AI agents are currently being sold the way most new technology gets sold, which is as a replacement for whatever you are already paying for. The pitch is that an agent will handle your support, your sales follow up, your reporting and your operations, and that your headcount can shrink accordingly.
The honest picture is narrower and considerably more useful. Agents are genuinely good at a specific shape of work. Outside that shape they are expensive, unpredictable, and they fail in ways that cost more to clean up than the work would have cost to do properly.
What an agent actually is
Strip away the marketing and an agent is a language model that has been given tools and permission to use them in a loop. It can read a request, decide which tool to call, look at the result, and decide what to do next. The difference from a chatbot is that it can act rather than only answer.
That capability is real. It is also the source of the risk, because an agent that can act can also act wrongly, at speed, without anyone watching.
Where agents earn their cost
The work that suits an agent has three properties: the inputs are reasonably structured, the task repeats often enough to matter, and a mistake is recoverable.
Triage and routing. Reading inbound messages, classifying them, and putting them in front of the right person with a summary attached. The human still decides. The agent removes the sorting.
Drafting from a source. Turning a call transcript into a structured note, a set of requirements into a first draft specification, or a data export into a written summary. A person edits the output, which is far faster than starting from nothing.
Data cleanup and enrichment. Normalising messy records, matching entries across systems, filling gaps from a known source. This is work nobody enjoys and where consistency beats judgement.
Answering from your own documents. Internal questions where the answer already exists somewhere in your files and the cost is finding it. This works well when the agent is restricted to your material and made to cite where an answer came from.
Where they waste money
Agents struggle where the task requires judgement about consequences, where the inputs are ambiguous, or where being confidently wrong is expensive.
Handing an agent unsupervised authority over anything financial or contractual is the clearest example. So is putting one in front of customers with the ability to make commitments about price, delivery or refunds. The failure mode is not that it refuses. It is that it answers plausibly and incorrectly, and the customer reasonably believes it.
The second common waste is building an agent for a process nobody has bothered to define. If a task is currently done differently by four people, an agent will not resolve that. It will encode one version of it and make the inconsistency harder to see.
The questions to answer before you build
- What exactly is the task, and can you write down the steps a competent person follows today?
- How often does it happen? Under a few times a week and the payback is usually poor.
- What happens when it gets it wrong, and who notices?
- What does it need access to, and what should it be prevented from touching?
- How will you tell whether it is working, in numbers, three months from now?
If the last two have no answer, you are not ready to build. Access boundaries and a success measure are not paperwork, they are the difference between a tool and a liability.
Keep a human in the loop, deliberately
The most reliable pattern we see in production is not full automation. It is an agent that does the preparation and a person who approves the action. The agent drafts the reply, assembles the quote, or proposes the classification. A human presses send.
That arrangement captures most of the time saving while keeping the failure modes visible. It also gives you a natural path to widening autonomy later, once you have evidence about where the agent is reliable and where it is not.
Frequently asked questions
Do I need my own model?
Almost certainly not. For nearly all business tasks the sensible route is a capable commercial model with your own data and tools connected to it. Training your own is a large undertaking that rarely pays back for ordinary operational work.
What about our data?
Decide early what the agent may read and what it may never see, and enforce that in the system rather than in the prompt. Check your provider's terms on whether your inputs are used for training, and make that an explicit procurement question rather than an assumption.
How do we start without a large commitment?
Pick one bounded task that happens daily, keep a person in the approval seat, and measure the time saved for a month. A narrow deployment that works will teach you more than a broad one that stalls.
A sensible first step
Start with the work that is repetitive, well understood and low consequence, and expand only where you have evidence. The businesses getting real value from agents are not the ones that automated the most. They are the ones that picked carefully and kept a person accountable for the outcome.
We build this kind of automation into the systems we deliver rather than bolting it on afterwards. If you have a process in mind, describe it to us and we will tell you honestly whether an agent is the right answer.
