Choosing an AI vendor for SMEs: a checklist with 15 questions

Key points: Choosing an AI vendor is not a price comparison. It is a risk evaluation. The wrong partner costs you 3-9 months of time and the AI project ends up on a shelf. This checklist with 15 concrete questions separates the agencies that will actually deliver your project from those that just talk a good game. Plus red flags that tell you immediately to look elsewhere.

Before you request quotes: define scope

Avoid the most common mistake: asking an unclear question to three parties who each interpret it differently. Before you send out quote requests, answer internally:

1. What is my concrete problem? (not: "we want to do something with AI", but: "we lose 8% of leads because follow-up isn't prioritized")
2. What data do I have? (be specific: which systems, what history, what quality)
3. What is my budget range? (share it; parties won't waste time on unrealistic scope)
4. How do I measure success? (one measurable KPI, not "satisfaction")
5. What is my deadline? (a specific month, not "as soon as possible")

With this on a single page, you can approach 3-5 parties and get meaningful quotes back.

The 15 questions for every candidate

A, Experience (questions 1-4)

1. Which comparable projects have you delivered in the last 24 months? Watch out for: only pilots, no production. An agency that has only built pilots doesn't know what production looks like.

2. May I speak with two clients who are currently running your work in production? Real clients, real conversations. An agency that dodges this: red flag.

3. How many years of experience does the person leading my project have, with this specific technology? Not "our team has 20 years of experience", but who is standing at your kick-off?

4. Which SME projects have you done (versus enterprise)? SMEs have a different reality: fewer people, less budget, faster decision-making. Enterprise agencies often don't fit.

B, Technical approach (questions 5-8)

5. What is your preferred technical stack and why? Good answer: "We first look at your existing stack and adapt to it; we work with Python/scikit-learn/PyTorch and deploy on AWS/Azure/GCP depending on where you are." Bad: "We build everything in [specific proprietary platform] that we made ourselves."

6. How do you handle data quality if it turns out to be poor? Good: "We start every project with a data audit. Unexpectedly poor data results in a scope adjustment or a feasibility phase." Bad: a promise that data will always be "good enough".

7. What happens if the model doesn't work technically? An honest partner has an exit strategy. An overly optimistic partner denies this can happen.

8. What monitoring do you set up after deployment? Production AI requires continuous monitoring: model drift, data drift, performance degradation. An agency that doesn't mention this doesn't deliver production work.

C, Commercial matters (questions 9-12)

9. What is your approach to IP and code ownership? Standard: you get ownership of all code and models built specifically for you. An agency that wants to keep the IP: red flag (unless you accept a licensing model).

10. How do you deal with scope creep? Good: "We work in work packages with fixed prices. Additions are agreed separately." Bad: "We're flexible", which often translates into budget overruns.

11. What IS and IS NOT included in the quote? Be specific: training your team? Six months of production monitoring? Documentation? Updates in case of regression?

12. What percentage of projects do you deliver within budget? An honest agency: 70-85% within budget, scope creep in 15-30% of projects. An agency that claims "100% within budget" is either lying or only does fixed-price scopes.

D, Subsidies and broader strategy (questions 13-15)

13. Do you help with subsidy applications (WBSO, MIT, SLIM)? Agencies that say "no" here leave 30-50% of your budget on the table.

14. What is your experience with EU AI Act compliance? Legally required as of 2 August 2026. An agency that isn't familiar with this can land you a fine.

15. What does your support look like after go-live? Production AI needs maintenance. An agency that "delivers the project and is done" leaves you with an aging model.

Red flags, stop immediately if you hear this

• "We can improve your X by 50%", before they've seen your data.
• "Our AI is different from other AI", without a concrete technical distinction.
• "Tell us your budget, then we'll make a quote", without a discovery / scope conversation.
• "We have a patent on our proprietary algorithm", usually marketing.
• "Our projects are always successful", statistically impossible; honest agencies acknowledge failed projects and what they learned.
• "We have an AI platform that solves everything", doesn't exist. A use-case-specific approach wins.
• Unclear ownership of data or model.
• No client that has been running in production for 18+ months.

The scorecard system

Make a simple matrix:

Under 50: don't do it. 50-60: possible, but negotiate. 60-75: a serious candidate. No agency scores 75. That is a warning sign, not a recommendation.

Stratalytic in the checklist

Evaluating an agency also means evaluating us. Our answers to the 15 questions:

• Projects: Masparts, Travelaround, construction demand forecasting, predictive maintenance. All live in production.
• Technical lead: Rutger Geerlings, MSc Computer Science. Present in person at every kick-off.
• Stack: Python, a modern ML stack, deployed on your cloud (Azure, AWS, GCP, or on-prem).
• Data audit: every project starts with this, no exceptions.
• IP: you own all code and models.
• Subsidies: WBSO, MIT, SLIM built into every quote as standard.
• EU AI Act: we run compliance assessments and build in risk classification.
• Post-launch: 3 months of monitoring as standard + a one-hour-per-month check-in.

Schedule a 30-minute introductory call and evaluate us against this checklist.

Frequently asked questions

How many quotes should I compare? At least three. More than five becomes unmanageable.

What is a red flag with an AI agency? Promises without a data audit, dodging technical questions, no reference clients, unclear IP arrangements.

Should I go for a large consultancy or a specialist? For SMEs up to 250 FTE: almost always a specialist. Faster, sharper, cheaper.