Choosing an AI vendor is harder than it should be, because almost everyone now claims to do AI and the demos are uniformly impressive. A polished slide deck and a smooth proof of concept tell you a vendor can sell; they tell you little about whether the thing will work on your catalogue, integrate with your stack, or still be supported in two years. This guide lays out the questions that cut through the pitch, the red flags worth walking away from, and how to tell a genuine partner from a vendor who disappears once the contract is signed.
Get clear on what you are buying before you shop
The most expensive vendor mistakes happen before you ever talk to a vendor. If you have not defined the problem, you will be sold a solution to someone else’s. Start by writing down, in business terms, what outcome you want, what success looks like in numbers, and what constraints are non-negotiable, such as data residency or a specific platform. This sits within your broader AI strategy and your build-versus-buy decision, which you should settle first. Buying makes sense when the capability is not your differentiator and a good product exists; building makes sense when it is core to how you compete.
Walk in with a clear brief and you control the conversation. Walk in vague and the vendor will define the problem to fit what they sell.
The questions that actually reveal quality
Anyone can answer “can your AI do X?” with a yes. The useful questions are the ones that expose how a vendor works when things are not ideal.
On the technology
- How does it perform on data like ours? Ask for a trial on your catalogue or tickets, not a generic demo. The gap between a curated demo and your messy reality is where most disappointment lives.
- What happens when it is wrong or unsure? A serious vendor has a clear answer about confidence thresholds, fallbacks and human handoff. A vague one has not thought about it, which tells you plenty.
- What are the realistic results? Be wary of anyone promising dramatic, precise gains. Credible vendors talk in ranges and conditions, not guarantees.
On data and security
- Where is data processed, and do you train on it? This matters legally and competitively. You generally want your customer data excluded from their training, and you need it to satisfy your GDPR obligations. No DPA, no deal.
- Who are your sub-processors? You are responsible for the whole chain behind them.
On integration and operation
- How does it connect to our stack? Ask about your specific platform and systems. “We have an API” is not an integration plan.
- What does it cost at our volume? Get the full-volume figure, not the pilot price. Usage-based pricing can scale faster than revenue if you are not careful.
- What does support actually look like? Response times, who you reach, and what happens when something breaks at the weekend.
Red flags worth walking away from
Some signals reliably predict trouble. None is automatically disqualifying on its own, but a cluster of them should give you pause.
- AI as a black box. A vendor who cannot or will not explain in plain terms how their system reaches an output. You do not need their source code, but you need to understand the approach well enough to govern it.
- Guaranteed precise results. “We will increase your conversion by 34%” is a sales figure, not a forecast. Real outcomes depend on your data and execution, and honest vendors say so.
- Demo-only proof. Reluctance to trial on your real data suggests the demo is the best it gets.
- Lock-in by design. Proprietary formats, no data export, contracts that make leaving painful. The harder it is to leave, the less incentive they have to keep earning your business.
- No clear ownership of failure. When you ask who is accountable if it goes wrong, the answer is a shrug or a redirect to your team.
- Selling the future. Heavy emphasis on a roadmap of features that do not exist yet, light on what works today.
Vendor, partner, or platform?
These words get used interchangeably, but the relationship type shapes everything. Be honest about which you actually want.
- A platform gives you tools and largely leaves you to run them. Cost-effective and flexible if you have the in-house capability, frustrating if you do not.
- A vendor sells you a packaged product. Predictable and quick to start, but you are one of many customers and the product evolves on their timeline, not yours.
- A partner works with you on outcomes, adapts to your context, and shares responsibility for results. More involved and usually more expensive, but the right choice for anything strategic or complex.
The mistake is buying a platform when you needed a partner, then being surprised that nobody is steering. Match the relationship to your internal capability and the importance of the project.
Running a fair evaluation
When you have a shortlist, structure the comparison so the decision is evidence-led rather than driven by whoever pitched best.
- Define your criteria and weights first, before you see any demos, so a slick presentation cannot quietly rewrite your priorities.
- Run a paid pilot on real data with each finalist where feasible, against the success metrics you set.
- Talk to reference customers who resemble you in size and sector, and ask them what went wrong, not just what went right.
- Test the support relationship during the trial. How they treat you as a prospect is the best version of how they will treat you as a customer.
- Score against your weighted criteria, and only then factor in price.
Pay attention to how a vendor handles the pilot’s inevitable hiccups. A partner who is transparent about a limitation and works through it is worth more than one whose demo was flawless but who goes quiet when reality intrudes. The pilot is also your first real test of whether you can move from pilot to production together.
Protecting yourself in the contract
A few terms are worth insisting on regardless of the vendor:
- Data export rights in a usable format, so you are never trapped.
- A clear DPA and defined data handling, including deletion on exit.
- Service levels with real consequences, not aspirational language.
- An exit clause that lets you leave without ruinous penalties.
- Pricing transparency, with a cap or clear mechanism for how costs scale.
Choosing well
The best AI vendor is rarely the one with the flashiest demo or the boldest promises. It is the one who understands your problem, is honest about what their technology can and cannot do, integrates cleanly, and stays accountable after the contract is signed. Slow down enough to test on real data, check references properly, and read the contract, and you avoid the expensive mistakes that come from buying a slide deck.
If you would like an independent view on a shortlist, help structuring an evaluation, or a partner who works to your outcomes rather than a one-size product, get in touch and we will help you choose with your eyes open.