PRACTICAL GUIDE
Choose one process before adding an AI agent
Start with a task you can explain on one page. Give it an input, an owner, a review point and a clear output before choosing tools.
Where I would start
When a task feels messy, write down what a good handoff would contain. For a quote request, that might be the requested item, quantity, missing specifications and the source for each detail. That gives you something concrete to review.
My proposed first test is deliberately small: one draft brief, checked by a person, with no customer message sent. The map below is a planning aid. It is not a tested product or a report of client results.
A six-part process map
Follow the request from input to output. Tools support the steps; the owner and approval point tell you who remains responsible.
- Input
- One customer request, with its source and date.
- Steps
- Extract the request, find missing details, prepare a review brief.
- Tools
- An approved inbox export and product reference. No live systems needed for the first test.
- Owner
- One named person who checks the brief.
- Approval
- A person verifies product fit, price, availability and any reply.
- Output
- A traceable draft brief, with unknowns still marked.
A fictional distributor inquiry
Illustrative example only. This is not a customer request or a completed experiment.
A buyer asks for 20 replacement pumps by Friday but provides no model number. The assistant’s permitted task is to prepare an internal brief from that message and an approved reference sheet.
- Known: quantity 20; requested delivery Friday. Keep the original request attached.
- Unknown: model, operating requirements, shipping address and whether Friday is possible.
- Draft next step: ask the buyer to confirm the missing specifications.
- Human check: verify suitability, current stock, pricing and delivery before preparing a quote or sending anything.
A useful draft preserves the unknowns. A confident-looking model number that came from nowhere makes the review harder and could lead to a wrong quote.
The same map in other teams
For a SaaS campaign, start with an approved brief and end with draft copy whose claims and audience are checked by a marketer. For a consulting project, start with consented notes and end with a research brief that separates client statements, sources and open questions. These are examples to explore, not evidence of demand or results.
Run a small, honest comparison
- Choose a few fictional or approved requests and record how many you use.
- Complete the brief manually first. Save the output and record time spent.
- Use the same requests and instructions with the assistant. Record the tool, model if visible, date and access limits.
- Have a reviewer check both versions against the same criteria: completeness, unsupported facts, missing details and corrections needed.
- Compare accepted drafts, review minutes and recorded costs. Keep failures in the report.
Do not turn a small example into a claim about company-wide savings. A brief can be faster to produce but take longer to check. Token usage belongs in the comparison only when the records are available and comparable.
A checklist you can copy
Task and reason:
Approved inputs and sources:
Steps and tools:
Owner and reviewer:
Actions requiring approval:
Definition of an acceptable output:
Stop and ask when:
Baseline, test date and sample size:
Observed result and remaining questions:
Why start this way?
Anthropic distinguishes fixed workflows from agents that choose their own steps and tools, and recommends starting with the simplest approach that fits the task. Its December 2024 guide now notes that tooling has changed. I use that architectural distinction here, not its older tool examples as current buying advice. Read Anthropic’s Building effective agents.
The map, fictional example and comparison checklist above are proposed teaching tools. We still need to run the test before reporting what worked.
What process would you try first?
Curiosity Insights by Azhar shares practical ideas, experiments and the questions behind them.
Get Curiosity Insights