Most AI projects leave nothing behind

Ask an executive how many AI initiatives their organisation has run and you often get a number larger than three. Ask what remains from the first one and the answer is usually a slide deck, a decommissioned pilot environment, and a vendor relationship nobody wants to renew.

This is not a story about models being disappointing. The models are fine. The problem is the shape of the engagement.

The standard shape

A typical initiative goes: pick a use case, assemble a corpus, wire up a model, demonstrate it, then attempt to put it in production. It demos well and degrades in operation, and the reason is always the same set of missing things — stable context, identity and authority, evidence and lineage, safe execution, and a way for a human correction to matter beyond the current session.

Teams usually diagnose this as a model problem and try a better model. It is not a model problem. It is that the system has no representation of the organisation to operate against, so every answer is reconstructed from whatever text happened to be retrieved.

The part nobody accounts for

Here is the expensive bit. When the pilot ends, the organisation is exactly as legible as it was before it started.

Nothing accumulated. The connectors were scaffolding. The corpus was assembled for one use case and is not reusable for the next. The understanding built up during the project lives in the heads of the people who did it, most of whom were contractors.

So use case two begins with a fresh round of discovery workshops, at roughly the cost of use case one. Run that four times and you have spent real money to arrive back where you started, which is precisely what the executive is describing when they say AI has not delivered for them.

What the alternative requires

The alternative is not a better pilot. It is refusing to treat organisational context as project scaffolding.

If the first engagement produces a durable, machine-readable model of how the organisation actually works — systems, integrations, data objects, processes, capabilities, cost, ownership, controls — then the second engagement starts from that model instead of from interviews. The third starts from a model that has been corrected twice. The curve bends the right way.

That requires three properties most AI work does not have:

1. The model is the deliverable, not a by-product. It is exported, owned by the customer, and readable without the tool that produced it.
2. Corrections mutate the model. When someone says “no, that system was retired last year,” that changes the shared representation rather than one conversation.
3. The reasoning engine is replaceable. If capability is welded to a specific model, the next model generation is a rebuild rather than an upgrade.

The sentence that matters

Conventional AI implementations depreciate. The engagement ends and the value decays from its maximum on delivery day.

An engagement that leaves a model behind appreciates. It is worth more in year two than year one, because it has been used, corrected, and extended — and because the next piece of work no longer has to pay for rediscovery.

That is not a claim about technology being better. It is a claim about where the asset sits when the invoice is paid. It is worth asking, of any AI proposal in front of you: when this finishes, what do we still have?

If the honest answer is “a report and some learnings,” you have seen this film before.