To choose an AI agent development company, ask how they limit what the agent can do, how a person approves risky actions, how they test on your real data, who owns the code and accounts, and who will actually do the work. A good team answers plainly, shows working examples, and is happy to start with one task.
The ten questions below go into each of those in more detail. After them you will find the answers that should worry you, a real example of good answers in practice, and how we answer the same questions ourselves.
What should an AI agent actually do for your business?
An AI agent is software that reads a request, decides which tools to use and completes a task step by step. Unlike a chatbot, it takes actions: it updates a record, drafts a reply, books an appointment or creates a task. That is what makes agents useful, and it is also why choosing who builds one matters.
Before talking to any vendor, write down the one task you want the agent to handle, the tools it would need to touch, and what should happen when it is unsure. Vendors who are worth hiring will ask you about exactly these things. You can see the kinds of agents we build on our AI agents page.
What are the 10 questions to ask?
The ten questions fall into four groups: what the agent may do, how people stay in control, who owns what, and who does the work. Ask all ten in your first conversation. A team that has built real agents will answer each one quickly and specifically.
1. What exactly will the agent be allowed to do on its own?
Ask for a plain list of actions the agent can take without a person, and a list of actions that always need approval. If the answer is vague, the agent will be too.
2. How does a person approve risky actions?
Anything involving money, contracts or customer messages should wait for a person. Ask to see what the approval looks like: who gets it, where, and what information they see before they decide.
3. What happens when the agent is unsure or wrong?
A good agent hands uncertain cases to a person instead of guessing. Ask how the agent decides it is unsure, where those cases go, and how mistakes are traced and fixed afterwards.
4. How will you test it before it goes live?
The answer should involve your real data: real emails, real documents, real calls, including the awkward ones. Ask how many examples they test on and whether the test set is kept so it can be run again after every change.
5. Is every action logged?
You should be able to see what the agent received, what it decided, what it did and who approved it. Without a log, you cannot audit it, debug it or trust it.
6. Who owns the code, the accounts and the API keys?
They should be yours from the first day. If the model accounts and keys are in the vendor’s name, your costs, your data and your ability to switch vendors are all in their hands.
7. Which tools and models will you use, and why?
You want a clear answer, not a sales pitch. Ask which language models they plan to use, which automation platform or framework, where memory and logs are stored, and why those choices fit your task. Ask how they connect to your existing systems.
8. Who will actually build and support it?
Ask whether the person you are talking to will be involved in the build, and who you contact when something needs changing. Being handed from a salesperson to an unnamed team is a common source of frustration.
9. Can you show a working agent you have built?
Case studies and demos tell you more than a slide deck. Ask for an example close to your use case, and ask what the hardest part of that project was. Real builders have real answers.
10. How do we start small?
A sensible vendor suggests one task first, a working version on your data, and a decision point before anything larger. Be cautious of anyone who proposes a long programme before you have seen anything work.
Which answers should worry you?
Some answers are warning signs. They do not prove a vendor is careless, but they do suggest the agent will behave in ways you did not expect, and in software that takes actions on your behalf, unexpected behaviour is costly. Be careful if you hear any of these:
- The agent will “handle everything” with no clear list of limits.
- Approval steps are described as optional extras.
- Testing will happen “in production” or on made-up examples.
- The accounts and keys will be held by the vendor.
- Nobody can tell you who will write the code.
- Big promises about speed or savings with no example to back them up.
What do good answers look like in practice?
Here is a real example. We built an AI voice agent for a service business that needed calls answered, leads qualified and appointments booked without someone staffing the phone at every hour. The agent answers the call, asks the qualifying questions, books a slot directly into the calendar and records each conversation and lead status in the CRM.
When a caller needs something the agent should not handle, it hands off to a person.
In terms of the questions above: the agent’s allowed actions were limited to qualifying and booking (question 1), anything outside that went to a person (questions 2 and 3), every call and outcome was logged in the CRM (question 5), and the stack was chosen for the task: a voice platform for the calls, a workflow tool for the logic, and a database for records (question 7). You can read the full AI voice agent case study.
Agents work best alongside ordinary automation, where fixed steps run on rules and the agent handles only the parts that need judgment. Our guide to what AI workflow automation is explains how the two fit together, and construction teams can see six workflows contractors automate first.
How do we answer these questions?
As an AI agent development company, here is how we answer the ten questions. Every agent gets a written list of what it can and cannot do alone. Money, contracts and customer messages wait for a person. Uncertain cases go to a person and every action is logged.
We test on your real examples and keep the test set for later changes. The code, accounts and keys are yours from day one. You talk to the engineer who builds it. And we start with one task, so you see it working before deciding on the next.
If you want to put these questions to us directly, our page on custom AI agents with approval steps has more detail, and how we work explains the process from first call to handover.