Most AI enablement proposals arrive as two pages with a price at the bottom and a deliverable line that reads “AI training workshop, half day.” That sentence binds nobody. It does not say how many people are in the room, who stands at the front, how much of the clock is hands-on, what you supply, or what would count as the vendor failing to deliver. Everything a buyer cares about lives in those gaps, and they get filled after signature by whoever has the stronger hand. Below is a read of the document itself, the statement of work (SOW): what each section of an AI enablement scope of work has to say, the sentence that signals a webinar in disguise, and the edits to make before you countersign. It assumes you have bought ChatGPT, Claude, Gemini, or Microsoft Copilot seats for a company of 50 to 500 people, and someone has now asked what the licenses bought. Why that question has a capability answer rather than a license answer is the argument in the case for building capability instead of buying seats.
What This Document Has to Do That a Software SOW Does Not
Buyers have gotten good at scoping AI builds. In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. Procurement responded by tightening build contracts: named evaluation cases, acceptance thresholds, model ownership, handoff runbooks. That discipline never reached the training line, where the same companies still sign a paragraph.
It should, because this deliverable is harder to pin down. A build either runs in production or it does not. An enablement engagement produces a changed person, and a person can sit through four hours and change nothing. Three clauses carry nearly all the weight, and all three are missing from the templates that rank for this search: how a session is defined, what acceptance measures, and what you owe. The rest is contract furniture.
You are buying capability in named people, not access to a course library, and the two get sold under one word. How that difference shows up on an invoice is in our breakdown of what you are buying when a vendor says enablement.
Background and Scope: Capability, Not Curriculum
The background section should run four sentences and contain a number. It names the business problem, the roles affected, the tools in place, and the state of the work the training is meant to change. A paragraph that could sit in any other company’s contract tells you the program predates you.
Then the scope. A usable AI enablement scope of work states, per role, what a person will be able to do afterward that they cannot do now, with a verb and a time. Something like: finance analysts will produce a first-draft monthly variance commentary from the close pack in 30 minutes, and will know which figures the model must never be trusted to restate. Three or four of those sentences are the whole scope, and they double as your acceptance criteria, so they belong in the document before the price is agreed.
What arrives instead is a module list. “Module 1: Introduction to Large Language Models. Module 2: Prompt Engineering Fundamentals.” That is a table of contents, and it commits the vendor to talking for a set number of hours and nobody to an outcome.
If your draft reads that way, ask for capability statements per role before you negotiate anything else. Most vendors can do this in a day. The ones who cannot have told you something.
Deliverables, and the Clause That Defines a Session
Deliverables are countable things with dates: a baseline readout per role, the sessions, exercise materials built on your documents, a prompt library your team keeps, a 30-day reinforcement touchpoint, and a closing readout against the baseline. Six items, each with an owner and a date.
Then the clause almost nobody writes, which defines what one delivered session is:
| Term | What the clause should say | What proposals usually say |
|---|---|---|
| Cohort cap | No more than 16 to 20 attendees, with a named consequence if exceeded | Nothing, or “up to 50 participants” |
| Hands-on share | At least 60% of scheduled time with attendees working in the tool on their own files | “Interactive format” |
| Facilitator | Named in the SOW; substitution on written notice only, with equal delivery history | “A senior member of our team” |
| Duration | Clock hours of delivery, exclusive of breaks and setup | “Half day” |
| Materials | Role-specific, built on client artifacts listed in an appendix | “Course materials provided” |
Every row is cheap to honor if the vendor means to, and impossible if the plan was a webinar. The cohort cap is the highest-value edit in the document: above roughly 20 people, hands-on practice stops being possible and the event becomes a lecture with a chat window. Naming the facilitator comes second. The person who sold the engagement is often not the one who delivers it, which is fine, as long as the document says who does.
Acceptance: The Clause Most Training Contracts Skip
Acceptance is where an enablement SOW borrows the wrong language. Build contracts accept against a test set, so the training document copies the shape and accepts against attendance: delivery is complete when the session occurs. Under that clause, a vendor who reads slides to 60 bored people has performed.
Tie acceptance to the baseline instead. Three pieces make it work:
- A pre-session baseline, per role, 15 minutes per person: how long the named tasks take now, what people do with the tool today, where they do not trust it. A deliverable with a date, before the first session.
- A materials review window. Executed training contracts have done this for decades. Sample acceptance-of-training clauses require manuals and a syllabus to be submitted a set number of business days before the first class for the buyer’s approval, not to be unreasonably withheld. Ten business days is a fair ask, and those materials tell you whether your documents made it in.
- A 30-day re-measure against the same baseline tasks, written into the closing readout.
Add a remedy. If a session breaches the delivery terms, cohort cap exceeded or facilitator swapped without notice, the vendor re-delivers at its own cost. Without a named remedy, the acceptance clause is a sentiment.
Keep milestones few, three to five, each releasing a payment. How that sequence fits a rollout from licenses to measured capability sits in our enablement playbook.
Your Side of the Document, Priced in Staff Hours
The assumptions-and-responsibilities section is where the real budget hides. For a software build, client obligations are credentials and sample data. For an enablement engagement they are payroll.
Your side should name two real artifacts per role and who clears them, a redaction decision in writing, the tool and account people work in, an internal owner with allocated hours rather than a volunteer, and an attendance commitment from the managers whose teams are in the room.
Then do the arithmetic the proposal leaves out. Eighteen people in a four-hour session is 72 paid staff hours. Add document prep, manager time to hold the follow-through, and the senior review load afterward, and the internal cost can rival the fee.
Put that number in front of your CFO yourself. Our look at what these engagements include and quietly leave out walks the costs that land on payroll instead of the invoice.
Exclusions, Fees, and Change Control
The out-of-scope section protects both sides, and a vendor with a point of view writes it unprompted. For a fluency program it should exclude platform licenses, policy drafting, custom builds, data migration, and production access. Ours is narrow by choice: fluency with the tools your team already pays for, applied to the work it already does. A proposal with an empty exclusions list will absorb whatever you ask for and bill it later.
Three fee structures dominate, and each pays the vendor to behave differently. Per-seat rewards headcount, so the vendor wants the biggest possible room, which is what the cohort cap exists to stop. Per-cohort rewards sessions delivered. A retainer funds reinforcement, and can also bill for months of nothing.
Market ranges for a hands-on fluency program run roughly $2,000 to $15,000 by cohort count and depth; automation work that sometimes follows sits at roughly $25,000 to $150,000. Whichever structure you pick, release payment against accepted milestones, not the calendar.
Change control needs one line most drafts omit: the per-cohort rate for sessions added later, fixed for the term. Programs expand when they work, and setting that number early beats a discount on the first cohort.
Your Documents and the Literacy Duty
The exercises run on your files, so the document has to say what happens to them: no retention past the engagement, no use of your material to train or tune any model, no reuse as teaching examples for other clients, and a named tool and account where the work happens. Ask for written deletion confirmation at close.
Name the compliance thread in the background section if it reaches you. Article 4 of the EU AI Act has applied since 2 February 2025 to providers and deployers of AI systems, including contractors acting on your behalf. The Digital Omnibus rewrote it with effect from 27 July 2026, so the duty now runs to taking measures that support staff AI literacy rather than ensuring a level of it.
Nobody has to score your people, but a record of what was delivered to whom, and when, is worth having. That puts attendance logs and a copy of the materials in your deliverables list rather than in a vendor’s drive.
Materials, IP, and the Exit
Four lines decide whether you bought capability or rented it.
Materials license. Perpetual, internal-use, royalty-free, covering slides, exercises, and the prompt library. Materials licensed “for the term” leave when the vendor does.
Recordings and artifacts. You own the recordings and anything your people create in the room, including prompt patterns built on your documents.
An internal-facilitator path. Priced in the SOW or as a named option, so your own people can re-teach the material to new hires next year. A vendor whose model needs you to keep buying will resist this line, and the resistance is your answer.
Termination. Either side on 30 days’ written notice, pro-rata refund for undelivered cohorts, no claw-back on materials already licensed.
Those four, plus the cohort cap and the baseline, are what separate an enforceable document from a hopeful one. The vetting questions upstream of the paperwork sit in our twelve rules for choosing an enablement partner.
A One-Hour Redline Pass
If a vendor SOW is on your desk, you can fix the load-bearing parts in an hour. In order:
- Rewrite the scope as one capability sentence per role, with a verb and a time, and ask the vendor to sign the rewrite.
- Insert the cohort cap, the hands-on percentage, the clock hours, and the facilitator’s name.
- Add the pre-session baseline as a dated deliverable and the 30-day re-measure as the last one.
- Add the ten-business-day materials review window, and the re-delivery remedy for a session that breaches the delivery terms.
- Fix the per-cohort rate for added sessions, for the term.
- Change the materials license from term to perpetual internal use, and price the train-the-trainer option now.
None of this costs an honest vendor anything. We have seen proposals come back with every edit accepted inside a day, and one come back with the cohort cap struck through. The second response told us more than any reference call.
Frequently Asked Questions
What should an AI enablement scope of work include?
Eight sections: background with a baseline, scope as capability statements per role, countable deliverables with dates, delivery terms defining a session, acceptance tied to the baseline, your own obligations including staff hours, exclusions and fees with a change-control rate, and IP with a materials license and exit. Three decide the engagement: the session definition, acceptance, and your obligations.
What are good acceptance criteria for a training engagement?
Delivery against the named capability statements, measured before the sessions and again 30 days after, plus the materials review before the first class. Attendance and satisfaction scores are not acceptance criteria: accept on attendance and you have agreed that showing up counts as working.
Should an AI enablement engagement be fixed price or time and materials?
Fixed price per cohort, with payment released against accepted milestones. The scope is knowable in advance, so the vendor is better placed to carry the variance. Never pay hourly for a facilitated session whose length you can count.
Who owns the materials and prompts afterward?
You should own what your people create and hold a perpetual internal-use license to the vendor’s materials. Vendors keep their own frameworks and reuse them elsewhere, which is fair. A license that expires with the engagement is not: it turns the program into a subscription.
What should the client obligations section say?
Which documents you supply and who clears them, the redaction rules, the tool and account people work in, the internal owner with hours allocated, and a manager-backed attendance commitment. Write the staff-hour total beside it: three cohorts of 18 is well over 200 paid hours, and that belongs in the budget conversation.
How many milestones should a first engagement have?
Three to five. Baseline accepted, materials accepted, delivery complete, 30-day readout accepted is a clean four. Fewer than three leaves one payment event, which removes your only hold on the vendor after signature.
Can the scope of work help us meet an AI literacy obligation?
It can help you evidence one. If Article 4 of the EU AI Act reaches your staff, the useful output is a record: who was trained, on what material, when. Name attendance logs and a copy of the materials as deliverables. The duty is to take measures supporting AI literacy, not to certify scores, so a documented program plus retained materials is the practical form of it.
Our vendor says this is too much paperwork for a workshop. Is it?
The edits fit on one page and cost nothing but specifics. Pushback on all of them usually means the delivery model is a large room and a slide deck. Pushback on one, with a reason, is a normal negotiation. Facilitators who have run these sessions tend to welcome the cohort cap, because it lets them do the job properly.
Where to Start
Read three places in your draft: the scope section, the deliverable line for the sessions, and the client responsibilities. If the first is a module list, the second has no cohort number, and the third is one sentence about a meeting room, you are holding a webinar contract. The redline pass above turns it into something you can hold a vendor to.
When you want a second read on a draft, book a free AI-readiness call. We will go through the clauses, say which are worth pressing for, and tell you plainly if you do not need a paid program yet.
Arthur Wandzel
