AI implementation: what gets delivered?

Ask what you get from a build and the answer is usually "a working system", which is not a specification. Here is the inventory, with the four commonly missing items marked.

The thing itself, running where you can reach it

Live, doing the job, in an environment you control.

That last clause matters more than it sounds. If it runs on the supplier's infrastructure under the supplier's accounts, your operation depends on their continued goodwill and continued existence. Ask what would still be running on Monday if you parted ways on bad terms tomorrow.

Connections into the systems you already use

Writing to the calendar your team actually looks at, the records your CRM holds, the inbox somebody monitors.

The failure mode: a system that works beautifully in its own interface and requires somebody to copy results into the place where work happens. That is not automation, it is a second place to check.

The configuration, in readable form

Usually missing. The prompts, rules, thresholds, escalation conditions, and the list of things the system must never do.

This is the actual product in AI work, and it frequently lives only inside a platform the supplier administers. Require it exported, readable, and kept current rather than delivered once at handover.

Escalation paths that have been tested

Not described. Tested.

What happens when the system does not know, who is contacted, how quickly, and what happens if that person does not respond. Ask to see this demonstrated rather than explained, because it is where the real engineering sits and where cheap builds are hollow.

Documentation aimed at somebody who never met the builder

Usually missing, or delivered as a handover call.

Enough for a competent person to understand what runs, on what trigger, and what to do when it breaks. Tie a payment milestone to it. It is the most promised and least delivered item in this business, and a milestone is the only mechanism that reliably works.

Accounts and credentials in your name

Every third-party service the solution depends on, registered and billed to you, with administrator access your side.

Set up at the start rather than migrated later, because migrating accounts after the fact is a project of its own. Why this protects you more than any exit clause.

A baseline and a metric somebody looks at

Usually missing. The number the work was supposed to move, captured before it launched, plus a report that puts the current figure in front of a named person on a schedule.

Without it, success is a matter of opinion in six months, and opinion favors whoever is presenting. Measurement design.

A named owner on your side

Usually missing, because it is your job rather than the supplier's.

One person, by name, accountable for the thing. Not a team, not a committee. This is the single strongest predictor of whether an automation is still working in a year.

Evidence it was tested against your awkward cases

Not a demo of the easy path. A record of it being run against real, unusual examples from your business: the ambiguous request, the customer who is not who they claim, the case at the edge of your rules.

The phase most often cut when a price is squeezed, and the reason you find out about the exception in front of a client.

What is reasonable not to get

So the list is not read as a set of grievances.

A system that handles everything. Everything real escalates, and a supplier promising otherwise is describing something that will confidently mishandle your unusual cases.

Zero errors. These systems are consistent, not infallible. What you should get is a known error profile and a review requirement for anything expensive to get wrong.

Source code, in the ordinary sense. Much of this work is configuration on top of existing platforms rather than bespoke software, which is normal and correct. The configuration is what you need.

Perfect documentation. Good enough for a stranger to maintain is the standard.

The one-sentence test

Could a competent person you hire next year keep this running without contacting the supplier?

If yes, you were delivered an asset. If no, you were delivered a dependency, whatever the invoice called it.

Our builds go live in about ten days for a contained first project, with the accounts in your name and the configuration handed over readable. We run it afterward when that suits you and hand it over when it does not. How the consulting side decides what to build in the first place, and what the scope boundary looks like.