The moat isn't the model.

01

Why enterprise AI feels slow.

The model is not slow. The substrate underneath it is. This is the part nobody wants to write a deck about because the deck does not photograph well.

When an operator says "the AI is not delivering," what they usually mean is something quieter. A salesperson who asked for the cohort report on Tuesday is still waiting on Thursday. A pricing analyst noticed a leak on a Friday but the recommendation lands at the next Wednesday review. A claims handler escalated a fraud signal at 2pm; the policy team sees it on a 9am dashboard the following day. The model fires in two seconds. The decision still takes a week. Nobody on the front line confuses the two.

This is decision compression. The time between a business noticing a revenue signal and shipping the decision that responds to it. Compress it and the business gets faster. Stretch it and the business loses ground to the one that compressed first.

The compression problem is not solved at the model. It is solved at the layer beneath the model. Call it the substrate. The shared state. The ontology. Whatever the operator calls it on a Tuesday is fine. What matters is whether one canonical view of what a customer is, what a deal is, what a journey is exists in a place the agent and the operator both read.

Without that substrate, every AI deployment is a renovation on rented land. The model gets smarter. The decisions still take a week. The dashboard still says different things to different people. The agent drafts a beautiful recommendation that references a customer ID that does not exist in the system the operator opens to approve it.

This is the part most consulting decks skip. The interesting part is not "look how smart the model is." The interesting part is "look how fast the shared layer makes the operator." A 4-billion parameter model in a five-data-source mess will lose every time to a 70-billion parameter model in a one-ontology shop. The gap is widening as the models get cheaper. The substrate is where the moat moves.

02

Five sectors, one shared layer.

Five sectors where this is being decided right now. The names of the systems change. The pattern does not.

Property. The deal-decision lag.

Deal data is split across the CRM, the valuation models, the data room, and a dozen inboxes. The picture rebuilds by hand every week because a new comparable lands on Saturday and the model catches up on Wednesday. The lever moves before the recommendation lands. The operator is deciding on last week's read of the market. The substrate move is one identity-resolved view of the deal that updates faster than the market does, with the agent drafting the go or no-go Friday at noon for an operator review at 2pm. Decisions that took a week ship in a day, and the operator approves every move with the math attached.

Banking and fintech. The activation cohort problem.

Activation data lives in three tools that do not reconcile. The web app reports one cohort. The mobile app reports another. The partner channel reports a third. The customer activation team spends ten to fifteen hours a week reconciling them by hand. By the time they have a clean cohort report, the cohort has already churned through its 30-day decision window. The substrate move is one customer ontology that survives the channel split. The agent surfaces at-risk cohorts before the operator opens the dashboard. Cycles compress from three weeks to two days. The LTV exposure on a typical 1,400-user at-risk cohort is north of one and a half million dollars.

Life insurance. The adviser handoff problem.

Quote to bind crosses three systems. The adviser captures context on the call. The system asks for it again at pre-medical. The underwriter asks for it again at decision. Context evaporates at every handoff. The substrate move is one policy ontology that carries the adviser context through to underwriting. The agent drafts lapse-risk saves 58 days ahead of the renewal. Cycles compress from ten days to one. ASI has seen adviser productivity rise sharply when the context captured on call survives the next two systems.

Property and casualty. The loss-to-pricing lag.

Claims, broker, and pricing systems do not reconcile. Loss ratios lag thirty to sixty days. Premium leakage shows up in quarterly review, never in real time. The substrate move is one policy and claim ontology that carries the loss signal to pricing inside three days, not sixty. The agent drafts the repricing memo with the loss signal attached. Cycles compress from sixty days to three. Mispricing on a single motor segment, caught early, can recover two to four percent of premium.

Operator-led services. The margin visibility problem.

Pipeline lives in the CRM. Time lives in the timesheet system. Billing lives in the finance stack. Margin sits where nobody looks. Utilization is rebuilt manually every Friday. The substrate move is one project ontology that joins pipeline, time, and billing in one frame. The agent flags margin gaps the same week they open instead of the same quarter they close. ASI has seen single-project recoveries of twenty to fifty thousand dollars when the scope correction lands inside the engagement window instead of after.

Same primitive. Five compositions. The named systems differ. The decision compression is identical.

03

Two ways teams get this wrong.

Two failure patterns surface again and again. They are easy to name once they show up inside a few engagements. Most operators are running one of these without knowing it.

Anti-pattern A. Bolt-on AI.

The team buys a vendor agent or wires up an LLM to a Slack channel. The model is fine. The wiring is fine. The substrate underneath is not. The agent runs against the same five disconnected systems the operators were already losing in. Output: drafts that reference customer IDs that do not exist in the operator's view, suggestions that contradict the dashboard, time saved on writing emails but no movement on revenue.

This is the version that gets the budget. It is also the version that gets pulled in twelve months because nobody can defend a single revenue decision the model influenced. The model gets the blame. The substrate was the problem the whole time.

Anti-pattern B. Ontology theatre.

The team agrees the substrate is the move, then commissions a six-month enterprise data model. Three consultancies are engaged. Stakeholders are interviewed. The slide deck is gorgeous. Production never ships. Twelve months later, the org has a 200-page ontology document, no agent in production, and a CFO asking what the spend bought.

This is the version that gets executive air cover. It is also the version that produces zero decisions. The fix is to constrain scope ruthlessly: one revenue line, one ontology, one agent, ninety days. Refuse the second sector until the first is shipping decisions weekly.

The two anti-patterns look like opposites. They are siblings. Both originate in the same mistake: failing to constrain the engagement around a named revenue line. Bolt-on AI fails because the substrate underneath is too messy. Ontology theatre fails because the substrate is over-scoped before any of it has to defend a decision.

04

Field notes from the build phase.

Anonymized notes from inside the work. Patterns that surface in every pilot, regardless of sector.

The operator approves the math, not the agent.

Operators do not trust AI agents in the abstract. They trust the math the agent attaches to a recommendation. Show the rollback window, the confidence interval, the prior, the alternative considered. Operators will approve a 0.91 confidence draft with a clean rollback in under a minute. They will reject a 0.99 draft with no math even when the model is more right. The substrate makes the math possible. Without it, every recommendation is a black box pretending to be a colleague.

The audit trail is the trust artifact, not the dashboard.

Every operator asks the same second question: "show me what happened the last time the agent did this." Not "show me what the agent thinks today." The substrate has to record every recommendation, every approval, every rollback, every operator note. Once the audit trail exists, the operator approves drafts twice as fast. Once it does not, the operator second-guesses every draft until somebody pulls the agent.

The first revenue line is the only revenue line that matters.

Every pilot that drifts dies at month four when the executive sponsor asks what the spend bought. The pilots that ship are the ones that pick one revenue line on day one and refuse to add a second one until the first is shipping decisions weekly. The expansion conversation happens after the lift is proven with the math attached. Not before.

The data team is the wrong owner.

Decision compression is an operator problem. When the data team owns the substrate roadmap, the roadmap optimises for legibility and reproducibility. Both are good. Neither is decision compression. When the operator owns the roadmap, the roadmap optimises for the next decision they have to make on Friday. That is what ships.

Security is the wedge, not the friction.

In banking and insurance, the moment the firm says "this moves through a security review" the room exhales. The teams that have been burned by previous AI engagements are the ones that ask about IAM, threat modelling, and audit logging in the first call. Treat the security conversation as a wedge, not a hurdle. Banking-grade by default is how the firm earns the second pilot, not just the first.

05

The one move to make this quarter.

Most operators reading this already have an AI agenda. Three vendors. A pilot that stalled. A board paper drafted but not shipped. A consultant pitching a six-month data architecture.

The move this quarter is to compress that agenda to one sentence: ship one clean ontology against one revenue line, with one agent and one operator, in ninety days.

If the operator can name the revenue line out loud, the pilot can run. If not, it is not an AI problem. It is a strategy problem. The pilot waits.

Phase 0 is two weeks of mapping and the build spec. Phase 1 is forty-five days building the data layer, pure infrastructure. Phase 2 is fifteen days shipping the agents and measuring the lift. The pilot runs at break-even, the operator approves every move, and ASI never charges more than it returns. The business keeps the substrate. The business keeps the operator playbook. The business keeps the audit trail. The downside is fourteen days of operator time. The upside is the first compression pattern in the sector, with the math attached.

The teams that ship this in 2026 will compound for three years before the rest of the market catches up. The substrate is the moat. Whoever holds the ontology wins the speed race. Everyone else is renovating rented land.

Sameer Chib, editor. Sydney. May 2026.