AI Consulting Vendor Shortlist: How to Compare Proposals, Teams and Delivery Risk
More than 80% of AI projects fail, twice the rate of normal IT work (RAND, 2024). The difference between a partner that ships and one that stalls is visible in the proposal. Here is how to read it.
By Ivan Pylypchuk, CEO of SoftBlues. Has led Claude and Gemini implementations for finance, legal and healthcare teams across the UK and Ireland.
More than 80% of AI projects fail to deliver their intended value, roughly twice the failure rate of IT projects that do not involve AI (RAND, 2024). By the time you have two or three AI consulting proposals on the desk, most of the information that predicts which way yours will go is already in front of you. You just have to know where to look.
To shortlist an AI consulting partner, compare three things side by side: the proposal (is the problem, scope and success measure specific, or vague?), the team (will senior people actually do the work, or just sell it?), and delivery risk (fixed scope, clear data and security answers, a real exit). The vendor who is precise about all three is the safer bet, almost regardless of price.
At SoftBlues, an AI consulting firm working with regulated mid-market companies across the UK and Ireland, we sit on both sides of this table: we write these proposals, and we are often brought in to audit AI work that another supplier left half-finished. This guide is the scorecard we wish more buyers used before they signed.
Key facts
Who this is for, and who it isn't
This is for a 50 to 500-person UK or Ireland company, often in a regulated sector, that has already decided to buy AI help and now has a few proposals to compare. You are the CEO, COO, CFO or head of operations who will sign, and you want a way to compare suppliers that is fairer than gut feel and a day rate.
It is not for a solo founder wanting a weekend prototype, or for a team that has not yet defined the problem. If you are still deciding whether to use a consultant, a consultancy or an in-house hire at all, start with how to choose an AI consulting firm in the UK and come back here once you have proposals to compare.
Why shortlisting an AI partner is different from normal procurement
In ordinary procurement you compare a known thing: three quotes for the same window, the same audit, the same managed service. AI work is not like that. Two suppliers can quote the "same" project and mean completely different things, because the scope depends on assumptions about your data, your systems and how much change your team can absorb.
That is why price is a weak first filter here. The cheapest proposal is often the one that quietly assumes your data is clean, your process is documented and integration is somebody else's problem. The RAND research is blunt about where projects actually die: misaligned purpose between leaders and technical teams, poor data, and a focus on the technology rather than the problem. None of those are price problems. All of them are visible in a proposal if you read for them.
What belongs in every proposal you compare
Score each proposal against the same checklist. If a section is missing, that absence is itself a finding. Use a simple traffic-light scale (clear / partial / missing) per row so the proposals line up.
| What you are scoring | What a strong proposal shows | What a weak one shows |
|---|---|---|
| Problem definition | One specific business problem, named, with the cost of leaving it unsolved | "Leverage AI to drive efficiency" |
| Success measure | A number you could verify in 90 days (hours saved, error rate, turnaround time) | No measure, or "improved productivity" |
| Scope and deliverables | A fixed list of what is and is not included | Open-ended "discovery" with no end state |
| Data and security | Where your data goes, who can see it, what is retained, what is not | Silence, or "enterprise-grade security" with no detail |
| Team | Named people, their seniority, who does the build vs the sales | A logo, an "expert team", no names |
| Pricing model | Fixed price or capped, tied to the deliverables above | Open day rate with no ceiling |
| Exit and ownership | You own the code, data and prompts; you can leave | Lock-in, or unclear IP ownership |
A proposal that scores "clear" across all seven rows is rare, and worth a premium. One that is "missing" on data, success measure or team is not cheaper, it is just deferring the cost to later.
How to run team diligence: who will actually do the work
The most common gap between a proposal and a project is the team. You are sold by senior people and delivered to by whoever is free. The fix is to make the team part of the contract, not a surprise.
Ask each shortlisted supplier these, and compare the answers.
1. Who, by name, will build this? You want the actual engineers and their seniority, not "a dedicated team". A good answer names two or three people and what each will do.
2. Have they shipped this in your sector before? Ask for one example close to your problem, with the outcome and what went wrong. A practitioner will tell you about the hard part. A salesperson will tell you it went perfectly.
3. What is the ratio of senior to junior time? A proposal that is mostly junior hours at a blended senior rate is the classic margin trap. You are paying for experience you will not get.
4. Who owns the relationship if a key person leaves? Continuity matters more on AI work because so much context lives in the people. A good answer has a named second point of contact from day one.
5. Can you speak to a past client directly? Not a logo on a slide, a reference call. The willingness to arrange one is a strong signal in itself.
When we are asked to audit AI work that stalled, the root cause is rarely the model. It is usually that the people who understood the problem were never the people who built the thing. We saw exactly this on a Claude Code audit and secure rebuild for a food producer, where an app built quickly by a non-specialist team needed a proper review before it could be trusted in production.
The delivery-risk signals that predict a failed project
Some warning signs are reliable enough that one of them should drop a supplier down your list, and two together should drop them off it. These are the patterns we see most often in projects that later need rescuing.
1. No fixed scope. "We will scope it as we go" with no cap means the budget is open and the end state is undefined. Discovery is fine as a paid first phase, but it must produce a fixed proposal for the build.
2. Vague data answers. If a supplier cannot tell you where your data will be processed, what is retained and who can see it, they have not thought about it, and you will inherit that gap.
3. No success measure. If nobody can state how you will know the project worked, you have bought activity, not an outcome.
4. A pilot with no path to production. Many AI pilots look great and then never ship. Ask, at proposal stage, what it takes to move from pilot to production, and who pays for it.
5. All upside, no honesty. A supplier who names no risks and no situations where their approach is the wrong fit is selling, not advising.
6. Lock-in by design. If you would not own the code, prompts and data, or could not switch supplier without starting over, that is a commercial risk dressed up as a technical one.
How to compare pricing and engagement models
Once two proposals are close on substance, the engagement model usually decides it. The model determines who carries the risk if the work takes longer than planned, which on AI projects it often does. Compare the model, not just the number.
| Engagement model | Best for | Avoid if |
|---|---|---|
| Fixed price, fixed scope | A well-defined problem with a clear deliverable and success measure | The problem is still genuinely exploratory |
| Capped time and materials | Exploratory work where you want flexibility but need a ceiling | You need a guaranteed final cost to get sign-off |
| Paid discovery, then fixed build | Most mid-market projects: de-risk the unknowns, then commit | You already have a tight, validated spec |
| Staff augmentation | You have the in-house lead and need extra hands | You need someone to own the outcome, not just the hours |
| Open day rate, no cap | Almost nothing at shortlist stage | Always, unless the scope is tiny and you are managing it closely |
The model we recommend for most mid-market buyers is a short paid discovery that produces a fixed-price build proposal. It turns the biggest unknowns into a small, bounded spend before you commit to the large one. Whatever model you choose, tie payment to deliverables you can verify, not to elapsed time.
What to ask on the shortlist call, and what a good answer sounds like
By the call you have read the proposals. The call is for the things a document hides: how they think, and how they handle a hard question.
"What would make you walk away from this project?" A good answer is specific: bad data, no executive owner, a success measure nobody will commit to. A weak answer is "nothing, we can do anything".
"Show me a project that went wrong and what you changed." Practitioners have these stories and tell them plainly. The absence of any failure story is a flag, not a strength.
"What do you need from us to succeed?" A serious supplier asks for a named owner on your side, access to the right people and honest data. One who needs nothing from you has not thought about delivery.
"What happens after go-live?" You want support, monitoring and a handover, not a supplier who disappears once the invoice clears.
"If we stopped after discovery, what would we own?" The answer should be: a clear plan, and anything built so far, with no lock-in.
Shortlisting in regulated sectors
If you are in financial services, legal or healthcare, add one more layer: the supplier has to understand the framework you answer to, at the level of knowing what your compliance team will ask. For UK financial services that means the FCA and the Senior Managers and Certification Regime; for law firms in England and Wales, the SRA; for healthcare in England, the CQC and clinical-safety standards (DCB0129 and DCB0160). Across all of them, data handling sits under UK GDPR and the ICO.
You are not asking the supplier for legal advice. You are checking they will not design something your compliance team has to unwind later. A supplier who asks about your regulator before you raise it is one who has done this in a regulated setting before.
Frequently asked questions
How many suppliers should be on a shortlist?
Two or three. One is not a comparison and gives you no leverage. More than three or four and the diligence becomes shallow, because you cannot run proper team and reference checks on six suppliers at once. Shortlist the two or three closest to your problem and go deep.What is the difference between choosing and shortlisting an AI partner?
Choosing is deciding the category and budget: consultant, consultancy or in-house, and roughly what you will spend. Shortlisting is comparing the specific suppliers in that category on a single scorecard. This guide is about the second step; how to choose an AI consulting firm covers the first.Should I pick the cheapest proposal?
Only if it is also the most specific. On AI work the cheapest proposal is often the one that has assumed away the hard parts: clean data, easy integration, a path to production. Score substance first, then let price decide between proposals that are genuinely close.How do I compare proposals fairly when they are structured differently?
Ask every supplier to map their proposal to the same scorecard (problem, success measure, scope, data and security, team, pricing, exit). Normalising the proposals onto one grid is the only way to compare them like for like, and the willingness to do it is itself a signal.What is the single biggest delivery risk to check for?
A pilot with no funded path to production. Most AI work fails not at the demo but in the gap between a working prototype and something your team can run every day. Make the route from pilot to production an explicit part of the proposal.Do I need a technical person to run this comparison?
It helps, but the scorecard is designed so a non-technical buyer can use it. The questions that matter most, around problem definition, success measures, team and exit, are commercial, not technical. For the deeper technical checks, bring in an adviser or use the 10 questions to ask before you sign.How long should shortlisting take?
For a mid-market project, a week or two: time to read the proposals against the scorecard, run two or three reference calls, and hold a shortlist call with each supplier. Rushing it to days is how the team and delivery-risk checks get skipped, which is exactly where projects come unstuck.We built SoftBlues as a practitioner shop, not a slide factory: a registered Anthropic Partner Network member and Google Cloud Partner that puts AI into production and is often called in to fix the work that did not. If you have proposals on the desk and want a second pair of eyes on the team and the delivery risk, or you want to see how we structure a fixed-scope build with a clear exit, book a discovery call. You can also see how we approach business process automation end to end.
See it in production
Systems we have built and run for clients, with the numbers that came out of them.
Related Articles
Build vs Buy AI: What UK Mid-Market Companies Should Build in 2026 (and What They Should Not)
Shadow AI at Work: What UK Companies Should Do About Unapproved AI Tools
