Every company that puts AI agents to work needs one place those agents read from: the procedures, rules, decisions, product facts and approved answers. We call it the second brain. We asked four AI models, Grok 4.7, GPT-5.6, Claude Opus 5.5 and Gemini 3.1 Pro, how a business should build one at five company sizes, and to tear apart the design under review. Here is where they agree, where they split, and what we would build. Research date: 14 October 2026.
See it by company sizeAll four models converge on the same skeleton. People write and approve knowledge as plain files. Agents search an index built from those files, which can be thrown away and rebuilt. Bookings, invoices and stock stay in the systems that own them and are asked for at the moment of need. And nothing reaches the brain, or the customer, without passing a human and a test. What changes with company size is how much machinery sits around each layer.
Markdown with YAML frontmatter, one topic per note, organised by kind of knowledge: entities, procedures, rules, decisions, FAQ, glossary. Owned, versioned, portable to any model.
Keyword (BM25) plus vector search, built per audience, rebuilt from git in CI. If it is lost, it is rebuilt in minutes. Small companies may not need it at all yet.
CRM, ERP and PMS data is queried through narrow, read-only MCP tools with their own credentials. A price written into prose is a future wrong answer.
Agents propose changes, a named owner approves, and a known-answer test set runs before anything goes live. All four models say this is where projects live or die.
Knowledge lives as markdown in git. The vector or search index is a derived artifact you can rebuild, which also lets you swap embedding models without migrating knowledge.
Prices, bookings, stock and invoices are read at task time from the system of record, through scoped tools, not copied into notes or the vector store.
A poisoned ticket that can commit to the handbook is a permanent injection; one that can only open a pull request is a bad draft. Every model kept this rule.
A customer bot should have no credential that reaches the confidential index. Separate indexes beat a metadata filter in one index.
Grok, GPT and Claude insist the first index include keyword search, for SKUs, names, law articles and Greek inflections. Gemini puts hybrid only at mid-market.
The protocol standardises access, but identity, scopes and consent must be enforced by the host and server. Gemini treats MCP more as the answer than the risk.
Grok, GPT and Claude say add a graph only after evals show multi-hop failures; links and entity tables give most of the value. Gemini sees it as necessary from mid-market.
From roughly 100 people, suite search (Copilot, Glean, Notion) handles "find any document I can see", beside a smaller curated canon for agents that answer customers or act.
Read the files directly as long as possible, or index early? Placed by the size at which each model adds an index.
The reference design suggests 20 to 30 known-answer questions on every change. Placed by how large each model wants the set.
GraphRAG for complex questions, or links and tables first? Placed by how central each model makes the graph.
Pin positions are House of 47's reading of each report, not scores the models gave. "Synthesis" is our own view, which is opinion.
The synthesis below takes the point where at least three models agree, and flags where we chose a side. Architecture and effort are recommendations, so opinion. Prices are third-party list prices with sources, checked by the models on 14 October 2026; they exclude model tokens and staff time, which all four models say dominate the bill.
Stack. A private git repo of markdown, edited in Obsidian or any editor (Obsidian is free for commercial use; Sync is $4 a user a month). The agent always loads a short _index.md or "how to read this" file, then opens the notes it needs, or loads the whole folder with prompt caching while it is small. No vector database yet.
Structure. Folders by kind of knowledge, a 1–2 sentence summary at the top of each note, and frontmatter with at least id, type, audience, owner, status and review date.
Write path. The founder owns everything. Agents drop proposals into an inbox folder or a pull request. A weekly 30-minute review.
Permissions. A separate public folder for any customer bot, from the first day. Read-only, narrowly scoped keys for live systems.
Evals. 10–50 real questions, including ones the bot must refuse, run by hand after each change.
Cost. About $0–35 a month of infrastructure; days to set up; 1–2 hours a week. A $20–25 seat plan (ChatGPT Business, Claude Team) may be enough if you only need an internal assistant.
Fails when. Personal notes quietly become customer truth, or a price in a note goes stale.
Stack. The same repo with protected branches and code owners (GitHub Team, $4 a user a month). Add a hybrid index when a second audience, a few hundred notes or several agents appear: Postgres full-text plus pgvector (Supabase Pro from $25 a month) or a managed service such as Cloudflare AI Search (billing starts 1 Nov 2026, $0.75 per 1,000 semantic queries after 1,000 free). Serve it through one retrieval MCP server.
Structure. Add a glossary with Greek, unaccented and Greeklish aliases, and keep exact facts in small data tables, not prose.
Write path. An owner and a backup owner per domain. Agents open pull requests; the owner merges. A weekly stale-content queue.
Permissions. A physically separate customer index and credential. The caller never chooses its own audience tag.
Evals. 20–30 questions in CI at minimum, plus refusals, leak tests and one planted injection. Claude wants 80–150.
Cost. About $10–150 a month of infrastructure; days to a few weeks to set up; a few hours a week.
Fails when. Vector-only search misses exact codes, or non-technical staff avoid git and stop editing.
Stack. Git canon plus Postgres doing keyword and vector search in one place, rebuilt in CI. A central retrieval service, an MCP gateway, and audience-specific tools. Add a reranker only if evals show a gain.
Structure. A schema enforced by a linter: required frontmatter, no retired notes in the index, currency amounts flagged if the number is live elsewhere.
Write path. The bottleneck moves from search to authorship. Domain stewards, risk-based approvals (two approvers for pricing, legal, safety), and a web "suggest a change" form that produces the same pull request, because Obsidian shared vaults cap at 20 collaborators.
Permissions. SSO groups, database roles or row-level security, sensitive domains in separate repos, and end-user OAuth on live-data MCP so a customer agent sees only that customer's booking.
Evals. 50–400 questions by audience, kind and language, with retrieval scored separately from answers.
Cost. About $100–1,000 a month of infrastructure, or about $2,000 a month for 100 Copilot Business seats at $21 list. Weeks to set up; 0.2–1 FTE.
Fails when. "Connect everything" imports drafts and contradictions, or access metadata is lost during chunking.
Stack. Split the job. Buy permission-aware search for the long tail of documents, mail and chat (Microsoft 365 Copilot, Glean, Notion Business at $20 a member). Keep a smaller governed canon for agents that answer customers or act, exposed through MCP behind a gateway that passes the user's identity through.
Structure. Domain canons with stable IDs linked to CRM and ERP master data; derived entity links where multi-hop questions are real.
Write path. A knowledge owner, domain stewards and segregation of duties, wired into existing ticketing.
Permissions. SSO and SCIM, attribute-based access, separate runtimes for customer agents, a DPIA where personal data is involved. Residency is decided per product, not per vendor.
Evals. Per-domain sets built from real traffic and incidents, red-team suites for exfiltration and poisoned documents, quality tracked per language and department.
Cost. Licences dominate: 300 seats at $20–48 is roughly $6,000–14,000 a month. Canon infrastructure is low hundreds to $3,000. About a quarter to roll out.
Fails when. A central team becomes the approval bottleneck, or a suite connector keeps the documents but not the nested permissions.
Stack. Enterprise search or an existing cloud search stack for the unmanaged corpus; a curated, schema-enforced canon for policies, products and approved answers; agents that act get the canon and scoped tools, never "search the whole company".
Structure. A governed federation and catalog that preserves records management, legal holds and source permissions. Do not copy the company into markdown.
Write path. Agents draft from tickets and transcripts; stewards and a council approve; content approver, platform operator and auditor are different people.
Permissions. Zero-trust, policy-as-code, workload identity, regional deployment, customer agents in separate accounts with their own subprocessors.
Evals. A continuous programme: stratified golden sets, shadow evaluation, red teams, rollback thresholds.
Cost. 500 Copilot seats at $30 list is $15,000 a month before Microsoft 365 licences. Glean publishes no price; Vendr reports a median contract of $98,890 a year. A standing team of 3–10 FTE.
Fails when. "One enterprise brain" becomes one exfiltration surface, or the pilot ignores deletion, legal hold and group changes.
One row per decision, one column per size. Scroll the table sideways on mobile.
| Solo / micro 1–5 | Small 5–20 | SMB 20–100 | Mid-market 100–500 | Enterprise 500+ | |
|---|---|---|---|---|---|
| Source of truth | Private git repo of markdown | Same, protected branches | Git canon plus web suggest-a-change | Domain canons; sensitive domains in own repos | Curated canon; the rest stays in source systems |
| How agents read | Load index note, open files | Hybrid search, then open file | Hybrid search per audience role | Canon via MCP gateway plus bought search | Identity-aware platform and gateway |
| Live data | One read-only MCP tool | Two or three, separate secrets | MCP gateway, end-user OAuth | Per-agent credentials | Plus human approval on risky writes |
| Who writes | The founder; agents propose | Owners per folder; agent PRs | Stewards; risk-based approval | Knowledge owner and stewards | Stewards and a council |
| Permissions | Separate public folder | Separate customer index | Groups, row-level security | SSO, SCIM, attribute-based | Zero-trust, policy-as-code |
| Eval set opinion | 10–50, by hand | 20–150 in CI | 50–400, by audience and language | Per domain, from traces | Continuous release gates |
| Infrastructure excl. seats, tokens | $0–35 / mo | $10–150 / mo | $100–1,000 / mo | $500–3,000 / mo | Contract-dependent |
| People opinion | 1–2 h a week | 2–6 h a week | 0.2–1 FTE | 0.5–3 FTE | 3–10 FTE |
Ranges merge the four reports' planning estimates and are opinion, not quotes. Seat prices that swamp the table if you choose them: Microsoft 365 Copilot Business $21 and Enterprise $30 a user a month, Notion Business $20, Cloudflare Access $7 after 50 free users.
One audience's canon passes about 30–50k tokens, agents open more than about five notes per answer, recall drops on paraphrased or Greek questions, or a second audience appears. The audience is the boundary, not the agent count.
The first customer-facing agent, or the first content some staff may not see.
More than about ten regular editors, repeated sync conflicts, or non-technical staff avoiding git.
Staff ask agents about contracts, project folders and mail that will never be curated. Buy suite search; do not ingest the mailbox.
The first agent that acts rather than answers, the first regulated domain, or an auditor asking where a prompt is processed.
If the owner is real and the tests pass, a new product announcement is not a reason to move. Grok's line, and we agree.
Triggers are opinion, but each one is observable. Thresholds come mainly from Claude Opus 5.5 and GPT-5.6.
The design under review has three layers: people write and approve in an Obsidian vault (markdown and git); PostgreSQL with pgvector, or a Cloudflare stack of R2, AI Search, Vectorize, D1, Durable Objects, AI Gateway and the Agents SDK, serves as a rebuildable index; live data is queried through MCP. Six rules sit on top: files are the source, structure by kind of knowledge, live data stays in place, permissions from day one, agents propose and humans approve, and 20–30 known-answer tests on every change. All four models called the shape right. Grok's summary fits: most of what is good is organisational, most of what is weak is a missing contract underneath a sound diagram.
"Never the reverse" is, in every report, the most important decision in the design.
Keeping upstream secrets in the MCP server, not the agent, is the correct trust boundary.
The only write path that matches the knowledge-poisoning threat, using a review tool people already know.
Directionally the whole quality programme in one rule.
"The customer chatbot has no connection to the confidential index" is the strongest security idea in the notes (Claude, GPT, Grok).
Adding the index later matches the evidence on agentic search and long-context decay (Claude, Grok).
With 25 questions, one answer is worth 4 points; Claude computes a 95% interval of about 61–91% around a measured 80%. Add refusals, leaks, injection fixtures and Greek variants, and grow the set.
R2, D1 and Durable Objects can be pinned to the EU. Vectorize documents no EU jurisdiction, and Cloudflare's own table marks Vectorize, AI Search and Workers AI incompatible with Regional Services. Use EU-region Postgres for EU-only clients.
No SSO, no live co-editing, a 20-collaborator cap with owner-level access for all. Non-technical staff need a web form that makes the same pull request.
Add Postgres full-text (with the greek stemmer and unaccent), rank fusion and filters. Vectorize has no keyword search; D1 is SQLite with a 10 GB cap.
Agent memory is unreviewed, per-user and often personal data. Keep it in a separate store, read-only for agents that read untrusted input; it reaches the canon only by pull request.
No owner, status, review_by or supersedes fields; no stale-note report. Maintenance is the failure the notes themselves predict.
Webhook sync means a bad change is live before the score drops. Build a staging index, run evals in the PR, then swap (Claude, GPT).
| Area | In the notes | Change | Raised by |
|---|---|---|---|
| File contract | "Entities, procedures, rules, decisions, FAQ" | CommonMark plus YAML with id, type, audience, owner, status, lang, aliases, review_by, supersedes; linted in CI | Grok, GPT, Claude |
| Retrieval | pgvector, or Vectorize plus D1 | Hybrid from the first index; Greek normalisation; rerank only if evals say so | Grok, GPT, Claude |
| Audiences | Permissions added as the company grows | Public/internal split mandatory at every size; audience-first directories that fail closed; sensitive domains in separate repos | Grok, Claude |
| Live data | "Never copied", read-only where possible | End-user OAuth on MCP; allow labelled caches with TTL for outages (GPT) | GPT, Claude |
| Evals | 20–30 known answers per change | Start there, grow to 100–300 stratified; score retrieval and answers separately; gate a blue/green index swap | All 4 |
| Residency | "Check before the DPA" | EU-region Postgres for EU-only clients; name every processor, including the generation model, in the DPA | Grok, GPT, Claude |
| Write path | One-click PR approval | Provenance in every PR, two approvers on pricing and policy, instruction-text checks, a non-git proposal form | All 4 |
| Answer contract | Not in the notes | Cite note IDs, abstain below a threshold, escalate; disclose AI to customers (EU AI Act Article 50, since 2 Aug 2026) | GPT, Claude |
The notes' path from files only, to files plus an index, to per-audience indexes with memory and logs, is sound in order. The models' main objection is timing: audience separation, the schema and hybrid search belong at the start, not at a later step. The "when you move on" line (past a few hundred pages, or a second agent) is reasonable; Claude would trigger on a second audience rather than a second agent.
| Approach | Wins when | Loses when | Verdict opinion |
|---|---|---|---|
| Suite copilots and enterprise searchCopilot, Glean, Notion AI | Staff ask "what did we say in our own tools?" across many sources; permissions already clean | Customer-facing bots; seat cost grows with humans, not agents; inherits oversharing | Buy beside the canon from about 20–100 people; never make it the brain |
| Graph RAGGraphRAG, LazyGraphRAG | Global and multi-hop questions over large corpora | Lookup questions; full builds are costly and lost to BM25 at scale in one 2026 study | Links and entity tables first; graph only where evals prove it |
| Long context"just load the files" | A clean handbook that fits, with steady traffic and caching | Noisy corpora; context rot; 150k uncached tokens is about $0.30 a question on Sonnet 5.5 | Always load a short guide; retrieve the rest |
| Hosted file search and memoryOpenAI, xAI, Gemini, Anthropic | Prototypes and public FAQ bots; $2.50 per 1,000 calls at OpenAI and xAI | As the system of record: vendor coupling, less control, separate residency | Fine as a rebuildable index or scratch pad, never the only copy |
| MCP knowledge server | Many agent runtimes need one read contract: search, fetch, list changes | Each server is its own permission and audit surface | The right interface for the canon; keep writes on the PR path |
Synthesis estimates from the four models' outlooks, scenario estimates, not forecasts of record. The through-line: vector search becomes a commodity, and identity, policy and review become the product.
We gave the same research prompt to four AI models on 14 October 2026: Grok 4.7, GPT-5.6, Claude Opus 5.5 and Gemini 3.1 Pro. Each wrote its own report on what to build at five company sizes and assessed the same reference architecture. Here is what each one said, where they agree and where they split.
Method: one prompt, run once per model on 14 Oct 2026. The prompt asked for a source link behind every number, a clear line between fact and opinion, and honesty about each model's own maker. We compared the four reports side by side and quote their positions as written, shortened. Where a model's claim conflicts with another's sourced figure, the sourced figure stands. Gemini 3.1 Pro's report is the shortest, lists its citations as titles without links, and names older models (Claude 3.5 Sonnet, Gemini 1.5 Pro) as examples.
| Question | Grok 4.7 | GPT-5.6 | Claude Opus 5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| Solo / micro | Git repo plus a "how to read" file; index only when it outgrows a careful read | Markdown or existing workspace; direct reading with a manifest | Load the folder with caching, or index note plus grep | Long context, a local MCP server, or Notion AI |
| Small | Hybrid search, 20–30 checks on every merge | Managed hybrid or Postgres FTS plus pgvector | Hybrid index per audience behind MCP | Simple vector RAG on Vectorize or pgvector |
| SMB | Postgres lexical plus vector, owners per domain | Retrieval as a service, ACLs before retrieval | Authorship and permissions become the bottleneck | PR-based Postgres pipeline, or Copilot |
| Mid-market | Curated brain plus separate employee search | Federate sources; buy or build | Buy search, build a smaller canon | Hybrid plus GraphRAG, or Glean |
| Enterprise | Curated brain plus procured search; no whole-company crawl | Enterprise search control plane, or build for sovereignty | Identity-aware platform, federated canons | Glean, Copilot or custom multi-agent |
| Eval set | Start at 20–30, grow per domain | 20–30 is a smoke test only | 100–300 stratified | 20–30 is "highly mature" |
| Biggest gap found | No written file contract | MCP is not authorization; 11 missing controls | EU residency fails for the index layer | Obsidian friction; no graph |
| On its own maker | xAI Collections priced like OpenAI's; "a component, not the source" | OpenAI file search fine for prototypes, keep the corpus outside | Discloses it is an Anthropic model; memory tool is not governed knowledge | Not addressed |
Positions as each model wrote them, shortened. Cost ranges are each model's own planning estimates.
Files as source, a rebuildable index, live data via MCP and human approval all survived every review.
Each model says brains die from neglect, not from bad search, and that tokens and people cost more than infrastructure.
"Just load the files" is right for a micro business and wrong as the plan for thousands of notes.
Grok, GPT and Claude independently found Vectorize lacks an EU jurisdiction. Gemini did not check.
bge-m3 lists Greek but publishes no Greek retrieval score; run real Greek, unaccented and Greeklish queries.
Gemini wants GraphRAG from mid-market. Grok cites a 2026 study where graph builds stayed below BM25 at scale; GPT and Claude say links and tables first.
Gemini calls 20–30 questions highly mature. Claude calls it statistically unable to detect regressions.
Three models endorse "never copy". GPT says that is too absolute: allow labelled caches with a TTL for outages and analytics.
Gemini warns full rebuilds get slow and wants incremental indexing; Grok and Claude want full rebuilds in CI against a staging index, then a swap.
Gemini and Claude quote $21 (Business); GPT and Grok quote $30 (Enterprise). Both are right for their plan.
The single, governed knowledge base a company's AI agents read from: procedures, rules, decisions, product facts and approved answers. Live data such as bookings and invoices stays in its own systems.
Usually not at first. Under about 20 people, markdown files in git that the agent reads directly are enough. Add a hybrid index when a second audience, a few hundred notes or failing tests appear.
As an editor for the people who curate, yes. The real source of truth is the markdown in git; non-technical staff need a simpler proposal form, and shared vaults cap at 20 collaborators.
Keyword search catches product codes, names and Greek inflections that embeddings miss. Anthropic measured a 49% cut in retrieval failures from adding BM25, and 67% with reranking.
Not fully, as of October 2026. R2, D1 and Durable Objects can be pinned to the EU; Vectorize, AI Search and Workers AI have no documented EU jurisdiction. Use EU-region Postgres for the index if you need EU-only.
For employee search across mail and documents at 100+ people, often yes. For customer-facing or rule-bound agents, keep a curated canon you own beside it.
bge-m3 embeddings cost about a cent per million tokens. The scarce input is a person who can say of one page: this is the rule, I own it, it is still true. Fund that person before a second vector store.
The incident that hurts is a public bot quoting a margin or a salary. A separate public folder, index and credential costs almost nothing at five people and is very expensive to retrofit at fifty.
Hosted file search is priced almost identically across OpenAI and xAI, which is what a commodity looks like. Teams with their truth in git re-embed and continue; teams with it in a vendor index migrate again. For a broader view of model choice, see Best AI models of 2026.
"The bucket is in the EU" is not the whole story: the index, the embedding call, the generation call and the logs each have their own location. Write every processor into the DPA, as the design notes already say to check.
Plan for a small, typed, git-backed canon, a rebuildable hybrid index per audience, live systems behind read-only MCP, and a growing test set that gates every change. Buy suite search beside it once employees, not agents, are the main askers.
Stress-test for the swings. A poisoned document that slips past a tired reviewer, a cross-audience leak, or a regulator asking where prompts are processed would each force a redesign. Cheaper million-token context could also push the "just load it" line far higher for clean handbooks.
The brain that lasts is not the one with the best index. It is the one where every page has an owner and every change has a test.
Anthropic 2024, Contextual Retrieval (35% / 49% / 67% fewer failures), study
arXiv 2607.26497 2026, file agents vs BM25 vs graphs at enterprise scale, paper
arXiv 2607.21274 2026, CUP Greek retrieval benchmark (hybrid 0.673 nDCG@9), paper
Chroma, Context Rot across 18 models, research
Liu et al. TACL 2024, Lost in the Middle, paper
arXiv 2501.01880, long context vs RAG revisited, paper
Microsoft Research, LazyGraphRAG cost and quality, blog
arXiv 2502.11371, RAG vs GraphRAG systematic evaluation, paper
OWASP LLM01 Prompt Injection, Top 10
OWASP LLM08 Vector and Embedding Weaknesses, Top 10
OWASP RAG Security Cheat Sheet, cheat sheet
MCP specification 2026-07-28, OAuth 2.1 authorization, spec
European Commission, AI Act Article 50 transparency FAQ, FAQ
European Commission, GDPR principles, guide
Gartner 2025, over 40% of agentic AI projects cancelled by 2027, press release
Microsoft, Copilot oversharing guidance, blog
Cloudflare, Data Localization Suite compatibility, table
Cloudflare, Vectorize overview and pricing, docs, pricing
Cloudflare, AI Search limits and pricing (billing from 1 Nov 2026), docs
Cloudflare, R2 data location and pricing, location, pricing
Cloudflare, D1 limits, docs; Workers pricing, docs
Cloudflare, Workers AI pricing (bge-m3 $0.012/M), docs
Cloudflare, AI Gateway pricing and logs, docs
Cloudflare, Access pricing, pricing
SQLite, FTS5 diacritics handling, docs
PostgreSQL, text search configurations incl. greek, docs
MTEB, BAAI/bge-m3 model card, card
Obsidian, pricing and Sync collaboration limits, pricing, docs
GitHub, Team plan, pricing
Supabase, Pro plan, pricing
Pinecone, Builder and Standard, pricing
LangSmith and Langfuse, tracing plans, LangSmith, Langfuse
OpenAI, API pricing incl. file search, pricing; business plans, pricing
xAI, Collections pricing, pricing
Anthropic, Claude API pricing, pricing; memory tool, docs
Google, Gemini File Search pricing, pricing
Microsoft, 365 Copilot Business ($21) and Enterprise ($30), business, enterprise
Notion, Business plan ($20), pricing
Glean, Enterprise Flex (no public price), docs; Vendr median contract, marketplace
First published 14 Oct 2026