What to Ask an AI Vendor Before Hiring One

Signing the wrong AI vendor contract can lock your company into recurring license fees, vendor-owned code, and integrations that break the moment you scale. Before you approve a proposal or shake hands on a pilot, there are 12 questions every decision-maker should ask — and specific red-flag answers to watch for.

This guide is built for CTOs, product leaders, and operations executives who are evaluating AI development partners, not off-the-shelf SaaS tools. The stakes are higher: you're commissioning something custom, and the terms matter as much as the tech.

Why the Vendor Interview Matters More Than the Demo

Demos are polished. Contracts are permanent.

A vendor can show you a slick prototype in 30 minutes that took their team three weeks to stage. What they won't volunteer: who owns the code when the engagement ends, what happens to your data, or whether the quote you're looking at will double once you need "additional integrations."

The questions below are organized into four categories: ownership, economics, execution, and fit. Work through all four before advancing any vendor to a final proposal stage.

Ownership Questions: Who Controls What You Pay to Build

These are the questions most buyers skip — and the ones that cost them the most.

1. Who owns the IP and source code at the end of the engagement?

This is non-negotiable. Some vendors retain ownership of the underlying architecture and license it back to you on a monthly basis. You pay to build it and then rent it indefinitely.

What to listen for: "Full IP transfer" and "you own 100% of the source code" stated explicitly in the contract, not just verbally.

Red flag: Any language like "perpetual license," "platform fee," or "proprietary framework license" that persists after go-live.

2. Will your proprietary data be used to train shared models?

If the vendor runs on top of a foundation model (GPT-4o, Claude, Gemini), ask whether your inputs, outputs, or fine-tuning data are retained by the model provider or by the vendor for any purpose beyond your project.

What to listen for: Clear data isolation policies, reference to enterprise API agreements (e.g., OpenAI's zero-data-retention option), and contractual data deletion terms.

3. Can you fork or migrate the codebase to another team or cloud provider?

Vendor lock-in through infrastructure is as dangerous as IP lock-in. If the system only runs on the vendor's proprietary cloud layer, you're dependent on them for uptime, pricing, and future development.

Economics Questions: What You'll Actually Pay Over Time

AI vendor pricing is frequently obscured in the first proposal. These questions surface the real number.

4. What is the total cost of ownership at 12, 24, and 36 months?

Ask for a written projection, not a verbal estimate. Include: development fees, hosting costs, API usage at your expected volume, maintenance, and any per-seat or per-call fees.

Why it matters: A $50,000 build with $8,000/month in API and platform fees costs $338,000 over three years. A $120,000 build with no recurring license and $800/month in cloud costs totals $148,800.

5. Are there recurring license fees — and what triggers them?

"No license fees" can still hide per-API-call charges, support tier fees, or mandatory upgrade fees when the vendor deprecates a version.

What to ask: "Show me a client invoice from month 18." Real invoices tell you what reality looks like after the honeymoon period.

6. How is scope change priced?

Most AI projects evolve after the first sprint. Ask whether change requests are billed hourly, whether there's a fixed-fee buffer, and what the process is for reprioritizing scope mid-engagement.

Execution Questions: What to Ask an AI Vendor About Delivery

Understanding how a vendor actually ships is where most due diligence stops being polite and starts being productive.

7. What is your delivery timeline, and what milestones can I verify?

Vague timelines — "approximately 3–6 months" — are a planning liability. Ask for a week-by-week milestone map with clear deliverables at each checkpoint.

Concrete benchmark: Custom AI-native software built to production standard should be deliverable in 12 weeks for a full-scope engagement. Faster, focused scopes (a single workflow, a single integration) should be measurable in days or weeks, not quarters.

8. Who specifically will work on my project?

"Our team" is not an answer. Ask for the names and LinkedIn profiles of the engineers, ML practitioners, and project lead assigned to your account. Find out whether they're full-time employees or contractors, and whether the person presenting the pitch is the person writing the code.

Red flag: Vendors who can't commit to named team members before signing.

9. Show me a production system you've shipped — not a demo.

Request access to a live product or a detailed case study with measurable outcomes. If a vendor can't point to something in production, they're selling you a hypothesis.

What to look for: Specific metrics — "reduced processing time by 40%," "automated 1,200 manual reviews per week," "cut customer support response time from 6 hours to 14 minutes."

10. How do you handle model drift, hallucinations, and failure states?

Production AI systems degrade. Foundation models get updated. Outputs drift. Ask how the vendor monitors for quality degradation post-launch and what their SLA is for critical failures.

What to listen for: Evaluation pipelines, automated regression testing, defined escalation paths, and a maintenance retainer option with explicit response time commitments.

Fit Questions: Culture, Communication, and Long-Term Alignment

11. What does your post-launch support model look like?

The first 90 days after go-live are when most issues surface. Ask whether post-launch support is included, how it's scoped, and what happens when you need a feature change six months after delivery.

Some vendors disappear after handoff. Others offer structured retainer models that let you continue iterating on the product without renegotiating a full engagement each time.

12. Have you worked in my industry or with my tech stack before?

Domain knowledge accelerates delivery. A vendor who has shipped AI-native software for logistics companies will ask different questions — and make fewer mistakes — than one adapting a generalist template to your use case.

Ask for a reference from a client in your vertical or with a similar stack. Call that reference. Ask them: "What did the vendor miss in the scoping phase?"

How to Score the Answers

After the vendor interview, rate each answer across three dimensions:

• Specificity: Did they give a concrete answer or a vague commitment?
• Contractual backing: Is it in writing, or only verbal?
• Risk transfer: Does the risk sit with you or with them if the answer proves incorrect?

Any vendor who answers more than three of the twelve questions vaguely deserves a follow-up round — or a pass.

What a Strong Vendor Answer Looks Like

Here's the contrast in practice:

Build Custom, Own It Completely

At Catalizadora, we answer all 12 of these questions in writing before any engagement starts — because we've watched clients arrive after a previous vendor left them with a licensed codebase they couldn't modify and a $15,000/month platform fee they couldn't escape.

Our engagements are structured for full ownership:

• Catalizadora Core — Full custom AI-native software, delivered in 12 weeks. 100% IP and source code transfer. No recurring license fees.
• Solo — Focused builds shipped in 15 days for a specific workflow or integration.
• Forge — Scoped by complexity for enterprise-scale or multi-system projects.

Every client owns their code. Every deployment is cloud-portable. And the team that pitches is the team that builds.

Ready to compare options? See our pricing and engagement models →