What should a non-technical founder look for in an AI agent development company?

What is an AI agent, compared with a chatbot or a workflow?

A chatbot answers when someone asks. It waits for a message, replies, and does nothing in between. A workflow (also called an automation) runs the same steps in the same order every time: a form arrives, the data goes to a sheet, an email goes out. An AI step can sit inside it, such as summarising the form, but the path is fixed.

An AI agent decides its own steps. Anthropic, which makes the Claude models, draws the line this way: in a workflow the model and tools follow predefined code paths, while an agent directs its own process and chooses which tools to use¹. That freedom suits open-ended jobs where you cannot predict the steps. It also costs more and lets mistakes compound, which is why Anthropic recommends sandbox testing and guardrails¹.

For a founder the practical point is simple. Many jobs sold as agents are really workflows, and a workflow is cheaper, more predictable and easier to fix. A good company will tell you when you need the simpler thing.

Which kind of AI agent setup do I need?

Every agent proposal falls into one of three setups, even when the proposal does not name them. Ask which one you are being sold. The setup is set by what the agent's logins allow, not by what its instructions say, because a model can be talked out of an instruction but cannot get past a login that refuses.

• Read and report. The agent reads your chosen sources and writes only to its own output: a summary, a weekly report, a sorted list. Good for most first agents. The risk is a wrong summary that someone acts on.
• Read and draft. The agent reads and prepares drafts, such as reply emails or follow-ups, that a person reviews and sends. Check that the draft permission does not quietly include sending.
• Read and act. The agent sends, updates or pays by itself. OWASP calls the danger here excessive agency: more functionality, permissions or autonomy than the task needs². Keep it to narrow, reversible jobs with a hard limit, and have a person approve anything high-impact, as OWASP recommends².

How do I scope an AI agent in plain words?

You do not need technical words to scope an agent. You need one page that answers these, and the company should help you write it:

• The job: what the agent does, in one sentence, and how often.
• The sources: exactly which inbox, folder, sheet or system it reads, and nothing else.
• The output: where its work lands, and who looks at it.
• The setup: report, draft or act, and which actions always wait for a person.
• Done means: three or four real examples it must handle correctly before you accept it.
• Off switch: how you turn it off, and who you call when it misbehaves.

How do I judge an AI agent company's proof?

Ask to see it working, not slides. A demo on sample data you choose tells you more than any deck, because a deck shows the best case and a live run shows what really happens. Bring two or three ordinary examples from your business and one awkward one, and watch what the agent does with the awkward one.

Then ask to see the logs of that run: what the agent read, which tools it called, and what it produced. A company that logs every step can explain a mistake. One that cannot show logs cannot explain it either. Also ask for something the company built and still runs itself, and how it handles failures in that system. Case studies with no named, checkable detail count for little.

Who should own the code, prompts, accounts and data?

You should, from day one. The four things to name in the contract:

• Code. In a repository in your account, with copyright assigned to your company in a signed document. Indian law is covered in who owns the code; other countries have their own rules, so check locally.
• Prompts. The written instructions the agent runs on are part of the product. They belong in your repository, not only in the builder's head or account.
• Accounts. The AI provider account, the hosting and every tool login sit in your name, billed to you. Then the builder can be replaced without rebuilding.
• Data. The contract says which provider sees your data and on what terms. OpenAI, for example, says data sent to its API is not used to train its models unless you opt in, and abuse monitoring logs are kept for up to 30 days by default⁴. Ask for the equivalent statement for whichever provider is proposed.