Late October into early November 2026, a blunt framing of enterprise AI kept circulating on X: the models are already smart enough for a lot of office work, but companies still cannot automate whole functions because the real operating manual lives in people's heads. a16z's post quoting Databricks CEO Ali Ghodsi put it plainly — frontier progress is not the binding constraint if agents lack internal context from meetings, processes and key employees.
That sounds like a Fortune 500 problem. It is also an 18-person plumbing company's problem. Rohan Shah's reply called out the need to map the business into a reliable context layer. Shereef Mohammed put the operator version: founders winning with agents are not running the most agents — they are writing the most down.
TL;DR
- a16z amplified Databricks CEO Ali Ghodsi's point that enterprise AI is bottlenecked on company context, not on waiting for smarter frontier models.
- Owner-led service firms feel the same gap: exceptions, customer quirks and Friday finance habits live in heads and Slack, not in clean fields.
- Before you add agents, write the tribal knowledge into short cards an agent and a new hire can both follow.
- AI drafts, your team approves — and the approval rules must be written down too.
What company context means on a real job site
In a service business, context is not a data lake. It is knowing that Mrs. Alvarez always wants the crew to call before entering the side gate, that manufacturer SKU 4481 is sold under three names in your catalog, that weekend emergency rates do not apply to existing maintenance-plan customers, and that the office never texts estimates after 8 p.m. even if the CRM allows it.
None of that is in the model weights. Some of it is in a dusty SOP PDF. Most of it is in the office manager's muscle memory.
When you drop an agent onto CRM notes or inbox triage without those rules, you do not get a junior employee. You get a fast intern who never attended the Monday standup.
Why smarter models will not fix a missing playbook
Teams keep buying the next model the way they used to buy the next phone. Capability climbs. Reliability on your exceptions does not, until someone encodes the exceptions.
That is why "we tried ChatGPT and it was wrong about our pricing" is usually a context failure, not proof that AI is useless. The tool answered a generic question. Your firm does not sell generic work.
The same pattern shows up when agents meet undocumented deploy steps in software firms. Coding agents hit tribal-knowledge walls on day one. Your quoting agent will hit the same wall on day one if discount rules exist only as hallway knowledge.
A context pack beats a tool bakeoff
Before you compare vendors, write a one-page pack for a single workflow:
- Trigger: what starts the work (form, phone note, email label).
- Happy path: the boring steps when nothing is weird.
- Top ten exceptions: the cases that create rework today.
- Systems touched: inbox, CRM, estimating sheet, accounting.
- Approval gates: what a human must sign off before send, schedule, or charge.
- Definitions: what urgent, warranty, and VIP mean here.
If you cannot finish that page, you are not ready for an agent. You are ready for a documentation afternoon.
Table: model upgrade vs context upgrade
| Move | Typical cost | What it improves | What it ignores |
|---|---|---|---|
| Jump to a newer frontier model | Seat or API bump | Fluency, longer documents | Your unpublished exceptions |
| Clean CRM fields for one pipeline | Staff time | Retrieval quality | Still needs exception cards |
| Write a 10-exception context pack | Half a day | Agent reliability on real jobs | Does not replace approval |
| Add a second agent without docs | Tool sprawl | Demo energy | Error rate under load |
For most shops under 75 people, the third row wins first.
How to harvest tribal knowledge without a binders project
Sit with the person who currently does the job. Ask them to narrate while they work. Every time they say "except when," write a card.
Limit the first harvest to one workflow. Inbox triage for service requests is a good start. So is turning a completed job into a draft invoice.
Store the cards where both humans and tools can read them: a short Notion page, a repo markdown file, or a Google Doc with a clear owner. Review the cards monthly. Stale context is how agents earn a bad reputation.
Keep humans on the consequential steps
Context packs are not a license to remove judgment. They are how you tell the agent where judgment starts.
Money movement, legal language, medical or financial advice flavored wording, and any promise that creates a schedule commitment should stay behind an approval button. The brand line still holds: AI drafts, your team approves.
If your context pack has no approval gates, it is incomplete.
Verdict: Buy context work before another model seat
Verdict: Buy a written context pack and exception list for one workflow. Verdict: Hold on multi-agent rollouts until that pack exists. Verdict: Skip "just dump Slack into the model" as a strategy.
The X thread around Ghodsi's comments is useful because it names the boring bottleneck. Service businesses win the same way software platforms do: by making the invisible rules visible before asking software to act.
Interview prompts that surface exceptions fast
Ask the current owner of the workflow: What do you always double-check? Who gets a special rate? Which customers require a phone call instead of email? What broke last month that the SOP did not cover? What would you tell a smart temp on day one?
Record answers as cards, not essays. One exception per card. Tag each card money, safety, reputation, or convenience. Automate convenience last.
Where context should live
Pick one system of record for the pack. Duplicating the same exception in Notion, Google Docs and a pinned Slack message guarantees drift. Link out from the agent config to that single page. Review ownership monthly the way you review bank reconciliations.
Testing context quality with a new hire drill
Hand the pack to someone who does not know the workflow. Ask them to process three sample jobs using only the pack. Every question they ask becomes a missing card. This drill is cheaper than discovering the gap through a customer complaint after an agent ships a bad draft.
A day-in-the-life map before any agent install
Follow the ticket from first signal to closed job. Write every system touch: phone note, SMS, CRM stage change, estimate sheet, purchase order, invoice. Circle the steps that require judgment. Those circles are approval gates. Everything else is a candidate for drafts.
Owners who skip this map buy tools that automate the wrong slice. The map also shows where context is missing because you will find steps nobody documented.
Metrics that prove context work paid off
Track edit rate on drafts, time-to-approve, and repeat failure codes week over week. If edit rate does not fall after the context pack ships, the pack is incomplete or unread. Do not blame the model first. Sit with the reviewer and watch them correct three drafts.
Field note
Run the pilot with a named owner, a written kill switch, and a Friday review. If those three artifacts do not exist, pause new tools and build them first. AI drafts, your team approves only works when the queue has a human name on it.
For this workflow around AI agents company context, write the kill switch criteria in numbers: edit rate, error count, or hours saved. Numbers beat vibes when you decide to keep or cut the pilot.
FAQ
Want help turning tribal knowledge into agent-ready context?
In a free 30-minute assessment we map one workflow's hidden exceptions and show which pieces must stay human before any agent touches them.
What is company context for AI agents?
It is the unwritten know-how of how your firm actually runs: who gets weekend rates, which customers always pay late, which SKU names are wrong in the catalog, and which messages need a person.
Did Databricks' CEO say models are not smart enough?
The point amplified on X is the opposite: frontier models are capable enough for much of the work, but enterprises have not put proprietary context where agents can use it reliably.
How do we capture tribal knowledge without a year-long documentation project?
Shadow one workflow for a week. Write exception cards as you go. Stop at the cards that change money, safety, or reputation. That set is enough to start.
Can Slack and Notion replace a context layer?
Only if the useful bits are structured and current. Dumping every channel into a model usually produces confident nonsense on old threads.
Where should humans stay in the loop?
Anything that spends money, changes a contract, promises a date to a customer, or touches regulated data should require approval. AI drafts, your team approves.
Is this only an enterprise problem?
No. A 12-person shop has denser tribal knowledge per person. Agents feel that missing context immediately.
What is the first deliverable?
A one-page context pack for a single workflow: triggers, happy path, top ten exceptions, systems touched, and approval gates.