The Ten Jobs an AI Agent Is Actually Good At
Vishal Sachar
Co-Founder & CEO of CLRT
Most people ask the wrong question about AI. They ask what it can do for their industry, as though a law firm and a logistics company and a clinic each need a different kind of intelligence. They do not. An AI agent does not see industries. It sees jobs, and the same handful of job-shapes recur in every business under different names.
A law firm reviewing a contract, a bank running a know-your-customer check, and a clinic processing a prior authorisation are, to an agent, the same job: pull structured facts out of a document and test them against a set of rules. The clothes differ. The shape is identical. Once you learn to see the shape, the industry stops mattering, and you can finally tell which of your work is ready for an agent and which is not.
There are roughly ten of these shapes, and almost every business is full of them. Triage and routing, sorting incoming emails, tickets, and applications to the right place. Extraction, pulling structured data out of unstructured contracts, invoices, and forms. First-draft generation, producing the reply or memo a human then finishes. Research and synthesis, gathering across many sources and condensing them into something usable. Reconciliation and checking, comparing two things and surfacing the mismatch or the exception. Monitoring and alerting, watching a stream of activity and flagging only the thing that needs a human. Following up, the persistent, multi-step chasing that drops no item until it is resolved. Answering from your own knowledge, responding to questions from your documents, policies, and history. Multi-step coordination, orchestrating a process across several systems and people end to end. And classification and tagging, labelling things at scale by a consistent rule.
The strategic move is not to ask which of these sounds exciting. It is to look at your business and notice which job-shape you are drowning in, and then to ask a second question that decides where to start safely: of the shapes you have in volume, which ones produce an answer you can check.
The best first agent is the job you have most of, where being wrong is cheapest to catch.
A deeper dive
Underneath, all ten reduce to the same loop, retrieve the relevant context, reason about it, take an action with a tool, check the result, repeat, which is why one architecture serves them all. What separates the safe ones from the dangerous ones is verifiability. Reconciliation, extraction, and classification are deterministically checkable: the output is right or wrong against a source, so you can verify mechanically and automate aggressively. Drafting, research, and synthesis are judgment-laden: the output is plausible-or-not rather than right-or-wrong, which puts them squarely in the zone where AI is most dangerous precisely because it looks most competent, the argument in Where AI Almost Works Is Where It Hurts You. The practical sequencing falls straight out of this. Automate the verifiable, high-volume shapes first, because they pay quickly and fail visibly. Put the judgment-laden shapes behind a human or a separate checker. The reason this matters more than picking the cleverest model is that the job-shape, not the model, determines whether the work is safe to hand over at all.
Key terms
- Triage and routing
- Sorting incoming emails, tickets, and applications to the right place.
- Extraction
- Pulling structured data out of unstructured contracts, invoices, and forms.
- First-draft generation
- Producing the reply or memo a human then finishes.
- Research and synthesis
- Gathering across many sources and condensing them into something usable.
- Reconciliation and checking
- Comparing two things and surfacing the mismatch or the exception.
- Monitoring and alerting
- Watching a stream of activity and flagging only the thing that needs a human.
- Following up
- The persistent, multi-step chasing that drops no item until it is resolved.
- Answering from your own knowledge
- Responding to questions from your documents, policies, and history.
- Multi-step coordination
- Orchestrating a process across several systems and people end to end.
- Classification and tagging
- Labelling things at scale by a consistent rule.
Work with CLRT
The hard part is not building the agent. It is seeing which of the ten jobs you are full of, and which is safe to start with. That is precisely what a CLRT diagnostic identifies, in your business, in dirhams. That is where to begin.
Vishal Sachar
Vishal Sachar is the Co-Founder and CEO of CLRT, where he helps UAE businesses make sense of applied agentic AI and put it to work. He writes on agentic systems, AI governance, and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.
