. What share of CEOs report measurable financial returns from agents so far? Is ROI evidence improving in 2026? D3.4. Which budget lines are underestimated (data preparation, integration, change management, model/token costs, oversight staff)? D3.5. Will agent budgets keep growing in 2027 if returns stay unclear? Why or why not?

D4 · Barriers: skills, data, governance, regulation. D4.1. Rank the top barriers to agent adoption for European mid-sized companies (skills, data quality, legacy systems/integration, cost, security, legal uncertainty, employee resistance, lack of use cases). Give rough weights. D4.2. How ready are company data and knowledge (documents, procedures, ERP/CRM data) for agents? Estimate the share of mid-sized firms whose data is 'agent-ready'. D4.3. What does the EU AI Act actually require of a company deploying a typical back-office or customer-service agent, and on what dates? How well do CEOs understand this? D4.4. How much do GDPR, works councils (e.g. German Betriebsrat, French CSE) and labour law slow agent rollouts in practice? D4.5. Is the skills gap mainly at the top (CEO/board), in IT, or among staff who must work with agents? D4.6. Which barrier is the most overrated and which the most underrated?

D5 · Mid-sized vs large companies. D5.1. How far apart are mid-sized and large companies in agent readiness and adoption? Quantify the gap. D5.2. What advantages do mid-sized companies have (speed, owner-CEO, fewer systems) and what handicaps (no data team, budget, vendor attention)? D5.3. Should a mid-sized company build, buy off-the-shelf agents, use agents embedded in its existing software, or hire an agency/integrator? Under what conditions does each make sense? D5.4. Are multi-agent systems realistic for mid-sized firms in 2026-2027, or should they stay with one or two narrow agents? D5.5. Does the gap between mid-sized and large firms widen or narrow by 2028? Why?

D6 · Country differences: Greece, Italy, France, Germany vs US. D6.1. Rank Greece, Italy, France, Germany and the US on CEO awareness of AI agents and on actual adoption in mid-sized and large firms, with a rough % adoption estimate for each. D6.2. What is specific about Greek companies (size structure, family ownership, tourism and shipping, digital maturity, funding) for agent adoption? D6.3. What is specific about Italy (family PMI, generational handover, manufacturing districts) and France (PME/ETI, compliance culture, CNIL, CSE)? D6.4. What is specific about Germany (Mittelstand, labour shortage, works councils, data protection culture, ERP landscape)? D6.5. Why does Europe trail the US (if it does), and is the gap narrowing or widening? How much is due to regulation versus other factors? D6.6. Which of the four European countries is likely to move fastest in 2026-2028, and why?

D7 · Which functions and levels agents enter first. D7.1. In which business functions do agents enter first in mid-sized and large companies (customer service, sales, finance/AP, IT, HR, procurement, logistics, legal, marketing)? Rank them and say why. D7.2. Using three levels of work (Executive: decision support for leadership; Operational: day-to-day department flows; Task: high-volume repetitive work), where should a company put its first agent and why? D7.3. Which processes look attractive but usually fail as first agent projects? D7.4. What share of a typical back-office employee's work could agents handle reliably in 2026, and by 2028? D7.5. Do customer-facing agents (e.g. bank assistants) or internal back-office agents spread faster, and why?

D8 · Ownership and governance: CEO vs CIO. D8.1. Who actually owns AI agent decisions in mid-sized and large companies today: CEO, CIO/CTO, CDO/Chief AI Officer, COO, or business unit heads? Give rough shares. D8.2. Should the CEO personally own the agent agenda? What goes wrong when it is delegated entirely to IT? D8.3. What governance does a company need before agents act in its systems (permissions, human approval, logging, budgets, incident handling)? What share of companies have it? D8.4. Who is accountable when an agent makes a costly mistake, and are boards and insurers ready for that question? D8.5. How involved are boards in AI agent oversight in 2026?

D9 · Timeline to daily use. D9.1. In what year will a majority (more than 50%) of large European companies use at least one agent in daily operations? And of mid-sized European companies? And in the US? D9.2. In what year will agents be used daily by a majority of office employees in large companies? D9.3. What are the key milestones or triggers (technology reliability, prices, regulation dates, success stories, labour shortages) that set the pace? D9.4. What could slow the timeline sharply (a major agent failure, a security incident, regulation, an AI market correction)? D9.5. Where are we on the hype cycle for AI agents in October 2026?

D10 · What a ready company looks like, and the risks. D10.1. Describe concretely what an 'agent-ready' mid-sized company looks like (people, data, systems, processes, rules, budget). Give 6-10 checkable criteria. D10.2. What share of mid-sized and of large European companies meet most of these criteria today? D10.3. What are the main operational risks of agents in daily use (errors, hallucinated actions, security/prompt injection, data leakage, cost overruns, vendor lock-in, skill erosion)? Rank them. D10.4. What are the main workforce effects over 2026-2028 (job losses, role changes, morale, hiring freezes) in mid-sized vs large firms? D10.5. What should a CEO do in the next 90 days if the company has no agent yet? Give a concrete plan. D10.6. What is your single most important message to European CEOs about AI agents?

Limits

Diversity is limited: four models were used, three of them from the Grok family plus Cursor's Composer, so agreement between them is weaker evidence than it looks (two gave identical country tables). The prompt included a summary of theagency47's guides, a House of 47 company, for section R. Every model answered on its own and none had live data, so recent events (the AI Omnibus, the 2025 statistical releases) were unknown to them.

Sources · figures checked against primary reporting · read on 17 October 2026

Sources

Grok 4.7 (grok-4.7), answers to the identical questionnaire, Cursor cloud agent, 17 Oct 2026, raw/grok-4.7.md

Grok 4.6 (cursor-grok-4.6-high-fast), answers to the identical questionnaire, Cursor cloud agent, 17 Oct 2026, raw/grok-4.6.md

Grok 4.5 (cursor-grok-4.5), answers to the identical questionnaire, Cursor cloud agent, 17 Oct 2026, raw/grok-4.5.md

Composer 2.5 (composer-2.5), answers to the identical questionnaire, Cursor cloud agent, 17 Oct 2026, raw/composer-2.5.md

PwC, 29th Annual Global CEO Survey, 19 Jan 2026 (press release; 4,454 CEOs), link

PwC Greece, CEO Survey 2026, Greek cut (n=62), via Parapolitika, link

KPMG, 2025 Global CEO Outlook, Oct 2025 (1,350 CEOs, >$500M revenue), link

KPMG US, AI Quarterly Pulse Survey Q3 2026, 24 Sep 2026 (314 leaders, ≥$1B), link

IBM IBV, 2025 CEO Study, 6 May 2025 (2,000 CEOs), link

BCG, AI Radar 2026, 15 Jan 2026 (press release; 640 CEOs), link

EY-Parthenon, CEO Outlook 2026, Jan 2026 (global PDF), link

EY France, CEO Outlook 2026, French CEOs, link

PwC US, AI Agent Survey, May 2025 (308 executives), link

Microsoft, Work Trend Index 2025, 23 Apr 2025, link

McKinsey, The state of AI in 2026, 25 Aug 2026 (1,719 respondents), link

McKinsey, The state of AI in 2025, 5 Nov 2025, link

Deloitte, State of AI in the Enterprise 2026, 21 Jan 2026 (3,235 leaders), link

Capgemini, Rise of agentic AI, 16 Jul 2025 (1,500 executives, >$1B), link

Il Sole 24 Ore, Capgemini Italy data on agents, 25 Jul 2025, link

Gartner, Over 40% of agentic AI projects canceled by end-2027, 25 Jun 2025, link

Gartner, 40% of enterprise apps with task-specific agents by 2026, 26 Aug 2025, link

MIT Project NANDA, The GenAI Divide, Jul 2025 (small, non-random sample; often misquoted), link

RSM, Middle Market AI Survey 2025 (US & Canada), link

Eurostat, 20% of EU enterprises use AI technologies, 11 Dec 2025 (dataset isoc_eb_ai), link

Eurostat, Use of artificial intelligence in enterprises, Statistics Explained, link

ELSTAT, ICT usage in enterprises 2025 (Greece), link

Istat, Imprese e ICT, anno 2025 (Italy), link

Insee, Insee Première n° 2120, 21 Jul 2026 (France), link

Destatis, ICT use in enterprises (page timed out on 17 Oct 2026; figures via Eurostat), link

Bitkom, Erstmals nutzt die Mehrheit der Unternehmen KI, 14 Sep 2026 (603 firms), link

Politecnico di Milano, Osservatorio Artificial Intelligence, 5 Feb 2026, link

Politecnico di Milano SoM, Gen-AI & Agentic-AI nelle imprese italiane, Apr 2026, link

Bpifrance Le Lab, L'IA dans les PME et ETI françaises, 4 Jun 2025, link

AWS / Strand Partners, Unlocking Greece's AI Potential 2026, Oct 2026 (vendor-commissioned), link

European Commission, AI Act Service Desk, implementation timeline, link

European Commission, AI Omnibus enters into force, 27 Jul 2026, link

European Commission, Enforcement and transparency requirements from 2 August, 31 Jul 2026, link

theagency47 (group company), Do you need an AI agent? Guide for Greek businesses, 17 Oct 2026, link

theagency47 (group company), Guide for Italian businesses, 17 Oct 2026, link

theagency47 (group company), Guide for French businesses, 17 Oct 2026, link

theagency47 (group company), Guide for German businesses, 17 Oct 2026, link

More House of 47 research
Changelog · what changed in this brief, newest first

Changelog