How to Choose Your First AI Agent
The best first AI agent isn't the most impressive one. It's the one you can finish, measure and trust.
Almost every team starts with the wrong question: "What's the most impressive thing AI could do for us?" The better question is: "What's the most useful thing AI could do for us that we can actually finish?"
Your first agent has two jobs. It has to deliver value, and it has to prove to your team that AI can be trusted with real work. A flashy but fragile first project can set the whole effort back.
The five-question filter
Score each candidate workflow on these five questions.
| Question | Good sign | Warning sign |
|---|---|---|
| 1. Frequency: How often does it happen? | Daily or weekly | A few times a year |
| 2. Clarity: Can you describe the steps and what "done" means? | Yes, in plain language | "It depends who does it" |
| 3. Data: Is the information it needs accessible? | In one or two systems with access | Scattered across people's heads |
| 4. Stakes: What if it's wrong? | Mistakes are easy to catch and fix | Money, health, legal or reputation at stake |
| 5. Measurability: Can you tell if it's better? | Time, speed or error rate is trackable | "We'd just know" |
A worked example
A services company is deciding between three ideas.
- Replying to common customer questions. Frequent, clear, data in a help centre, low risk, easy to measure. Strong.
- Negotiating enterprise contracts. Infrequent, ambiguous, high stakes. Poor first choice.
- Building the weekly sales report. Weekly, clear, data in the CRM, low risk, measurable. Strong.
They choose the support agent first, in "suggest only" mode, then the reporting agent second.
Red flags
- "We want AI to run the whole department." Too broad. Break it into workflows.
- "We don't have the data written down anywhere." Fix that first (see our data-readiness checklist).
- "Nobody owns the process." Agents need a human owner.
- "It has to be perfect before launch." Plan for supervised launch and improvement instead.
What to do next
- 1List five candidate workflows.
- 2Score them with the five questions.
- 3Pick one. Name an owner. Decide your success measure.
- 4Start in a supervised mode and widen autonomy as trust grows.
Have a workflow in mind?
Tell us about it and we'll show you where AI could help.
