AI deployments that will not fail in 2026

Renting means starting from zero on every query. The general-purpose model rediscovers your context, burns tokens reassembling what your teams have known for fifteen years, and bills you for that reasoning every month. Cancel the subscription and nothing stays in the building.

Skin-deep verticalisation rents you the same thing under a different brochure: standard product, prompt engineering, a line that reads “built for banking”. So does the connector. SharePoint, Notion, Copilot, Gemini: chop up the corpus, index it, and the question brings back a few passages by semantic proximity. In a demo it is spectacular. In production the system does not know what it failed to retrieve, that the 2023 procedure was repealed, that this paragraph is classified and the person asking is not. It hands back a plausible summary nobody can verify. In front of a credit committee, that is called a liability.

A system that commits a decision has to know what it queried and what it ruled out. Language can stay probabilistic at the input. The path to the data cannot. Which assumes your organisation has been modelled, and nowhere is it.

An AI that knows insurance is worth nothing to you: your competitors have the same one, at the same price, on the same day. Nobody ever built a durable advantage on an asset the whole market rents. What sets you apart sits in your internal naming, your unwritten rules, the exceptions you tolerate, the way a case escalates in your house. No training corpus contains that: it is in your databases, your documents, and the heads of thirty people, some of whom retire this year.

Hence bespoke, the only viable form of enterprise AI.

The ontology

Take the yield rate on your lines. Every site in the group counts stoppages its own way, handles changeovers its own way, pulls planned maintenance out of the calculation or leaves it in. A model wired to your documents will guess, well or badly depending on the day. An ontology declares the rule, traces it down to the machine counters, and makes it binding in the room.

The word is worn out, so let us be exact. Not a data catalogue, not a glossary with extras. The executable model of your organisation: the entities that really exist in your house, a Batch, a Non-conformity, a Counterparty, a Monitored Asset, their relationships, the operations permitted, the rules that constrain them, the classification that governs every link.

And it moves. A frozen ontology becomes, inside eighteen months, a fiction the business works around. Every new source, every exception from the field, every regulatory change amends it, along with the effect on past decisions. It compounds instead of depreciating.

Palantir proved the principle with Foundry, at the old world’s price: months of mapping, engineers in residence, an SAP migration invoice. The whole point of 2026 is getting the same asset in a few hours.

Monarch

We hand the mapping to agents. They read your schemas, profile your data, interview your experts, test the model against the real thing, and the ontology comes out of that. Then it has to be filled: that is the job of Galahad Mercure, our ingestion component, which industrialises the extraction and processing pipelines all the way to the declared entities. Read-only, no migration.

Then come the operational agents, and that is where projects are won or lost. Nobody serious hands the keys to the kingdom to autonomous agents. So Monarch works as a control plane: they never reach the data directly, they go through the ontology, which carries the clearances. An agent inherits the level of the operator who invokes it and sees nothing beyond it, including where classified and unclassified sit side by side. Every write is previewed, reversible, logged: which data, which rule, who decided, at which second. The models run inside your perimeter, on your own weights if you want them. Holding the weights is holding the means of production. Short of that, you are handing your knowledge to a vendor who will resell it to your competitor, with your money.

What it looks like

Manufacturing. A supplier flags a drift on a batch of parts. Today, two departments spend three weeks reconstructing which units are affected, working through bills of materials, delivery notes and engineering changes by hand. With the model in place, the list lands in minutes, sourced, and holds up in front of the certification authority.

Insurance and banking. Cover, exclusions, internal precedent, counterparties and exposure limits, declared once and for all. Every conclusion rests on the clause it comes from, and an auditor walks the same path six months later. Fraud becomes a pattern in the graph.

Intelligence. Twenty compartmented sources, each at its own level. The agents correlate across silos without ever pushing classified material down where it has no business being, and command gets the traceability it will demand before committing to anything.

None of this standardises, which is what makes it defensible. What standardises, your competitors buy too.

In three years, the model of your organisation will belong to someone. It may as well be you.