The promise of an AI agent is that it does work rather than describing it. That promise is real, but it holds for a narrower band of work than the demos suggest — and the projects that go badly are almost always the ones that got the boundary wrong rather than the technology.
So: where does the line sit?
What agents are genuinely good at
Three properties make a task suitable. It helps to insist on all three.
It is repetitive. The same shape of task, many times. Not identical — an agent handles variation fine — but recognisably the same job.
It is bounded. There is a clear definition of done and a limited set of things that can happen. "Chase the timesheets that are missing on Monday morning" is bounded. "Improve team productivity" is not.
The evidence is in your data. The agent can reach everything it needs to decide correctly. This is the one that quietly fails: if half the context lives in someone's inbox or in a conversation nobody logged, the agent is guessing, and a confident guess is worse than no answer.
Work that passes all three tends to look mundane. Chasing missing timesheets. Reconciling delivery notes against purchase orders and flagging the mismatches. Drafting the monthly client report from actual project figures. Raising a purchase requirement from the production plan. Watching every account for budget drift. Triaging inbound maintenance requests by urgency and routing them.
None of it is exciting. All of it is real hours.
What they are bad at — and should not be given
Judgement calls with consequences. Whether to extend credit to a customer. Whether to accept a late delivery or invoke a penalty. Whether this employee's absence pattern is a problem. These need accountability, and accountability needs a person.
Anything where being confidently wrong is expensive. Language models produce plausible output regardless of whether they have the facts. In a chat that is a minor annoyance; in a system that acts, plausible-and-wrong writes a bad record into your database and everything downstream inherits it.
Work whose context isn't written down. If the reason a rule exists lives only in the head of someone who has been there fourteen years, the agent will apply the rule as stated and get it wrong at exactly the edge case that mattered. Sometimes the useful outcome of an automation project is discovering these rules and writing them down — which is worth something on its own.
First contact with an unhappy customer. Technically feasible. Usually a bad trade.
Where the guardrails belong
The pattern that works in practice is narrower than "autonomous agents": most agent work should be proposing, not committing.
The agent assembles the purchase order and puts it in front of a person. It drafts the client report and marks the three figures it is least sure about. It writes the follow-up email and leaves it in drafts. The human step is a review, not a rebuild — and reviewing is dramatically faster than producing.
Three things make this safe:
- Explicit permissions. What it may do alone, what needs approval, what it may never touch. These belong in the same authorisation system as your human roles, not in a prompt.
- An audit trail. Every agent action logged as an agent action, with the evidence it used, and reversible. If you cannot tell agent writes from user writes, you cannot audit the automation.
- A confidence floor. Below a threshold, escalate instead of acting. An agent that says "I'm not sure, look at this" is more valuable than one that always produces an answer.
The goal is not an agent that never needs a human. It is an agent that knows which cases need one.
Why this needs one system
Every property above depends on the agent seeing a complete picture. An agent reconciling a delivery against an invoice needs purchasing, inventory and finance to mean something precise and connected. If those live in three products, the agent works from partial data — and partial data is where confident errors come from.
This is the practical reason agentic automation tends to arrive alongside consolidation rather than before it. Not because consolidation is fashionable, but because an agent cannot act reliably across boundaries it cannot see across. We unpack that distinction in what AI business management software actually is.
How to start
Pick the least glamorous task on the list. The one everyone complains about on Friday afternoon, that has a clear definition of done, and where a mistake is cheap and visible.
Run it in propose-only mode for a few weeks and measure how often a human changes the output. If they rarely change it, widen the authority. If they change it often, you have learned something valuable about a rule nobody had written down — fix that first.
What you should not do is start with the highest-value, highest-risk process. The first agent's job is to teach the organisation what agent work feels like and where its judgement fails.
Different operations start in different places — a retail chain usually starts with demand forecasting and re-ordering, a services firm with timesheets and reporting, a property manager with maintenance triage. You can see how that plays out per sector on the industries pages, or ask us directly via the contact form.
