AI Agents · Basics
What Agents Are Genuinely Good At Today
The shapes of work where agents earn their cost, described by property rather than by industry — so you can recognise yours.
Most lists of agent use cases are by industry, which is not much help — your work is not "retail", it is a specific set of tasks. This one is by property, so you can recognise the shape.
The properties that make a task suit an agent
Variation that defeats a script. Each instance differs enough that fixed steps break, and not so much that there is no pattern. This is the essential one — without it, automation is better. See when an agent is the wrong answer.
For a vendor-side perspective, Monitask also has a page on task switching cost.
A checkable answer. Someone can tell whether it was right, reasonably quickly.
Recoverable errors. A wrong result costs a review rather than a customer.
Reachable, clean data. See integration is the hard part.
Enough volume to amortise setup and oversight.
A shape a person could describe in a few sentences, even if they could not write the steps.
Where these properties cluster
Reading and extracting
Pulling structured information out of unstructured documents — invoices, contracts, reports, correspondence.
Why it suits: the documents vary, the target fields are fixed, the answer is checkable against the source, and errors are visible.
The commonest genuinely working deployment.
Triage and routing
Sorting incoming things into categories — tickets, emails, applications, requests.
Why it suits: high volume, variation that resists rules, and a misroute is corrected rather than catastrophic.
And it can act as preparation rather than decision: classify and summarise, a person handles.
Research and synthesis
Gathering from several sources and organising it.
Why it suits: the search path differs every time, and the output is checkable if sources are cited.
Caveat: verify the sources. Plausible references that do not exist are a known failure.
First drafts
Proposals, replies, summaries, documentation, code.
Why it suits: a person reviews everything, so the oversight question answers itself, and starting from something beats starting from nothing.
This is where most individual use actually sits. See agents for a one-person business.
Software work
Writing, modifying and testing code.
Why it suits: tests provide external verification, version control makes errors recoverable, and the work has enough structure to be checked automatically.
The most mature area by some distance, and the one where autonomy is furthest advanced.
Reconciliation and cross-checking
Comparing records across systems and flagging mismatches.
Why it suits: tedious, high volume, and the flag is checked by a person rather than acted on.
Answering from a defined body of knowledge
With retrieval, over documentation, policies, product information.
Why it suits: the questions vary, the source is fixed, and answers can cite where they came from. For broader independent background, see NIST AI Risk Management Framework.
Borderline agent — much of this is retrieval plus generation rather than a loop. Which is fine, and worth knowing so you buy the right thing. See vocabulary.
Monitoring and alerting with judgement
Watching for conditions that resist a simple rule, and escalating.
Why it suits: the pattern is fuzzy, and the output is a flag rather than an action.
The pattern across all of these
Notice what they have in common.
Most produce something a person then uses, rather than completing a transaction unsupervised.
Most have external verification available — a source document, a test, a second system, a human review.
Most are preparation rather than execution.
That is where agents currently earn their cost, and deployments that skip past it into unsupervised action are where the incident data comes from. See where agent projects fail.
Where it is heading
Autonomy is extending gradually, in areas where verification is strong — software being the clearest.
The constraint is not model capability. It is verification, integration and the compounding of small error rates across long chains. See what agents are still bad at.
Expect this list to widen slowly rather than transform. And check the date on this article.
The short version
The essential property is variation that defeats a script — without it, automate instead.
Reading and extracting, triage, research, first drafts, software work, reconciliation. Those are where the properties cluster today.
Most working deployments produce something a person then uses, rather than completing transactions unsupervised.
External verification is what makes autonomy safe — which is why software work is furthest ahead.
And the list widens with verification, not with model capability.