Short answer
Pick the work that is repeatable, rule-based, reversible and measurable. Repeatable means it happens weekly or more. Rule-based means two people would do it the same way. Reversible means a mistake is caught and corrected before it reaches a customer. Measurable means you can prove the change in four weeks.
Key takeaways
- The first agent should be boring. Boring is what makes the second one possible.
- Anything that reaches a customer unreviewed fails the reversible test, whatever it saves.
- If two people on your team do the task differently, document it before automating it.
- Choose a workflow you can measure by week four, or you will be arguing about opinions in month three.
Owners almost always nominate the wrong first workflow. They pick the thing that annoys them most, which is usually the judgement-heavy work that only annoys them because it is hard. The right first pick is the dull, frequent task nobody defends, and that is the one that earns the trust to do the interesting ones later.
The four-question screen
Run every candidate through four questions before choosing. Does it happen at least weekly? Would two people do it the same way? If it goes wrong, is it caught before a customer sees it? Can you measure the difference within four weeks? Four yeses means start here. Three means it is a second or third project.
| Division | Typical first workflow | What stays human |
|---|---|---|
| Executive | Meeting notes into decisions, owners and dates | The decisions themselves |
| Administration | Intake, filing and scheduling handoffs | Exceptions and anything contractual |
| Marketing | Drafting from an approved brief and brand rules | The brief, the offer and final approval |
| Sales | Follow-up drafting and pipeline hygiene | The conversation and the pricing |
| Production | Quality checks against a written standard | The standard, and judgement calls on edge cases |
| Finance | Reconciliation prep and exception flagging | Approval, payment and anything a regulator reads |
| Quality Assurance | Reading output against the checklist | Deciding what the checklist should say |
Why reversible matters more than impressive
An agent that drafts a proposal for a human to approve can be wrong on a Tuesday and cost nothing. An agent that sends the proposal cannot. Start where a mistake is visible and cheap, prove the quality over several weeks, and only then move the approval point. The order protects both the customer and the project.
This is the part people skip when they have seen a demo. The demo always shows the send step, because the send step is the exciting one. In a live business the send step is the one that costs you a client.
Write the process down first, even badly
Before anything is built, have the person who does the task write out how they do it, including the decisions they make without noticing. Half the value shows up here: most teams find steps that exist only because someone left in 2019. Simplify the process, then automate what is left.
How to tell in four weeks whether it worked
- Hours: the time the task took before, and after, measured the same way both times.
- Cycle time: how long the work sits waiting, which usually improves before hours do.
- Error rate: how often output needed correcting, counted for both the human and the agent version.
- Adoption: whether the team still uses it in week four without being reminded.
Adoption is the one that predicts everything else. An agent that saves four hours and gets quietly abandoned saves nothing, and the reason is nearly always that nobody agreed who owns it.
Work to keep away from an agent for now
Anything where the rules are actually judgement: hiring decisions, clinical calls, pricing exceptions, and conversations where a relationship is at stake. Anything a regulator reads without a human signing it. And anything that happens twice a year, because you will never get enough repetitions to make it reliable.
Questions people ask
What is a good first agent for a small team?
Usually the one that turns meeting notes into decisions, owners and dates. It runs weekly, two people would do it the same way, a mistake is visible immediately, and you can measure the difference in a month. It also gets the team used to reviewing agent output.
Should I automate the task I hate most?
Not first. The task you hate most is usually judgement-heavy, which is exactly what agents are worst at and what fails loudest. Start with the frequent, dull, rule-based work, then use the credibility from that to take on the harder thing.
How do I know if a process is documented well enough?
Hand the written version to someone who has never done the task and watch them try. Every question they ask is a missing step. If they finish without asking, the process is ready for an agent to be built on.
Who should own the agent after it goes live?
A named person, not a department. They approve output, notice when quality drifts, and raise it when the underlying process changes. Agents that belong to everyone get abandoned in about six weeks.
What if my team is worried about being replaced?
Say what the agent will do and what they will do, in writing, before it is built. In our work people move from doing repeatable tasks to directing the agents that do them. Projects that skip this conversation fail on adoption rather than on technology.
Sources
- delegAIte Eight Divisions map, delegaite.co (read 20 September 2026): The seven divisions listed in the table, and Division 8 as the layer that runs them.
Published . Last reviewed .
Want this in your business?
Book a call and we will scope exactly which part of your operation an agent should take first.
Book a call with the team