AI agents
AI agents is the word of the year and it is doing too much work. Depending on who is talking it means a chatbot with a nicer name, a model that can call a function, a system that acts unsupervised, or a research demo. The four have almost nothing in common except the word, and the gap between them is where budgets die.
The four things people mean
- A chatbot, renamed. Answers questions from a fixed set of sources. Useful, mature, and not an agent by any definition that means anything. Most of what is sold as agentic is this.
- A model that can call your systems. It reads a request, picks the right function, and comes back with the result. This is the one that works today, and almost every genuinely useful build is this shape.
- Something that acts without asking. It decides and does, then reports. This works inside a narrow, well-fenced job and fails outside one, and the fencing is most of the engineering.
- A demo. Several models talking to each other, planning, self-correcting. Impressive on stage. What it does on a Tuesday with your data is the question nobody in the demo answers.
What separates the two that work from the two that do not
The buildable ones share three properties: the job has a definite end, there is a way to tell whether the result was right, and the systems it touches can be read from and written to. Take away any one and the thing becomes a demo — usually the middle one, because measuring whether an answer was correct is harder than producing it.
What the word agentic adds, and what it hides
Agentic AI names a property, not a product: the software picks its next step instead of following a written one. That is a real distinction and it is not the one that decides a budget. The same build — a model that reads a request and calls the right function — is sold as agentic AI by half the market and as a chatbot by the other half, and the price differs by a factor nobody can justify from the code.
The question that actually separates them is permission. How far may it go before a person signs, what is it never allowed to decide alone, and what happens on the run where it stops. Those three answers get written by your business before any code exists, and a vendor who cannot state them about their own product is describing a demo with a better word attached.
The fortnight that settles it
- Run it beside the person first. It does the job, a person does the job, and the two are compared for a fortnight. This costs a fortnight and it settles arguments no demo can.
- Give it a written list of things it may never do. Written before it is built, tested as part of the build, and checked on every run. A never-do list discovered afterwards is an incident report.
- Make it explain itself. Every action with the reason recorded next to it, in a form a person who was not there can read. Without this you cannot debug it and you cannot defend it.
- Decide who owns it on a Tuesday. Not who built it — who is called when it does something odd, and who is allowed to turn it off. An agent with no owner is an incident with a delay on it.
When the answer is not an agent
- The job is a rule and nothing has to choose. A decision table is cheaper, faster, exactly repeatable and explicable to an auditor. The distinction is not whether a schedule is involved — it is whether anything decides.
- Nobody can say what a correct outcome looks like. Then you cannot evaluate it, which means you cannot improve it and you cannot know when it breaks.
- The systems it would need have no way in. An agent that cannot reach anything is a chat window, and you can buy one of those.
- The reason for it is that the board asked for AI. That is a real reason to do something, and it is not a reason to let software act unsupervised on your data.
Four of these running, each fenced differently
- Customer-chat assistant — the narrowest useful shape: one job, one set of systems, and a written list of what it may never say. It has been answering buyers in a working business every day for months, which is a different test from a demo.
- Promotion guard — runs on a schedule and acts on a rule, and the opt-out still happens on the owner's say-so rather than behind their back — which is what a fenced agent actually looks like in production, as opposed to in a demo.
- Returns & defects handler — judges each case against written rules and records why, which is the version an auditor can read.
- Text factory — an agent that runs a whole job — research, draft, fact-check, two languages — with the checking built into the machine rather than bolted on after, and a person still approving every publish.
What people ask once past the demo
What are AI agents, in one sentence?
Software that decides what to do next rather than following a fixed path — which in practice means a model choosing which of your systems to call, with a written boundary on how far it may go before a person is involved.
Are AI agents actually working in production anywhere?
Yes, in narrow jobs with clear boundaries. What is not working is the broad autonomous version from the demos, and the honest signal is that the working ones are boring to describe: one job, one set of systems, one written list of things it may never do.
Do we need a multi-agent system?
Almost certainly not, and the question is usually a sign the first one has not been built yet. Several models talking to each other multiplies the ways it can fail without multiplying what it can do. Get one narrow job right first, then find out whether the second one needs to talk to it.
How is this different from the chatbot we already have?
A chatbot answers. An agent acts — it books, updates, pays, refuses. That difference is entirely about permission and consequence, not about the model, and it is why the interesting work is the boundary rather than the intelligence.
What does one cost to build?
The build is quoted per project and fixed up front. What actually drives the number is the state of your data and whether your systems can be written into — the same job costs a fraction with clean inputs and multiplies when three systems disagree about the same record.
What are some real examples of AI agents, not demos?
The four listed above are running in a working business rather than on a stage: a customer-chat assistant with a second model blocking anything it should not say, a guard that pulls products out of forced discounts on the owner's word, a returns handler that judges each claim against written rules, and a text factory that researches and drafts in two languages with a person approving every publish. Examples of AI agents are worth reading for their fences, not their cleverness — each of those is one narrow job with a written list of things it may never do.
Name the job, not the technology
Try the sentence out loud: this thing should do X, and I will know it worked because Y. If the second half is hard to finish, stop — that is the finding, and it is worth more than any prototype, because nothing that cannot be evaluated can be improved. If it comes out easily, the shape of the build is already half decided. The first conversation is an hour and it is free.
Book that hour →