Generative AI Consulting: 10 Questions to Ask Before You Sign
MIT found 95% of enterprise generative AI pilots deliver no measurable return. The variable you control is who you hire: ten questions to ask a generative AI consulting firm before you sign, with what a good answer sounds like.
By Ivan Pylypchuk, CEO of SoftBlues. Has led Claude and Gemini implementations for finance, legal and healthcare teams across the UK and Ireland.
Before you sign with a generative AI consulting firm, ask how they define a measurable outcome, who actually does the work, what their last three projects shipped, how they handle your data under UK GDPR, and what happens if the pilot fails. The right answers are specific, numbered and backed by named references. Vague answers are the warning.
Here is a number worth pausing on. In 2025, MIT's NANDA initiative found that 95% of enterprise generative AI pilots delivered no measurable impact on profit or loss (MIT, via Fortune, Aug 2025). The same research found that buying from specialist vendors and building partnerships succeeded far more often than internal builds. The lesson is about selection. Who you choose, and how you vet them, decides which side of that 95% you land on.
At SoftBlues, an AI consulting firm that puts Claude into production for regulated mid-market companies across the UK and Ireland, we sit on both sides of these conversations: we answer these questions for buyers, and we have watched good firms lose deals because they could not. This is the procurement-grade checklist we would use if we were buying.
Key facts
Who this guide is for, and who it isn't
This is written for a 50 to 500-person regulated firm in the UK or Ireland choosing its first serious generative AI partner, where the work touches sensitive data and a failed pilot carries a real cost in time, budget and trust.
It is not for a solo founder who wants a weekend prototype, or a team with in-house machine-learning capacity that only needs extra hands. If you already know exactly what to build and just need delivery capacity, you want staff augmentation, not consulting.
What is generative AI consulting, and why is choosing a firm different?
Generative AI consulting is paid advisory and delivery work that helps an organisation put large language models, such as Claude or Gemini, into real use: scoping the problem, choosing the model, integrating it with your systems, and getting it safely into production. It sits between strategy (what should we do) and engineering (build the thing).
Choosing a generative AI consulting partner is harder than a normal software purchase for one reason: the field is young, the claims are loud, and most of the proof is private. You cannot rely on brand, and you often cannot see a rival's results. So the vetting has to come from the questions you ask and the specificity of the answers. The ten below are ordered the way a real evaluation runs, from outcome to aftercare.
The 10 questions to ask before you sign
1. What measurable outcome will this deliver, and how will we know it worked?
The single most important question. A serious firm reframes your request as a metric with a baseline and a target, and tells you how it will be measured.
A good answer sounds like: "We will cut first-draft time on your compliance file reviews from four hours to under one, measured on a sample of 50 files, with a go or no-go review at week six." Red flag: "AI will transform your productivity." No number, no baseline, no measurement.
2. Who actually does the work, and will I meet them?
In too many firms the senior person who wins the deal disappears after the kick-off, and delivery passes to people you never met.
A good answer sounds like: named engineers, their seniority, and a commitment that you meet the delivery team before signing. Red flag: the partner cannot say who will build it, or the team is "to be allocated".
3. Can you show three projects like mine that shipped to production?
Pilots are easy. Production is the hard part, and 95% of pilots never make the leap. Ask specifically for work that went live, in a sector close to yours.
A good answer sounds like: concrete examples with the problem, the result and a reference you can call, plus honesty about what did not work. Red flag: only demos, only logos, or "we can't share any of that".
4. How do you handle our data: where does it live, who can see it, and is it used to train models?
For a regulated buyer this is the question that can end the conversation. You need clear answers on data residency, access control, retention and model training.
A good answer sounds like: a Data Processing Agreement, data kept in a region you accept, no training on your content by default, and named sub-processors. They should reference UK GDPR and the ICO without being prompted. Red flag: "it's all secure, don't worry."
Models update every few months. A firm wedded to one model, or one that cannot explain why it picked it, is a risk.
A good answer sounds like: a reasoned choice (capability, cost, data terms, region) and a plan for version changes and regression testing. Red flag: "we always use X" with no reasoning.
6. How do you scope and price the work: fixed price or time and materials?
This is where budgets quietly bleed. You want to know what is fixed, what is variable, and what would make the price move.
A good answer sounds like: a fixed-price first phase with defined deliverables, honest day-rate ranges for the rest, and the cost drivers named. We cover realistic figures in our guide to AI consulting costs in the UK. Red flag: an open-ended retainer with no deliverable attached.
7. What does the pilot prove, and what is the path to production?
A pilot should be designed to make a decision, not to run forever. Ask what it proves and what happens next.
A good answer sounds like: a proof of value with explicit go or no-go criteria and a costed route to production if it passes. Red flag: a pilot with no exit criteria, which tends to become a permanent, billable science project.
8. What happens if it doesn't work?
The honest firms have an answer ready, because they have seen projects fail.
A good answer sounds like: named risks, kill criteria, and a commercial position such as a fixed-price pilot you can walk away from, or a money-back arrangement. Red flag: "it will work."
9. How will my team run it after you leave?
If the knowledge walks out with the consultants, you have bought a dependency, not a capability.
A good answer sounds like: training, documentation, a handover plan and no architectural lock-in. Red flag: a black box only they can maintain.
10. What are the ongoing costs: licences, usage and maintenance?
The build is rarely the whole bill. Licences and token usage continue, and someone has to maintain the system.
A good answer sounds like: an itemised view of licence cost, usage-based cost and a support or maintenance estimate. Red flag: the running cost is never mentioned until the first invoice.
Three ways to get generative AI built
The right answer to question 6 often depends on which model of engagement you actually need. Most buyers are choosing between three.
| Engagement | Best for | Watch out for | Indicative cost (our data, June 2026) |
|---|---|---|---|
| Consultancy (advise + deliver) | First serious project; you need scoping, model choice and production help | Vague scope; senior-then-vanish staffing | Fixed-price proof of value £10,000–£20,000; specialist work ~£300–£450/day |
| Staff augmentation | You know what to build and need vetted AI engineers in your team | Paying day rates without your own technical direction | ~£300–£450/day per engineer |
| In-house build | You have machine-learning capacity and time to learn | Highest failure rate in the MIT data; slow to first result | Salaries plus ramp-up; longest path to production |
What good and evasive answers sound like
If you only remember one thing, remember that specificity is the signal. The table below distils the pattern.
| The question is really testing | Green flag | Red flag |
|---|---|---|
| Outcome discipline | A metric, a baseline, a measurement method | "Transformational productivity" |
| Real delivery | Named engineers you meet before signing | "Team to be allocated" |
| Production track record | Three shipped projects with references | Demos and logos only |
| Data safety | DPA, no training by default, named region | "It's all secure" |
| Commercial honesty | Fixed-price phase, named cost drivers | Open-ended retainer |
A regulated worked example
A UK financial-advice firm wanted to speed up its monthly compliance file reviews. The honest scoping conversation did not start with a model; it started with the metric: hours per file, error rate, and which checks a human must always sign off. The relevant regulator (the FCA, with the firm's own SM&CR responsibilities) framed what could be automated and what could not. The output of that work was a costed proposal and a defined pilot, not a signed promise. You can see the shape of that engagement in our compliance file review example, which is a discovery and proposal piece, not a live production claim.
The point holds across sectors. In legal the regulator is the SRA; in healthcare it is the CQC plus clinical-safety standards. Name the right one early, and a good firm will already know it.
Red flags to walk away from
For the wider evaluation around these questions, our guide on how to choose an AI consulting firm in the UK covers team composition, procurement and pricing models in more depth.
Frequently asked questions
What is the difference between generative AI consulting and AI consulting?
AI consulting is the broad term for advisory and delivery work on any kind of AI. Generative AI consulting is the subset focused on large language models such as Claude and Gemini: chat assistants, document processing, drafting and agents. The vetting questions are the same; the technical specifics differ.
How much does generative AI consulting cost in the UK?
It varies by scope. A fixed-price proof of value typically runs £10,000 to £20,000, and specialist AI engineering is roughly £300 to £450 per day (our data, indicative June 2026), below the UK market mid of £650 to £900. Our AI consulting costs guide breaks down the bands.
How long does a generative AI pilot take?
A well-scoped proof of value usually runs four to eight weeks, ending in a clear go or no-go decision. If a "pilot" has no end date or exit criteria, that is a sign it has been designed to bill rather than to decide.
How do I know if a firm is just reselling someone else's product?
Ask question 2 and question 3 hard. A firm that builds will name its engineers and show production work. A reseller will struggle to explain how the thing is built or maintained, and will avoid commitments on handover.
Is my data safe with a generative AI consulting firm?
It can be, if the contract says so. Insist on a Data Processing Agreement, confirmation that your content is not used to train models by default, a data region you accept, and named sub-processors. Reference UK GDPR and the ICO, and require it in writing.
Should we build in-house instead?
Sometimes, if you already have machine-learning capability and time. But internal builds had the lowest success rate in the MIT research. For a first regulated project, a fixed-price external pilot usually reaches a result faster and with less risk.
What is the single most important question on this list?
Question 1: the measurable outcome. If a firm cannot turn your request into a metric with a baseline and a target, every later answer is guesswork.
Choosing well is the difference between joining the 5% of generative AI projects that pay back and the 95% that do not. We answer all ten of these questions in writing before any pilot, scope a fixed-price proof of value, and stand behind it with a money-back position if it fails. As a registered Anthropic Partner Network member and a Google Cloud Partner, we build in production rather than slideware.
If you are vetting a generative AI consulting partner, bring these questions to us and judge our answers against them. Book a discovery call.
