What you should receive at the end of an AI enablement engagement

Ask a vendor what you will be holding the week after the last session ends, and the honest ones name files. The rest name feelings. Search for AI enablement deliverables and you get ten vendor service pages naming a readiness assessment, a scored use-case list, a 90-day roadmap, and a closing deck, because those four things are what those vendors sell. If you bought ChatGPT, Claude, Gemini, or Microsoft Copilot seats a year ago for a company of 50 to 500 people, and someone has now asked what the seats bought, a roadmap is not a handover. It is a proposal for the next engagement. Below is the packet you should be holding instead, with the one test that sorts the real items from the packaging. The case for buying capability rather than more seats is made in our argument against license-first rollouts.

The Test That Sorts a Real Handover From a Thank-You Email

One question grades every item in the packet: could a person hired six months from now reach the same working level as your trained team, using only what the vendor left behind, without paying that vendor again?

Run it down the list and most close-out folders collapse. A satisfaction survey fails. A slide deck of screenshots fails, because screenshots of a chat window age out in a quarter. A recording of a live session passes weakly, since a new hire can watch it but cannot be corrected while doing the work. A prompt library written against your own close pack, your own proposal template, your own ticket queue, passes outright, because the next person can open it and work.

The test is unkind on purpose. Training spend gets approved on a promise of capability and reviewed on whether anything survived, and the gap between those two moments is where most programs quietly fail. The MIT Media Lab’s Project NANDA reported in July 2025 that 95% of enterprise investments in generative AI had produced zero return, pointing at brittle workflows and poor fit with daily work rather than at the models. Preliminary findings, not a peer-reviewed study, and worth reading with that caveat. The diagnosis still matches what close-out folders look like: nothing in them is tied to the work anyone does on Monday.

Vendors who have thought about durability tend to welcome the question. Vendors whose model needs you back next year will try to answer it with a login.

The Eight AI Enablement Deliverables a Close-Out Packet Should Contain

Eight items, in the order they get used after the vendor leaves:

1. A prompt and workflow library built on your documents. Not a generic list of 100 prompts. Twenty to forty patterns your own people wrote and tested in the sessions, against your templates, each with the task it serves, the tool it runs in, and a note on where it goes wrong. The most valuable file in the packet, and the one most often missing.
2. Role-level capability statements with before and after readings. One sentence per role, with a verb and a time: analysts draft a monthly variance commentary from the close pack in 30 minutes. Read in week one, read again later.
3. The raw measurement data and the instrument that produced it. Covered below, because it is the item vendors most want to keep.
4. A trust-and-escalation guide for the work the model must not do alone. Which outputs need a human check before they leave the building, which figures a model may never restate, what the team does when an answer looks confident and wrong. Two pages beats a policy nobody opens.
5. Session recordings and the exercise materials, in your storage. Delivered as files you hold, not as access to a portal. A portal that lapses is not a deliverable.
6. A named internal owner with a written scope. Not something the vendor produces, but something the engagement should force into being: whoever answers questions in month three and runs the next re-measure. Without that name, the questions that arrive in month three go nowhere.
7. A reinforcement schedule somebody owns. Two or three dated touchpoints over 90 days, each with a task attached. Reinforcement with no owner and no date is a wish.
8. A train-the-trainer path for next year’s hires. Either an internal facilitator who shadowed the sessions, or a written run sheet detailed enough for your owner to re-teach the material. Without it, every new cohort is a new purchase order.

Items one, three, and eight carry the weight. The other five cost a good vendor little and tell you a lot by their absence.

Two cautions. Eight tidy files with no measurement in them is still a failure, while three honest files that include the library and the data is a win. And the list assumes the sessions were hands-on work on your own material. If the engagement was a large room watching a demo, there is nothing to hand over and no document fixes that after the fact. What the paperwork should have said upstream sits in our read of an enablement scope of work.

What Strategy Vendors Hand Over Instead

Most pages selling AI enablement services describe the same close-out: four strategy artifacts and a call. Trust Insights publishes its version plainly: a task-level opportunity analysis, a per-role exposure report, a use-case scoring framework, and an implementation master plan, at a $25,000 fixed fee for 90 days with no recurring office hours in the phase. Those artifacts are well made, and the per-role task breakdown is useful.

It is also a different purchase from the one you think you are making. Strategy work answers where should we aim. Enablement answers can these 80 people work differently now. Buy the first if you have no tools and no direction. Buy the second when the licenses are already on the invoice and the usage dashboard looks fine while the output has not moved. Sorting those two budgets apart before you sign is the subject of the transformation-versus-enablement distinction.

Pay enablement money and receive a roadmap, and you have funded a plan to do the thing you thought you were funding. A corporate AI enablement packet should hold at most one forward-looking document, and it should be short.

The Measurement File Has to Be Yours

Vendor pages in this category list baseline metrics as a deliverable. Almost none hands over the instrument that produced them.

Three things make the measurement portable:

• The task list and the timing method, written down, so the same tasks can be timed the same way in twelve months by someone else.
• The raw per-person responses, anonymized if you like, not only the summary chart. A bar chart cannot be re-cut when your CFO asks a different question.
• A dated re-measure your own owner can run without the vendor in the room.

Keep the before and after figures narrow and plain. Minutes to first usable draft on a named task. Share of a role’s weekly output now passing through the tool. Number of people who can do a thing none of them could do in week one. Those are defensible in a budget review. Confidence scores and satisfaction averages are not, and they are what gets handed over when nobody asked for better.

We build our own close-out around that durability question, so the data file leaves with the client rather than sitting in a vendor dashboard. Low bar, and the line most often crossed, because a buyer who cannot re-measure alone is a buyer who renews without arguing. How a finance reviewer reads these numbers is covered in our guide to underwriting the budget, and the full arc from rollout to measured capability sits in the enablement playbook.

What You Own, What You License, What You Lose

The packet splits into three piles, and the split belongs in the contract before it matters in practice.

Pile	What belongs in it	What goes wrong
Yours outright	Prompts and workflows your people wrote, your documents and any redacted copies, baseline and re-measure data, recordings of your own sessions	Vendor claims joint ownership of prompts written on your material
Licensed to you	The vendor’s frameworks, slide templates, exercise designs, facilitator run sheets	License runs for the term only, so the material expires and the program becomes a subscription
Theirs, fairly	Methodology, internal tooling, anonymized aggregate learning across clients	Buyer over-reaches here and poisons an otherwise fair negotiation

Ask for perpetual internal use on the middle pile. A vendor who reuses a framework across twenty clients loses nothing by letting you keep teaching your own staff from it, and the ones who refuse have told you their renewal plan. Cheap engagements usually hold the squeeze here rather than in the headline price, a pattern laid out in the hidden cost of a cheap enablement vendor.

One more pile worth naming: a record you may need for a compliance file. Article 4 of the EU AI Act, applicable since 2 February 2025, obliges providers and deployers to take measures ensuring “a sufficient level of AI literacy” among staff who operate these systems. The duty is to take measures, not to certify scores. Who attended, on what material, when, plus a retained copy of the material, is the practical form of that evidence, and it costs nothing to ask for at close-out.

What You Should Not Expect to Receive

Knowing this list keeps you from burning goodwill on the wrong fight.

A certificate that means anything outside your company. A guarantee of behavior change, which no honest facilitator will sign, because managers and people decide that after the room empties. Custom software, a tuned model, or a built assistant, unless you bought a build, which is a different contract at a different price. Unlimited support folded into a workshop fee. Any claim that attendance made the team compliant with a regulation.

Also fair: a vendor declining to hand over a facilitator’s private notes, or an unedited recording of a session where your staff talked through live commercial matters.

Press hard on the data file, the prompt library, and the materials license. Shrug at the certificate.

Thirty Minutes Before You Release Final Payment

The last payment milestone is the only hold you have left. Spend half an hour on the packet before it clears.

1. Open the prompt library and run two patterns yourself, on a real document, unaided. If they fail in your hands, they fail for the new hire.
2. Check the measurement file for raw rows, a written task list, and a timing method. Summary chart only, send it back.
3. Confirm the recordings and exercise files sit in your own storage, not behind a vendor login with a renewal date.
4. Read the materials license for the word perpetual and the phrase internal use.
5. Find the internal owner’s name and the dates of the reinforcement touchpoints. If either is blank, fix it now while someone still owes you a deliverable.
6. Ask the owner, not the vendor, whether they could re-run the measurement alone next year. The answer grades the whole engagement.

Fixing any of these is hours of work for the vendor, not weeks, which is why the ask is fair and refusal is such a clean signal.

Frequently Asked Questions

What deliverables should an AI enablement engagement produce?

Eight: a prompt and workflow library built on your own documents, role-level capability statements with before and after readings, the raw measurement data plus the instrument that produced it, a short trust-and-escalation guide, session recordings and exercise files held in your storage, a named internal owner with a written scope, a dated reinforcement schedule, and a train-the-trainer path. Three decide whether the spend survives the year: the library, the data, and the trainer path.

What if the vendor just hands over a deck and a survey?

Treat it as an incomplete delivery and name what is missing in writing before the final payment clears. A deck records a session; it is not a capability anyone can pick up. Most vendors can build the library and the data file inside a week from material they already hold, because the work happened in the room. Refusal usually means the sessions were a lecture and there is nothing underneath to send.

Who owns the prompts our team wrote during the training?

You should, outright. Your people wrote them, on your documents, on your time. A vendor may fairly keep its own frameworks, slide designs, and facilitator run sheets, and may fairly reuse its methodology elsewhere. Joint ownership of work your staff produced on your material is not a standard term, and asking for it is a flag worth chasing before you sign anything further.

Is a recording library a real deliverable?

Only if you hold the files. Recordings help someone who joins after a cohort runs, and fail as the main handover, because few people pick up a hands-on skill by watching. Ask for raw files plus a transcript in your own storage. A vendor portal that lapses on renewal is a sales asset wearing a deliverable’s name.

How do we know the program worked?

Compare the same named tasks before and 30 to 90 days after, using the instrument the vendor hands over. Minutes to first usable draft, share of a role’s weekly output now moving through the tool, count of people who can do something none of them could do at the start. Usage dashboards and satisfaction scores both look fine in programs that changed nothing, which is why they get reported.

Can we train new hires on this ourselves afterward?

Yes, if you asked for the trainer path up front: an internal facilitator who shadowed the sessions, or a run sheet with timings, exercises, and the common failure points written out. Price it in the first scope, since a vendor quoting it after close-out quotes it as new work. Without it, each cohort is a fresh purchase order and nothing compounds.

Our vendor wants to keep the baseline data on their platform. Is that normal?

It is common and you should still refuse. The stated reason is usually dashboard tooling. The effect is that you cannot re-measure, or argue the number in a budget review, without buying another engagement. Ask for an export of the raw rows and the written method. A vendor confident in the result has no reason to hold the file.

Half our people skipped the sessions. Does the packet still help?

It helps the ones who came and gives you a way to reach the rest, which is more than a recording does. Low turnout is a management problem rather than a vendor one, and it is common: a LinkedIn survey reported by Fortune in August 2025 found 51% of professionals call the intensity and frequency of AI training excessive. The library and the run sheet let your owner re-teach a small group on real work, which lands better than the all-hands version anyway.

Where to Start

Pull the last close-out folder a vendor sent you and look for three files: a prompt library written against your documents, a data file with raw rows, and a run sheet your own owner could teach from. Most folders hold none of the three. That absence is the finding, and it tells you what to write into the next scope before anyone talks about price.

For a second read on a packet you already have, or on a proposal that promises one, book a free AI-readiness call. We will say which items are worth pressing for, and tell you plainly if your team does not need a paid program yet.