An AI-native company gives every agent the same current picture of the business. Routine work runs as automations people review, and approved work becomes the next person's starting point. A company brain provides the foundation. The automations are what you build on it.
At a consultancy, that might mean a meeting brief that knows what the last proposal promised, followed by a new proposal that starts from the team's approved research. The useful change is in how work passes between people.
This guide builds on What is a company brain? with a map of twenty automations, a recipe for building each one and a rollout sequence. It also asks who maintains the context and whether you can move it to another model.
The distinction matters even where AI adoption is high. Wharton Human-AI Research and GBK Collective's October 2026 study surveyed 1,416 US enterprise leaders and employees. Among leaders, 80% reported positive AI returns, but only 28% described them as strongly positive. Stronger returns were associated with workflow redesign, ownership and organisational alignment. That is survey evidence about how companies operate, rather than a test of company brains.
The Definition
What AI-native means
A useful definition has four tests.
Every agent reads the same context. Sales in ChatGPT, delivery in Claude and engineering in a coding agent read one shared, permissioned picture of customers, people, projects and decisions. They draw on the same underlying record. Their access determines which evidence they may use.
Routine work runs as automations people review. Briefs arrive before meetings. Forwarded documents come back processed. Conflicts are flagged while there is time to resolve them. People spend less time gathering and more time checking and deciding.
Work compounds. Finished analysis, decisions and approved drafts return to the brain with their evidence. The next person can reuse the work, correct it or extend it.
The context is portable. Company knowledge lives in a form the company can query and export. Changing models should preserve that knowledge, its permissions and its history. Each automation still needs testing against the new model.
The Failure Modes
Why automations fail without a shared brain
Someone builds a prompt for board packs. Someone else connects an agent to the CRM. A third team buys a meeting-notes tool. Each helps with a task, but the pilots may keep separate copies of the data and different accounts of who the customers are.
The pilots do not share what they learn. The meeting-notes tool records a promise on a call. The proposal agent never sees it. Twenty pilots can become twenty places to reconcile.
Every automation rebuilds context. Each run gathers documents and reconstructs the customer history before doing useful work. One fintech reported spending $1.5 million on tokens in one week in May 2026. Its account does not quantify how much went on repeated context assembly. That share needs measuring in your own workflows. The $10,000 analyst sets out an audit for finding it.
Context drifts. A company changes every week. An automation can keep applying a rule learned last quarter after the team has dropped it. Without dates and an evidence trail, people struggle to establish why the output is wrong.
Adoption stalls. People need useful outputs where they already work, with sources they can inspect. A separate workspace and unexplained answers make every automation harder to trust.
A shared brain gives these automations a maintained record to read and extend. It joins the promise from the call to the project, preserves the source behind it and records when a later decision replaces it. That foundation still needs owners, monitoring and good retrieval.
The Foundation
The foundation, and whether to build it
The automations below assume five capabilities.
- Sources joined around entities. Email, meetings, chat, documents, tickets and the CRM link around the same customers, people, projects and decisions.
- Permissions on every passage. Knowledge retains its source access rules. The trusted retrieval layer checks who the automation acts for through every query route.
- Dates on changing facts. A replacement retains the earlier version, the source and when the change applied.
- A shared interface. Agents and automations reach the maintained context through an interface such as MCP. The server enforces access; the model does not decide what it may read.
- Evidence-linked write-back. Approved work returns with its sources, owner and change history.
Should you build the brain or buy it? We sell one, so weigh what follows with that in mind. A folder of notes and a retrieval script can give a founder a useful memory. Serving a company adds permissions, identity matching, freshness, audit trails and export. Each needs someone responsible for keeping it working.
- Who maintains it in a year? If the answer is one engineer, document how someone else can operate it.
- Can it enforce access when the asker is a model? Test ordinary user accounts, shared credentials and every retrieval route. Instructions in a prompt are insufficient.
- Does it know what changed? Ask about a price or decision that was replaced, including what was true before the change.
- Can you take it with you? Export the evidence, relationships, dates and access rules in a usable form, and test querying them independently.
If you can answer all four for an internal build, build. Hold a vendor to the same questions. What you own when you own your AI explains the ownership test.
The Automation Map
Twenty automations on one brain
The map has five layers. Every layer reads the brain. Approved outputs can return to it under the write-back rules you agree.
Briefs come first because everyone sees them. A daily brief or drafted stand-up shows within days whether the brain has the right context. One client judges its pilot by how much people edit their drafted stand-ups before sending them.
Intake turns inbound material into structured knowledge. An analyst can forward a banker deck or supplier list, then review the extracted companies and their matches to existing records. A transcript can produce proposed decisions and commitments attached to the right project.
Monitors watch for change. A sales commitment, scope document and shipped change all attach to one project, so a conflict monitor can flag a mismatch. Sold, scoped, shipped follows that handover.
Agents that work draft proposals, onboarding briefs and memos from relevant precedents. Sources accompany factual claims, and a person reviews the result. Signed, not live examines where this can help with onboarding.
Write-back hooks preserve approved work. A merge hook can capture the design
decisions behind a change, with who made them and when. An instruction in a shared
AGENTS.md can describe that workflow, but the hook still needs an implementation,
access rules and checks. Capture the approved decision summary and its evidence rather than an
unreviewed session log.
| Automation | Trigger | Reads | Produces | Reviewed by |
|---|---|---|---|---|
| 1. Daily brief | Every morning | Changes in the person's accounts and projects | A short brief in Slack or email | The reader |
| 2. Stand-up draft | Every morning | Tickets, meetings, chat, calendar and agent work | A stand-up draft per person | Its owner, before posting |
| 3. Meeting brief | Calendar invite | Past interactions with attendees and their company | Context and three questions to ask | The attendee |
| 4. Weekly digest | Friday | Decisions, risks and changes across a team | A one-page summary | The team lead |
| 5. New-joiner brief | Start date | Projects, decisions and people around the role | A reading list with sources | The manager |
| 6. Document intake | Email to an agreed address | A forwarded deck, list or contract | Extracted items matched to records, with a reply | The sender |
| 7. Meeting to decisions | A transcript lands | The transcript and its project | Proposed decisions, commitments and actions | The meeting owner |
| 8. Questionnaire answers | An RFP or questionnaire arrives | Past answers, policies and contracts | A first draft with a source per answer | Each section owner |
| 9. Ticket context | A ticket is created | The thread, call or request behind it | The ticket linked to its reason | The assignee |
| 10. Conflict monitor | A commitment or product change | Promises, scope and shipped changes on the project | A mismatch flag to both sides | The account lead |
| 11. Freshness monitor | Daily | Each connector's last successful sync | An alert for a stale source | The brain's owner |
| 12. Watch list | Daily | Signals on tracked companies and people | A note when a threshold or role changes | The relationship owner |
| 13. Renewal and risk | Ahead of a renewal | Usage, tickets, sentiment and commitments | A risk note and suggested plan | The account lead |
| 14. Proposal draft | A new opportunity | Comparable projects, pricing and outcomes | A cited first draft | The partner or seller |
| 15. Onboarding brief | A contract is signed | The most similar past customers | Likely exceptions, risks, entities and contacts | The delivery lead |
| 16. Report and memo drafts | A template and subject | Relevant passages and past good examples | A draft with evidence for factual claims | The author |
| 17. Introduction draft | A target is added | Relationships and routes to the target | A route and draft note to the connector | The asker, then the connector |
| 18. Code factory | An agreed request | The request, current decisions and tests | A tested change in a preview environment | An engineer, before merge |
| 19. Merge hook | A pull request merges | Approved design decisions from the coding session | A decision record with who and when | Agreed policy and gated edits |
| 20. Finding write-back | An output is approved | The output and supporting sources | A finding the next person can reuse | The output's approver |
The Recipe
A recipe for each automation
Write down these eight steps before building.
- Trigger. Choose a schedule, calendar event, incoming email, new record or merge.
- Who it acts for. Use a named person's access or a company credential scoped to the task. Agree who may receive the output as well as what the automation may read.
- The context query. Define the entity and relationships, current facts, source filters and dates it needs. Start with excerpts, then fetch the supporting passages.
- The skill. Store the model instructions with their examples, source material, owner and version history. A skill can be an entity in the brain, linked to the evidence behind its rules.
- The output. Put it in Slack, email, the CRM or the document editor the team uses. Label drafts and separate facts from suggestions.
- Human review. Name the approver. Show the output, its sources and any changes it would make.
- Write-back. Specify what returns after approval, who owns it and which evidence supports it. Keep proposed findings separate from approved records.
- A test set. Use ten to thirty past examples with known good answers. Include missing evidence, changed facts and restricted sources.
Keep the model's job narrow. Fetching, matching and formatting often run on code or smaller models. Use a more capable model where the task needs judgement, and measure the difference. When evidence is missing, the automation should explain the gap and distinguish a nearby example from an answer.
The Rollout
The order to roll them out
Twenty automations make a useful map and a poor first sprint. Use the following sequence as a starting plan, adapting it to source access, risk and the team's capacity.
Weeks 1 and 2: the brain and one use case. Pick the question that costs the team most to answer. Connect the sources behind it with the narrowest useful access. Six months of history can be a starting point; choose the period the question needs. Turn on a daily brief for one team.
Weeks 3 and 4: more briefs. Add meeting briefs and stand-up drafts. Track edits. Heavy rewrites help identify missing sources, poor matching or instructions that need changing.
Month 2: intake and monitors. Add document intake where inbound volume is highest, then meeting-to-decisions for agreed recordings. Introduce conflict and freshness monitors. Expand department by department as the checks pass.
Month 3: agents that work and approved write-back. Add proposal and onboarding drafts, then merge hooks and finding write-back. Keep generated findings out of the trusted shared record until the approval and quality checks are ready. Connected source ingestion continues throughout.
After that, add one automation at a time. Each gets an owner, a test set and a business outcome. Retire it when it stops earning its place.
Connect the assistant each team already uses before asking people to change how they work. Give the first team a visible output it can judge.
Maintenance
Keep the context current
An automation built on last quarter's understanding can keep producing confident output after the business has moved on. Five practices help catch that drift.
Set freshness per source and measure it. Agree a refresh interval for each connector and installation. A priority channel may need updates within minutes; an archive may refresh overnight. Track both the last successful sync and the age of evidence reaching answers. A failed schedule can otherwise leave the system answering from old material.
Let newer facts replace older ones. Keep the date and source of each replacement. Current questions should retrieve the current version, while historical questions can still establish what was true in March.
Version skills and attach their reasons. A rule learned from past behaviour goes stale too. Link instructions to the examples and sources behind them, so a reviewer can find the skills affected when the business changes.
Run a daily health check. Exercise the tools, check source freshness and ask questions with known answers. One pilot's first health check found three stale sources whose schedules had never triggered. Monitoring the data path matters as much as monitoring whether the application is up.
Leave out the noise. Thousands of automated alerts can crowd useful evidence out of retrieval. In one pilot, removing two noisy channels helped answer quality more than adding sources. Test that decision on your own questions.
Adoption
Make people use it
The automations pay back when people use their outputs. Four choices help.
Arrive where people work. Put briefs in Slack, intake in email and drafts in the document editor. Each additional login adds friction.
Be right and show why. Cite evidence for factual claims. When an answer is wrong, the reader should be able to find the passage behind it and submit a correction. Review that correction before it becomes shared knowledge.
Keep the person in charge. An automation can draft an introduction. A person decides whether to send it. Begin with explicit approval for consequential actions, then agree which routine actions can run under a tested policy.
Lead from the top. Use the digest in the leadership meeting. Leaders then encounter the same missing context and quality problems as the team, and can help resolve them.
The brain keeps a team aligned to a decision once someone has made it. It cannot make the team agree. Its job is to help people reach the discussion with the same evidence.
Ownership
Keep it portable
Choose a deployment you control, or a host that gives you a usable export. Use an interface such as MCP across compatible agents. Keep changing knowledge with its evidence outside model weights.
Then test the arrangement. Change the model behind one automation and run its evaluation set. Check tool support, permissions, citations and output quality. Ask for a complete export and query it independently. The interface alone cannot prove portability.
We run open-weight models alongside company brains in clients' clouds, and our agent workspace can use them too. Teams can choose those models for routine work and call a frontier model for tasks that need it. Our permissions model explains how shared company knowledge stays separate from private personal context.
Measurement
Measure what changes
Measure in the order trust is earned.
| Measure | What to track | What good looks like |
|---|---|---|
| Freshness | Time from a source change to its appearance in answers, per source | Each source stays within its agreed lag |
| Accuracy | Each automation scored against its test set | An agreed pass rate held across repeated runs |
| Permissions | Access tests on every release and retrieval route | Zero unauthorised disclosures in the tests; investigate every failure |
| Use | Draft edit rates and weekly active users per automation | People use the output with light edits |
| Outcome | One business result per automation and cost per finished piece of work | A useful change against the baseline, including running costs |
Choose the outcome the automation exists for: duplicated work, meetings spent establishing facts, days from signature to go-live, or cost per finished memo. Count useful results. A system judged only on how many briefs it sends has an incentive to send more briefs.
The Limits
What does not change
Judgement stays with people. Two partners can read the same evidence and disagree. The brain can preserve their reasoning and the decision they make.
External delays remain. Security reviews, bank testing and contract negotiations have their own dependencies. Measure those separately from the time spent gathering knowledge.
Unrecorded knowledge remains missing. Transcripts can capture reasoning that never reached a document, but they do not capture everything. Interviews and human judgement still matter.
Capture changes behaviour. People may move sensitive conversations if they believe everything will be shared. Explain access and retention before connecting personal sources. Company search must respect the boundary around private personal context.
A First Use Case
Where to start
Pick the question your team spends most time answering and the automation that could answer it each morning. Build the brain around that question. Turn on the brief, watch the edits and fix the gaps before adding the next automation.
We install a company brain for an agreed first use case in two weeks. Choose 60x Cloud, hosted by us with EU or US data residency, or Your Cloud, inside your Azure, AWS or GCP account. Both can run on open-weight models. We agree the initial sources, users and automation scope before installation.
Book a demo and bring a job your team repeats every week. We will work through the sources, review step and outcome a useful automation needs.
Sources
Sources
- Wharton Human-AI Research and GBK Collective, The Adaptive Advantage: Who's Winning with AI and How, 13 October 2026. The primary study announcement reports the survey sample and return figures.
- Ramp, Introducing Glass: Learn things once, 13 October 2026. The token-spend figure is Ramp's account of one week in May; it does not establish the share spent rebuilding context.
- 60x, What is a company brain?, A company brain that knows who is asking, The $10,000 analyst, Sold, scoped, shipped and Signed, not live.
Client examples are anonymised. The automation map, meeting-brief example and ninety-day sequence are planning guides. Source access, refresh intervals, review rules and implementation scope are agreed for each installation.