Every company needs a digital AI brain.
The ones being built right now will not all be worth having.

In March, in Why We Exist, we described the hole in the AI stack: extraordinary models with no institutional memory underneath them. In July, in Every Company Will Need a Second Brain, we argued that the thing filling that hole is inevitable, and that a company without one pays for it every day.

We stand by both. Having no digital AI brain is the worst position a company can be in. Every AI deployment starts from zero, and every correction your people make flows to a model provider instead of into something you own.

But the conversation has moved. A lot of companies are no longer asking whether they need a digital AI brain. They are deciding whether to build one or buy one, and many have already started. That is the right decision. It also raises a harder question, one that neither of our first two articles answered: once you have a digital AI brain, how do you know it works?

Connected is not the same as working

Most digital AI brain projects start the same way. Connect the systems, index the documents, build a graph, put an agent on top. Within weeks there is a demo that answers questions about the company, and it looks like the job is done.

It isn't. A digital AI brain that has ingested your data knows what your systems say. It does not know which of those things are true, which are outdated, and which contradict each other. It will answer anyway, fluently and with confidence, because that is what models do.

The test of a digital AI brain is not whether it answers. It is whether the people who run the business are willing to act on the answer without checking it somewhere else first. If they check every time, you have built a slower search engine. If they act and the answer was wrong, you have built something more expensive than that.

So the standard is simple to state and hard to meet: a digital AI brain works when the people who know the business would sign off on what it says.

Ask three plants for their OEE

Established manufacturing players and business with physical operations show this problem more clearly than any other industry, which is one reason we work there.

Take OEE, overall equipment effectiveness. On paper it is one formula: availability times performance times quality. In practice, ask three plants in the same company for their OEE and you get three numbers that cannot be compared. One plant excludes planned maintenance from available time, another counts it. One treats changeovers as downtime, another as scheduled production. One measures performance against the machine's nameplate speed, another against the best rate the line has ever held.

None of these plants is wrong. Each definition was a reasonable local decision made years ago by someone who understood that line. The people in the building know which definition they use, and they know how to read the number.

A digital AI brain connected to all three plants does not know any of this. Asked which plant performs best, it compares the three numbers and names a winner. The answer is well-formatted, sourced from real systems, and meaningless. Worse, nothing about it looks wrong.

OEE is one example. The same thing is true of scrap rate, on-time delivery, first-pass yield, and cost per unit. Every number a manufacturer steers by carries a definition that lives outside the system that stores it.

The knowledge no connector can reach

The definition problem has a twin. A large share of what makes a plant run was never written down at all.

The process engineer who has run line 4 for twenty-five years knows that it has to be slowed down on humid days, that one supplier's material needs a different temperature profile, and that a certain alarm can be ignored unless it comes with a second one. None of that is in the MES, the ERP, or the maintenance log. It is in one person's head, and that person is closer to retirement than most companies like to think about.

A digital AI brain built only from system data inherits this gap in full. It can describe what the line did. It cannot explain why, or what to do when something changes.

The only way to get this knowledge is to ask the people who hold it. That is why interviews are a core part of how we build a brain: agents that sit down with the experts, ask the questions a sharp new colleague would ask, follow up where the answer is vague, and turn the result into structured knowledge that sits alongside the system data. Not a survey and not a documentation project nobody finishes, but a conversation that starts from what the brain already knows and looks for what is missing.

Captured is not the same as trusted

Getting knowledge into the brain, from systems and from people, is half the work. The other half is making it something an agent or a manager can rely on. In our experience, three things separate a fact the brain has stored from a fact the business will act on.

Provenance. Where does this come from? Which system, which document, which person, and when? A scrap rate without a source is a rumor. A scrap rate that traces to a specific report, a specific line, and a specific month can be checked, challenged, and corrected.

An agreed definition. What exactly does this number or term mean here, and who decided that? If two plants calculate OEE differently, the brain has to hold both definitions, know which applies where, and know whether the company has settled on one for group reporting. A definition nobody has signed off on is still an open question.

Context. When does this apply, and when does it not? The slower line speed is correct on humid days and wrong on dry ones. The discount rule holds for one customer group and not another. A fact stripped of the conditions it depends on gets applied where it does not belong.

A brain that stores facts without these three things will produce answers. A brain that stores them with all three produces answers someone is prepared to stand behind.

Someone has to sign

Provenance, definitions, and context do not appear on their own. A model can propose them. It cannot decide them, because deciding what a KPI means or which of two conflicting sources is right is a business decision, and it belongs to the people accountable for the business.

We use two words for the work this takes.

Ratification is the act of confirming. The brain surfaces a definition, a relationship, or a piece of expert knowledge, and the person who owns that topic approves it. From then on it is not something the brain inferred. It is something the company has agreed on, with a name and a date attached.

Rectification is the act of correcting. When the brain is wrong, outdated, or holding two versions of the same truth, the owner fixes it once, in one place, and the fix reaches every agent and every workflow that draws on it.

This puts humans in the loop as gatekeepers, and that is deliberate. It is tempting to see human review as a brake on automation. We see it the other way round. A brain nobody has ratified cannot be given real work, because nobody will take responsibility for what it does. A brain whose core knowledge has been signed off by the people who own it can.

It is also what makes the compounding we wrote about in July safe. A digital AI brain gets more valuable with every interaction only if what accumulates is correct. Without a gate, errors compound just as reliably as knowledge does.

Five questions for any digital AI brain, built or bought

Whether you are evaluating a vendor, including us, or reviewing what your own team has built, the same questions tell you whether the brain will hold up once people depend on it.

1. Can it show where an answer comes from? Not a list of similar documents, but the specific source, owner, and date behind each fact it used.

2. What does it do when two sources disagree? If it quietly picks one, you will not find out until the wrong one wins. It should surface the conflict and route it to someone who can settle it.

3. Who has approved the definitions it uses? For your ten most important KPIs, there should be a named person who confirmed what each one means.

4. How does it get what is not in any system? If the answer is "upload more documents," the most valuable knowledge in your company stays out of reach.

5. When someone corrects it, where does the correction go? One fix should update every agent and workflow. If each tool has to be corrected separately, the brain will drift apart within months.

How Along AI plays a role

We build digital AI brains for established companies, with a focus on manufacturers and business in physical operations. The brain combines two inputs: the data in your systems, and the knowledge of your experts, gathered through agentic interviews.

What we put most of our effort into is the layer between capturing that knowledge and using it. Every definition, KPI, and piece of expert knowledge carries its provenance, its agreed meaning, and the context it applies in. Your people ratify what is right and rectify what is not, and they stay the gatekeepers of what the brain treats as true.

Every company needs a digital AI brain, and the companies starting now are right to start. The ones that get the most from it will be those whose brain has earned the trust of the people it works for.