Questions to Ask Any AI Agent Vendor

If you've sat through more than one "AI agent" pitch as an SME owner, you've probably noticed they all sound roughly the same. Confident demo, a model name you may or may not recognize, a promise it'll save your team hours a week. What almost none of them tell you, because it's the unglamorous part and it doesn't fit in a five-minute demo, is what happens around the model: what it's allowed to do on its own, what it remembers, and how anyone would actually find out if it started behaving badly. That surrounding system is what we and others in the field call a harness, and it's the actual thing separating a vendor who's built something you can depend on from one who's connected a model to your CRM and is hoping for the best.

Why this is genuinely hard to evaluate from the outside

This isn't a knock on business owners who find AI agent pitches hard to compare, it's a genuinely hard thing to evaluate without a technical background, and vendors know it. Two pitches can use identical language, "AI agent," "automation," "connects to your existing tools," while one has spent real engineering effort on failure handling and permissions and the other has a working demo and nothing else. From the outside, in a single sales call, they can look indistinguishable. The gap only becomes visible once something goes wrong in production, which is exactly the wrong time to discover it.

Think about how differently this works when you're buying almost anything else for your business. A commercial oven, an accounting platform, a point-of-sale system, all come with specifications you can compare on paper, or at minimum a track record you can ask other customers about. "AI agent" has no equivalent yet. There's no standard spec sheet, no widely agreed benchmark a small business owner could reasonably check, and "trust me, it works" is doing more of the selling than it would for almost any other purchase of similar cost. That gap is precisely why a short, direct set of questions is worth more here than it would be for a purchase where the specs already do that work for you.

The five-minute-demo trap

It's worth naming why this gap exists, because it explains almost every disappointing AI rollout you've heard about secondhand. A demo only has to work once, in front of an audience, on inputs the presenter already knows will behave well. Real use has to work correctly on the hundredth run, the thousandth, including the malformed input, the system that times out, the edge case nobody thought to test beforehand. Almost everything that closes that gap is invisible in a five-minute demo, which is exactly why it's so easy for a vendor to sell around it, intentionally or not. The questions below are aimed specifically at surfacing what a demo can't show you.

Question 1: What happens when a step fails partway through?

Real work is rarely one action, it's a sequence: read an enquiry, draft a response, update a record, send a confirmation. Ask what happens if step three of five fails. A serious build picks up exactly where it left off, or fails cleanly with a clear record of what did and didn't happen. A shallow one either stops silently, leaving you to discover the gap later, or worse, restarts from the beginning, potentially repeating actions it already completed, double-sending a message or double-processing something it had already handled. If the answer is vague, hedged, or "it usually works fine," that's itself the answer.

Question 2: What is it allowed to do without asking you first?

An agent that can only draft a reply for a human to review is a fundamentally different risk than one that can send messages, move money, or change records entirely on its own. Neither is automatically wrong, but you should know which one you're buying, and specifically, which actions are unsupervised versus which ones pause for a human. A vendor who can't answer this precisely, tool by tool, action by action, likely hasn't actually built that boundary into the system; they've just described it in a slide.

Question 3: Does it remember anything, or does every conversation start from zero?

A lot of what gets marketed as AI "memory" is really just a longer context window within a single conversation, which forgets everything the moment that conversation ends. Ask specifically whether it remembers something across separate sessions, days apart, with no shared conversation history, and ask how you'd know that's actually true rather than a claim on a feature list. Genuine cross-session memory is a meaningfully harder thing to build than it sounds, and it's worth confirming it's been tested, not just described.

Question 4: How would you find out if it started behaving strangely?

An agent that runs for months accumulates behavior over time in ways a one-time demo or a single QA pass can't catch: a tool that starts failing intermittently, a pattern of confused or repeated actions, a task quietly stalling instead of finishing. Ask what actually watches for that on an ongoing basis, not just what was checked before launch. "We'd notice if a customer complained" is not a monitoring strategy, it's an admission that the monitoring strategy is your customers.

Question 5: Who's actually accountable when it gets something wrong?

This is less a technical question than an organizational one, and it matters just as much. Is there a specific, reachable person who built and maintains this system, or does an issue disappear into a support ticket queue and a rotating cast of account managers? A vendor confident in what they built usually answers this one immediately and specifically. A vague answer, "we have a great support team," "someone will get back to you," tends to mean the accountability, like the harness itself, hasn't really been built yet.

Phrases worth pausing on

Beyond the five questions themselves, a handful of specific phrases in a pitch are worth treating as a prompt to dig deeper, not necessarily a dealbreaker on their own, but a signal that the harder half of the build might not have happened yet.

"It's powered by [model name]", offered as the entire answer to how it works, with nothing said about what happens around the model
"It basically never fails", which is a claim about a system that's either never been genuinely stress-tested or is being described optimistically rather than accurately
"We'll add proper error handling / permissions / logging after launch", which means you'd be the one finding out what's missing, in production, with real customers
"Our AI handles that automatically" as the answer to who's accountable when something goes wrong, which quietly removes the actual human answerable for a mistake

None of these phrases automatically disqualify a vendor, plenty of good conversations start with an honest "we haven't built that part yet, here's the plan." What's worth noticing is the difference between that honest version and one of these phrases used to paper over the gap instead of naming it.

A vague answer is itself the answer

Notice that none of these five questions are really about which model is running underneath, and that's deliberate. Model choice is the least differentiated part of any AI pitch at this point; most serious vendors can reach for a broadly similar tier of capability. What actually varies, wildly, from one vendor to the next is everything these five questions are aimed at: failure recovery, permission boundaries, memory, monitoring, and accountability. If a vendor answers all five specifically and confidently, you're likely looking at something that's been genuinely engineered. If they redirect back to how capable the model is, or how impressive the demo was, that redirection is worth noticing, because it usually means the harder, less flashy half of the build, the harness, isn't there yet.

How we'd answer these about our own build

We hold ourselves to the same five questions before we'd pitch anything to a client, which is why we've written each answer up in detail rather than just asserting it. Failure recovery and duplicate-action prevention are covered in how our own agent asks before it acts and never acts twice. Permission boundaries show up throughout that same piece and in how we scoped its own private database. Memory, including how we verified it actually survives across sessions and stays bounded over time, is covered in how we built memory that doesn't grow forever. Ongoing monitoring is the entire subject of how we watch our own agent for warning signs. And the full picture of how all of it fits together sits in the case study on building the harness itself. We'd rather a prospective client read the actual engineering than take a claim about it on faith, which is the same standard this piece is asking you to hold any vendor to.

What it actually costs you to ask

Five questions, one call, and a willingness to notice when an answer is vague instead of letting confidence stand in for specificity. That's the entire cost of this filter, and it's cheap relative to what it protects against: months of quietly unreliable automation, or worse, a customer-facing failure that could have been caught by asking, upfront, what happens when something goes wrong. AI agents genuinely are ready for SMEs to rely on, the underlying technology has gotten good enough for that to be true. "Built properly" is doing the real work in that sentence, and these five questions are the fastest way we know to find out, before you sign anything, whether the vendor in front of you has actually done that work.

If you're at the stage of working out where AI genuinely fits your business in the first place, that's a separate, earlier conversation worth having before any of these five questions are even relevant, and it's the starting point of our own AI Solutions & Automation work: figuring out honestly where AI will save real time or money, and where it won't, before anyone builds anything.

See how we approach AI Solutions & Automation for Singapore SMEs ↗