Back to Guides
Enterprise Software 15 min read

Inside an AI training RFP: what good answers look like

Inside an AI training RFP: what good answers look like

The hard part of an AI training RFP is not writing the questions. Templates for those are free and mostly good. The hard part arrives six weeks later, when nine fluent, nearly identical responses sit in a folder and nothing in them says which provider will raise your team’s capability. That is an answer-quality problem, and it gets solved before the document goes out, by asking the few questions a weak bidder cannot fake.

What follows is written from the side that fills these documents in. Published guidance on the AI training RFP process is almost entirely question lists, and the response side exists only in material written to help vendors win. The useful view is the overlap: which questions produce signal, which produce paragraphs, and how the two sound different.

First decide whether you need an RFP at all

Training Industry’s reference on the instrument is blunt about weight: an RFP is “usually a lengthier process than the RFI (request for information) and the RFQ (request for quotation)” and runs “typically 10-20 pages in length.” For a workshop in the $2,000 to $15,000 market range, that document plus review plus scoring can cost more in salaried hours than the thing being bought.

It also tilts the field. A twenty-page response rewards providers who employ proposal writers. The firms whose facilitators have actually shipped work with these tools are usually small, and the better ones decline rather than spend four days on paperwork.

Spend Instrument Why
Under $15,000, one or two teams RFQ or a scoped brief to three providers One page of needs, the same questions to each, decide in two weeks
$15,000 to $75,000, multi-department RFI to shortlist, then an RFP to three The RFI filters cheaply; only finalists absorb the long document
Over $75,000, or procurement requires it Full RFP The process cost is proportionate and the governance is real

The same source puts step one of the sequence as “a thorough assessment of the organization’s existing training operation and its sourcing needs,” with the proposal stage at step two. Most AI enablement RFPs start at step two, which is why their requirements read as generic and the responses come back the same way. If you cannot name the three tasks you want measurably faster and the person accountable, no vendor answer will supply them. That assessment belongs in the buy-versus-build decision, before the document rather than inside it.

One caution from the vendor side. Writing in Training Industry in June 2025, Amy DuVernet, PhD, CPTM, notes that “in some cases, organizations issue RFPs to meet internal requirements, even if they’re already leaning toward a particular vendor.” Providers assume this and some price for it. If the decision is made, an RFQ documenting the comparison is cheaper and more honest.

Six questions that separate bidders

Most RFP questions get answered identically by every serious and unserious bidder. These six do not. The weak answer is what arrives when a question can be answered from a capabilities deck; the good answer needs someone who has run the work.

1. Who stands in the room, and what have they built with these tools?

Weak answer: “Our facilitators are certified instructional designers with an average of twelve years’ experience delivering enterprise learning programs.”

Good answer: Names the facilitator assigned to your engagement, says what that person uses the tools for in their own work, and commits the name in the statement of work.

AI enablement is a practitioner subject. Someone who uses ChatGPT or Claude daily on real work can answer the question a participant asks in minute forty: why the output went wrong and what to change. Someone who learned the material in order to teach it cannot, and a room of skeptical senior people spots the difference inside ten minutes. A provider who will not name the facilitator before signature is reserving the right to substitute, which is the most common cause of a disappointing session.

2. How many hours will you spend building from our material, and what will you build?

Weak answer: “All our programs are fully customized to the client’s context and industry.”

Good answer: Preparation hours, a list of artifacts, and a request for inputs. “Twelve hours: three exercises built from your own documents (a contract review, a monthly reporting pack, a customer email thread) plus a prompt library for your five most frequent tasks, with those documents two weeks before delivery.”

Customization is the only claim in a training proposal with a hard unit of measure, and almost no response quantifies it. “Fully customized” usually means the deck mentions your industry. Twelve hours of preparation against your own documents is a different product at the same price, and it decides whether people leave with a transferable skill or a demonstration. A practitioner answers in a sentence, because that figure is how they priced the job. A reseller cannot answer at all. It is also where the price genuinely moves, as negotiating the engagement shows.

3. What happens to the documents we bring into the room?

Weak answer: “We maintain SOC 2 Type II certification and ISO 27001 compliance across our platform.”

Good answer: Describes the handling. Which accounts the exercises run in, whether your material touches the provider’s systems at all, what is retained after delivery, and when it is deleted.

Procurement imports this question from software buying, where the certification answer is the right one. For a facilitated session it is close to irrelevant, and sometimes a tell: a provider leading with platform certifications may be a platform selling sessions.

The real exposure is narrow. Your team will paste live material, including draft contracts, pipeline data and personnel notes, into tools during the session. The questions are whose accounts those are and what survives the engagement.

The best answer is often the simplest: exercises run in your tenancy under your existing policies and the provider holds nothing. If not, ask for retention in days and deletion in writing. The NIST AI Risk Management Framework and its four functions (govern, map, measure, manage) belong in an RFP for AI software, not this one.

4. What will people be able to do on Monday that they could not do on Friday?

Weak answer: “Participants will gain a comprehensive understanding of generative AI capabilities and develop confidence in applying AI tools to their daily work.”

Good answer: Named tasks with a named role attached. “Finance analysts will draft the monthly commentary from the raw variance report in under thirty minutes. Account managers will turn a call transcript into a client-ready summary. Operations will build and reuse a prompt for the weekly exception report.”

Comprehension and confidence are not capabilities, and a session can deliver both while changing nothing. Most responses collapse into adjectives here, which makes this the sharpest discriminator in the document. Demand verbs and roles. Any provider who answers well has also handed you your own acceptance criteria, which is what the contract terms that decide whether training lands need to reference. Asking a vendor to define success is asking them to define their own failure, and the responses divide sharply on whether they will.

5. How will we know it worked, and who does the measuring?

Weak answer: “We provide post-session satisfaction surveys and a completion report.”

Good answer: Proposes a measure taken before and after, says who collects it, and separates what the provider can measure from what only you can.

Satisfaction scores measure whether people enjoyed the afternoon, which correlates with nothing you care about. Kirkpatrick’s four levels have been standard in the field for decades and most proposals stop at level one. A serious answer reaches behavior: a short baseline task timed before the session and repeated four weeks later, a count of how many people still use a built artifact at thirty days, or a named workflow with a cycle time attached.

The honest provider also says which parts they cannot measure, because business outcomes sit behind your systems and your managers, and that candor is a quality signal. SHRM’s 2026 research with 223 learning and development executives found function-level challenges centering on budgets, ROI and skill prioritization. You will be asked to prove this worked, and the proposal is the cheapest place to agree how.

6. What is explicitly out of scope?

Weak answer: “Our capabilities span AI strategy, governance, tool selection, custom development, change management and executive advisory.”

Good answer: Draws a line. “This engagement covers hands-on fluency with ChatGPT, Claude, Gemini and Microsoft Copilot applied to your own work. It does not include tool procurement advice, policy drafting, or building anything. Those would be scoped separately.”

A bidder who is good at everything is usually reselling other people’s instructors, and the breadth is a sales surface rather than a bench. Narrow scope, stated plainly, signals competence. It protects you later, too, because the most common mid-program failure is a provider gradually revealed to be out of depth, the situation behind most decisions about replacing a training vendor. A willingness to say “not us” is rarer than any certification.

Four questions to delete, and what to ask instead

Every template carries questions that fill pages and tell you nothing. Cutting them shortens the document, raises response quality, and keeps good providers in the process.

Delete Why it fails Ask instead
“Provide your full course catalog.” Breadth measures a content library, not a facilitator. Invites an appendix nobody scores. “Which three sessions have you delivered most often in the past six months?”
“How many learners have you trained?” Unverifiable, and large numbers usually mean self-paced seat counts. “How many facilitated sessions did the named facilitator run last quarter?”
“Describe your proprietary methodology.” Every provider has a named framework with four or five stages. The answers are identical. “Walk us through your last session, hour by hour.”
“Provide three client references.” You get three delighted contacts. The sample is curated by definition. “Give us one engagement that underperformed, and what you changed.”

That last substitution is the highest-yield line in a response. Serious providers answer it, usually about a session where the wrong people were in the room or where nobody owned follow-through, and the answer says more about their judgment than any success story. The same skepticism applies to anything presented as a case study, which needs its own verification first.

Weight the scorecard before you read a single response

Weights set after the responses arrive are rationalizations. Fix them in the document, publish them to bidders, and sum them to 100. This is the part of vendor evaluation that a shortlist cannot do for you: rankings of the best AI training companies are built on reputation, while a scorecard is built on what each bidder actually answered. A defensible set for an engagement at this size:

Criterion Weight What earns the points
Named facilitator, practitioner evidence 25 A named person who uses these tools on real work, committed in the SOW
Custom build, quantified 20 Preparation hours against your documents, artifacts listed
Outcome definition 15 Named tasks and roles, not comprehension
Normalized cost 15 Fee per participant per live facilitated hour
Measurement plan 10 A before-and-after measure, with an owner
Data handling 10 Retention and deletion of session material, in writing
Post-session support 5 Dated deliverables, not “ongoing access”

Note what price is worth. Fifteen points is deliberate: at this spend, the gap between a session that changes Monday and one that does not dwarfs the gap in fee, and a scorecard weighting cost at 40% reliably selects the stock deck. Normalize before scoring, because “a full-day workshop” covers anything from four to seven hours. Then score the written sections before anyone sees pricing, since reading price first turns the rest of the scoring into narrative.

The timeline that produces comparable answers

Compressed timelines produce boilerplate, because boilerplate is what a provider can supply in four days. Three mechanics matter more than duration.

Allow three to four weeks. Training Industry’s guidance on supplier response windows puts two weeks as the minimum expected and three to four as more standard. At two weeks, every answer to question 2 above comes back generic, because nobody can scope a custom build in that window.

Hold one bidders’ call and publish the Q&A. The same source places the vendor review meeting within a week of issuing the RFP and at least a week before the due date. One call with all bidders, answers circulated to everyone, removes the asymmetry that otherwise decides the outcome.

Cap the document and add a working session. Eight pages of written response, then ninety minutes with each finalist’s actual facilitator, four of your actual people, and one of your actual documents. This is the most informative hour in the process, and where the gap between a practitioner and a presenter becomes obvious. Where the stakes justify it, formalize it as a paid pilot workshop with one team.

Underneath all of it sits an assumption worth testing first: that the capability gap is a training problem rather than a licensing or ownership problem. The case that buying licenses is not the same as building capability cuts both ways. A good process buys a good provider. It does not buy an internal owner, and no response will volunteer that.

Frequently asked questions

Is an RFP necessary for corporate AI training?

Usually not below about $15,000. A one-page scoped brief sent to three providers with identical questions produces comparable answers at a fraction of the internal cost. Reserve the full document for multi-department rollouts or where policy requires it.

What is the difference between an RFI, an RFQ and an RFP here?

An RFI gathers capability information to build a shortlist. An RFQ asks for a price against needs you have already fixed. An RFP asks for an approach as well as a price, and is the heaviest of the three. For a defined workshop, an RFQ is often right.

How long should the AI training RFP process take?

Four to seven weeks: three to four for responses, a bidders’ call in the first week, then one to two weeks for scoring and the decision. Faster and you are comparing boilerplate.

How many providers should we invite?

Three to five. Below three there is no comparison. Above five, the scoring burden grows faster than the signal, and good small providers decline crowded processes.

Should the RFP state a budget?

Yes, as a range. Withholding it produces proposals scoped at random and wastes a round. A stated range of $8,000 to $12,000 gets comparable approaches instead of one bid at $4,000 and one at $40,000.

Can we ask vendors for AI tool recommendations in the RFP?

You can, but score it separately. A provider who both recommends tools and trains on them has an interest in the recommendation, and fluency training is a different purchase from tool selection advice.

Should we require SOC 2 or ISO 27001 from a training provider?

Only if they host a platform or process your data on their systems. For a facilitated session, ask which accounts the exercises run in, what the provider retains, and when it is deleted.

What is the clearest signal that a response is boilerplate?

Adjectives where numbers belong. “Fully customized,” “comprehensive understanding” and “extensive experience” are all unfalsifiable. A response carrying preparation hours, a named facilitator, participant caps and dated deliverables took real effort to write.

Key takeaways

  • The failure mode is not bad questions but a folder of indistinguishable answers. Prevent it by asking the few questions a weak bidder cannot answer.
  • Six do most of the work: the named facilitator, quantified build hours, handling of session material, Monday-morning capability in verbs, the measurement plan, and what the provider will not do.
  • Delete the catalog, learner-count and methodology questions. Replace the reference request with a request for an engagement that underperformed.
  • Publish weights summing to 100 before responses arrive, keep normalized cost near 15 points, and score the written sections before pricing is opened.
  • Allow three to four weeks, hold one bidders’ call with a published Q&A, and give finalists ninety minutes with your real people and a real document.
  • Below roughly $15,000, a scoped brief to three providers beats a formal RFP.

If you are assembling the requirements now and want a second opinion on what belongs in them, we run a free AI-readiness call: a straight conversation about where your team’s capability gap sits and what a proposal needs to contain to close it. Book a call.

Last Updated: Nov 17, 2026

AW

Arthur Wandzel

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles