- Guides
12 Questions to Ask an AI Agent Vendor Before You Sign
TL;DR: Before you sign with an AI agent vendor, get straight answers on four things: where your data lives, whether it trains the vendor’s models, what the agent can access and undo, and what happens when something goes wrong. Below are the 12 questions any honest vendor should welcome - with our own answers, because we should have to pass this test too.
What questions should you ask an AI agent vendor?
Ask twelve questions across four areas: data (where it lives, whether it trains the vendor’s models, who else sees it), control (what the agent can access, what gets logged, what can be undone), deployment (cloud versus on-premises, who owns the models), and the relationship (how pricing is scoped, what support and failure handling look like). Here’s the full list:
| # | Question | What a good answer sounds like |
|---|---|---|
| 1 | Where does my data live? | A specific environment you can point to - not “our secure cloud” |
| 2 | Is my data used to train your models? | No, full stop - or an explicit opt-in you control |
| 3 | Do you share my data with third parties? | No, with the boundary built into the architecture, not just the contract |
| 4 | Is there an audit trail? | Every action recorded, tamper-proof, reviewable by you anytime |
| 5 | What can each agent access? | Least privilege: only the data and tools its job requires |
| 6 | Can the agent’s actions be undone? | Yes - execution should have a safety net |
| 7 | What happens when the agent makes a mistake? | Detection, rollback, and a record of exactly what happened |
| 8 | Can I deploy on-premises? | Yes for regulated data - or an honest “no, we’re cloud-only” |
| 9 | Who owns the models? | Clear terms; ideally a model that can run on hardware you control |
| 10 | What happens when part of the system fails? | Workloads that recover and redistribute work on their own |
| 11 | How is pricing determined? | Scoped to your actual workflows, with a clear method |
| 12 | How do we start, and what does finding out cost? | A low-friction first conversation with no commitment |
Below, each group in detail - with our own answers on the record.
Where does my data live, and is it used for training?
Questions 1 through 3 decide everything else. If a vendor can’t tell you precisely where your data sits and who can touch it, stop there. Vague answers about “enterprise-grade security” usually mean your data flows through systems the vendor doesn’t fully control either.
Our answers:
- Where it lives: every agent runs in isolated containers on Google Kubernetes Engine (GKE), with models managed through Vertex AI. It’s a specific, inspectable environment - not a shared consumer product.
- Training: agents learn from results without ever sharing your private business data. Your data does not become someone else’s model.
- Third parties: no. Your information never leaves your environment without your say-so.
This is the core of what makes an agent private - the full breakdown is in what a private AI agent actually is.
What can the agent access, and what happens when it errs?
Questions 4 through 7 are about control. An agent takes real actions - posting content, sorting expenses, sending alerts - so “trust us” is not an acceptable governance model.
Ask for three things specifically:
- A tamper-proof audit trail. Every decision recorded so you can see exactly what happened and why - records you can review, replay, and understand. No black boxes.
- Least-privilege access. No agent should have more power than it needs. We treat this as its own discipline: our agentic security agents manage what data, tools, and systems each agent can access, and watch agent behavior in real time for unexpected actions.
- A safety net for mistakes. Agents will occasionally get something wrong - any vendor who says otherwise is the wrong vendor. In our architecture, every action has a safety net and can be undone, and the audit trail tells you exactly what to review.
Can I deploy on-premises, and who owns the models?
Questions 8 and 9 are about sovereignty and lock-in. Some data - patient records, contracts, regulated financials - shouldn’t pass through anyone else’s infrastructure. You also want to know the intelligence isn’t hostage to the vendor’s servers.
Our answers: cloud tiers run on Google Cloud with Gemini, and Enterprise deployments run on Google Distributed Cloud (GDC) or your own cloud, with Gemma 4 - Google’s open model - running entirely on your own infrastructure. The model, your data, and every inference stay on hardware you control. We compared the two paths honestly in cloud vs. on-premises AI agents; for most small teams, on-premises is more than they need, and a vendor who pushes it on everyone is selling hardware, not fit.
What happens when things fail, and what does it cost?
Questions 10 through 12 cover the relationship. Systems fail; the question is what happens next. Our platform is self-healing: if something goes wrong, agents redistribute work and keep running, with no single point of failure. Support scales by plan, from standard up to 24/7 enterprise support.
On pricing, be wary of one-size-fits-all quotes for agent work - the honest version is scoping against your actual workflows. We don’t publish flat prices; every plan is scoped in a free consultation, so you get a concrete recommendation before committing to anything.
When is this checklist overkill?
Honestly, sometimes. If you’re evaluating a chatbot to answer public FAQ questions, most of this list doesn’t apply - there’s no sensitive data or autonomous action to audit. If you’re piloting a plain SaaS tool on non-sensitive data, standard contract review may be enough. And if you have in-house engineers who want full control, building it yourself sidesteps vendor due diligence entirely - you just trade these twelve questions for hiring and maintenance ones.
Use the full list when an agent will touch data you’d never email to a stranger, or take actions you’d want to trace afterward.
If you’d like to run all twelve questions against us in person, book a free consultation - it’s a 45-minute call, we’ll answer every one, and we follow up within 24 hours.