Why enterprise AI cannot trust enterprise data.

AI does not have a reasoning problem. Enterprise AI has a reality problem. A frontier model will reason beautifully from three contradictory inventory counts, because nothing in its context tells it they are contradictory. It will draft the memo, size the reorder, and recommend the write-down, fluently, from a version of your company that does not exist. The model is not hallucinating. It is faithfully reasoning from an estate that never agreed with itself in the first place.

Here is the part most AI conversations skip: the systems are not wrong. Your ERP is correct about what was posted to it. Your warehouse system is correct about what was scanned. Procurement is correct about what was received against the PO. Each one is a truthful witness to its own records. The problem is that nobody in the enterprise is responsible for whether the witnesses agree, so the question is settled the old way: an export, a spreadsheet, a meeting, and whoever argues best. That process was already expensive when humans were the only readers. Now we are about to hand the same contradictory estate to software that acts on it at machine speed.

Point an agent at systems that disagree, and you have not automated the work. You have automated the argument.

I learned this selling light switches.

My first company made smart-home products. The insight that made it work was not a better device; it was killing fragmentation. Dozens of products, one interface, and revenue grew triple digits not because any individual switch improved but because the customer stopped managing forty apps that each told the truth about one corner of the house. Later, working in procurement and supply chain, I watched enterprises live the same disease with more zeros: forty systems, each truthful about its corner, and one heroic analyst with a spreadsheet playing referee between them, quarterly, after the money had already moved.

The pattern is identical. Fragmentation is never priced where it occurs. It is priced downstream, as working capital nobody can locate, safety stock sized to a number the floor cannot ship, cash paid against an over-receipt, and audit weeks spent refereeing witnesses. We publish one of these anatomies every week, with the arithmetic shown, because the fastest way to make a category intuitive is to make the cost visible.

Why the incumbents cannot fix this, structurally.

Every vendor in your stack will eventually ship an AI that explains its own data, and every one of those AIs will be useful and none of them will solve this, for a reason that has nothing to do with model quality: a system of record reconciling its own figures is grading its own exam. SAP cannot referee a disagreement between SAP and the warehouse; it is a party to the dispute. Neither can the warehouse. The referee has to stand outside every system it reads, authoritative over none of them, or the answer is marketing.

That is an independence argument, and independence is structural, not a feature. It cannot be added to an ERP in a point release, because the moment the referee owns one of the ledgers, it stops being a referee. This is also why I built the layer read-only, with no write method in the codebase at all: the referee that can rewrite the evidence is not one either.

What the trust layer actually is.

The stack that makes agents safe to point at an enterprise is boring, and the boredom is the point. Read-only connections to the systems you already run. One canonical state, where a disagreement is preserved with both values rather than silently resolved. Provenance on every figure: source, timestamp, the dated record behind it. Then, and only then, the AI, whichever provider you choose, receiving figures that carry labels: observed, derived, inferred, assumed, scenario. A model may reason across all of them; it may never promote one to another. Its output is a recommendation, which waits for a named human, whose decision is recorded, and execution returns to the systems of record, where it always lived. The whole stack is drawn, tier by tier, with the status of each stated honestly.

Model intelligence is becoming a commodity. Trusted enterprise context is not, and it compounds: every reconciled disagreement teaches the layer which patterns of divergence correspond to which economic problems. The models get cheaper every quarter. The corpus gets more valuable every customer.

The category, sized honestly.

People will file this company under enterprise data reconciliation. If the architecture holds, that label is substantially too small. Reconciliation is the wedge; the company is the layer between enterprise reality and machine intelligence:

SYSTEMS OF RECORD
↓
AGREEMENT / CONTEXT  (EOS)
↓
MODELS / AGENTS
↓
DECISIONS
↓
ACTIONS

The strongest version of this company is not better reconciliation. It is the trusted contextual substrate through which an enterprise increasingly understands itself. Every connector potentially makes the layer more useful. Every reconciled relationship enriches the graph. Every workflow adds context. Every interaction improves it. Every department widens the visibility. Context, then intelligence, then agents, then action: that flywheel is the company, and every wheel of it starts turning at the pilot, which is why the wedge ships first, and the word potentially holds these sentences to the standard the record will meet.

Where I actually am, stated plainly.

I hold myself to the standard I am selling, so here is my own estate, provenance attached. The engineering is real and independently reproduced. I published the whole arc in advance, the finished stages and the ones still ahead, with the definitions of finding, validated and captured committed ahead of the first result, so that when it lands, nobody has to take my word for what it means.

Those first two numbers used to be my opening line. They are not anymore, because verification is not the product; the found dollars will be. Verification exists so that when the dollar figure arrives, it arrives pre-trusted. The page where that figure will land is public today, empty, dated, and waiting: the record.

The bet, in one paragraph.

Every enterprise is about to put agents above its systems. Every one of those agents needs a single, provenance-carrying version of enterprise reality, and that layer cannot belong to any vendor whose figures it referees. Someone independent builds it. I think the founder who does should have spent years inside procurement's spreadsheets, should have already made one business work by killing fragmentation, and should be willing to prove everything in public before asking anyone to believe it. That is the bet I am making with my own time and my own balance sheet, and the scoreboard is a public page.

The model is not the advantage

Every buyer will soon have access to the same frontier models, which means the model itself stops being the differentiator. What differs between enterprises is what sits underneath: whether the figures the model reasons over agree with each other, and whether anyone can say where each one came from. A language model will produce a beautifully reasoned answer from contradictory inputs; the reasoning launders the disagreement instead of surfacing it, which is arguably worse than no answer at all.

That is the whole proposition, stated once more: establish decision-grade agreement across conflicting systems, with provenance and context, before downstream intelligence acts. The disagreement shows up everywhere the estate grew faster than its master data: procurement's numbers against finance's, the plan against what operations actually ran, two ERP systems still running side by side after an acquisition, and the same supplier, customer, SKU, or contract existing in several versions that no longer agree. The next advantage is not another model; it is the quality and agreement of the data underneath the model.