For the tasks that need judgement, not just execution.

Reviewing a contract for red flags. Researching a prospect before a discovery call. Triaging a hundred inbound emails into urgent / standard / noise. Summarising a sixty-minute meeting transcript into decisions and actions.

None of these fits a rigid workflow. The output depends on what's in the input, and the input is different every time. Traditional automation breaks on the irregularity. But the task itself is repetitive enough that doing it by hand, over and over, is a waste of the most expensive hours in your business.

This is what AI agents are actually for. Not "an algorithm that decides things." A system that reads, applies a clear set of instructions, and produces a draft for a human to review.

Every agent I build follows the same skeleton, scaled to the task:

• Role, who the agent is and what it's authorised to do. "A senior commercial solicitor reviewing a software licence." "A research analyst preparing a one-page brief on a prospect."

• Context, what it has access to. The document, the transcript, the CRM record, the latest news search. Scoped narrowly, never given the whole estate.

• Constraints, what it can't do. No facts not in the source. No promises about outcomes. Flag ambiguity rather than guess.

• Output shape, the exact format the result lands in. A review note with three sections. A bulleted brief. A one-paragraph summary. Predictable enough that downstream systems and humans can consume it.

The agent never ships output unsupervised. The point is to remove the reading, thinking and first-draft work, not the decision. The decision stays with you.

Four use cases that recur:

• Lead research, an agent that reads a prospect's website, LinkedIn, recent news, and a few internal notes, then produces a one-page brief before every discovery call. What was a 20-minute pre-call scramble becomes a 30-second read.

• Email triage, an agent that classifies every inbound email by urgency and category, drafts suggested replies for the urgent ones, and surfaces the noise so it can be ignored. The inbox isn't empty, it's sorted.

• Document review, first-pass review of contracts, briefs, applications, against a checklist of red flags. The agent surfaces the issues; the human decides what to do about them.

• Meeting summarisation, transcript in, structured summary out: decisions, actions (with owners and deadlines), open questions, follow-ups. No more "did anyone take notes?"

Each of these is a single-purpose agent. You're not buying a platform, you're commissioning the specific tool you need, scoped to do its one job extremely well.

A good fit if:

• You have a recurring task that requires reading and judgement, not just execution.

• You can specify the criteria for what "good output" looks like (a checklist, a template, an example).

• You're happy to keep a human reviewing the agent's output before it ships externally.

• The cost of doing the task manually is currently real (hours per week, or quality slipping under load).

Probably not the right move if:

• The task is too varied to specify any rules around it (creative work, nuanced strategy).

• You're looking for the agent to make the final decision, not draft it for a human to approve.

• The volume is so low that the build cost outweighs the time saved.

Part of the wider AI implementation work I do. The other outcome lanes:

• Workflow automation, for repeatable multi-step processes.

• Data extraction, for getting structured data out of messy documents.

Working in a specific sector? For law firms · For agencies