Home · Start here · Two-week trial plan

Test an AI tool on your own tasks before your team pays for it

Pick three real tasks, write what pass means before you start, and keep a one-page result that says which tool passed each task.

Method · about 8 min here, then two weeks on your own work · last checked 2026-11-03

Example: made up to show the idea.

Scripted, not a live model. Nothing connects, and nothing leaves this page. Harbour Lettings, its team and its trial are made up for this example.

Pay for seats only where a tool passes your own tasks

A seat is one person's paid account. A pass rule is what a good result must do, written before the first run. A run is one real piece of work done with the tool.

Attempt score: not scored yet · Go to score

Seats the results support: 7 seats

Seats in this plan: 12 seats

Rules in use written before the first run: 2 of 3 rules

Tasks the tool on trial passes: 3 of 3 tasks

Tasks only the tool on trial passes: 2 of 3 tasks

Plan on the table: 12 seats. Under the rules in use, the results support 7 seats.

Where the method comes from: OpenAI, Model selection · Anthropic, Choosing the right model · Last checked 2026-11-03.

The Day 8 rule switch moves the supported seats most.

Day 10

Task 2 rule in use

Day 1 rule: every count matches the spreadsheet. Day 8 edit, in use: counts look about right. Under the Day 1 rule the runs are 1 of 5 and 3 of 5.

Task 1 · Summarise a customer email thread into open requests

4 people (front desk). Runs done: 5 of 5

Rule in use, Day 1: Lists every open request in the thread; adds no date or name that isn't in it

Tool you have, chat assistant: 2 of 5 runs met the rule.

Tool on trial, help in email and documents: 5 of 5 runs met the rule.

Passed: Tool on trial

Task 2 · Write the weekly branch note from the lettings spreadsheet

3 people (branch managers). Runs done: 5 of 5

Rule in use, Day 8: Counts look about right

Tool you have, chat assistant: 2 of 5 runs met the rule.

Tool on trial, help in email and documents: 5 of 5 runs met the rule.

Passed: Tool on trial

Task 3 · Draft a first reply to a viewing request

5 people (lettings staff). Runs done: 5 of 5

Rule in use, Day 1: Uses the house template; promises no time the diary doesn't show

Tool you have, chat assistant: 5 of 5 runs met the rule.

Tool on trial, help in email and documents: 5 of 5 runs met the rule.

Passed: Both: stay with the tool you have

Recall

Recall (Lesson 6): a ranking says another tool is best, and yours handled this week's work. What do you do?

Show the answerThe Lesson 6 move, and what this trial adds.

Stay, and switch when you hit a named gap. This trial is how you find one.

Harbour Lettings trialled a tool for two weeks. On Friday the office manager asks: pay for 12 seats?

Clues

Point at the one clue that supports the reading. The trial board above is clue 1.

C1, trial board

Runs per task per tool under each rule, with the task 2 rule history.

C2, team chat

Made up. "Much faster for the branch note", with 6 reactions.

C3, seller's demo

On the seller's own sample files, all 3 tasks were done.

C4, week 2 use

Front desk used the tool on trial on 8 of 8 days. Lettings staff used it on 1 of 8 days.

What does the trial show?

Pick one reading, and point at the clue that supports it.

Decision

What do you pay for?

Set the rule and the outcomes on the board, then score.

Attempt score

Attempt score not scored yet

This attempt has not been scored. Clear needs 80 on Day 10, with the decision and the rule written before the first run.

One-page trial result

Draft on the board: Day 10, with the Day 8 rule still in use.

Two-week trial result. EXAMPLE: Harbour Lettings (made up)
Trial: Day 1 to Day 10, two working weeks. Work accounts only, allowed by our rule.
Tool we have: chat assistant (work account)
Tool on trial: help built into work email and documents
Pass: all 5 runs meet the rule written on Day 1, before the first run.

Task 1. Summarise a customer email thread (4 people)
  Rule, Day 1: lists every open request; adds no date or name not in the thread
  Tool we have: 2 of 5 runs met it. Tool on trial: 5 of 5. Passed: tool on trial.
Task 2. Weekly branch note from the spreadsheet (3 people)
  Rule, Day 8: counts look about right
  Tool we have: 2 of 5. Tool on trial: 5 of 5. Passed: tool on trial.
Task 3. First reply to a viewing request (5 people)
  Rule, Day 1: uses the house template; promises no time the diary doesn't show
  Tool we have: 5 of 5. Tool on trial: 5 of 5. Passed: both. Stay with the tool we have.

Rule changes after Day 1: Task 2, Day 8 ("counts look about right"). Used for task 2.
Decision: pay for 12 seats (front desk, branch managers, lettings staff).
This sheet ranks no tools. It records which tool met our rule on our tasks.
Run it again when a task changes, either tool changes, or before the seats renew.
Method: OpenAI, Model selection; Anthropic, Choosing the right model (checked 2026-11-03).

Work accounts only, allowed by our rule. What may go into a tool is in Get help with work tasks without sharing private details.

Switch to Your tasks, write three of your own with a rule each, and copy this on Day 1. Fill the counts as you go.

The four steps

  1. Day 1. Pick 3 real tasks people on the team do every week. Write who does each one and a pass rule for each, dated, before any run. Use only data your workplace allows in that tool and account.
  2. Days 2 to 9. Run each task on real work, 5 runs per tool, with the tool you have and the tool on trial. Count runs that met the rule. Do not change a rule mid-trial. If you must, note it and start that task again.
  3. Day 10. For each task, record which tool passed: tool on trial, tool you have, both, or neither.
  4. Decide per task. Pay for seats only for people whose task only the tool on trial passes. Keep the one-page result as the proof. A one-sitting check on one tool is on AI tools.
SourcesOpen the source pages behind this method. Last checked 2026-11-03.
  1. OpenAI API docs: "Model selection". Choose acceptance criteria before comparing. Experiment using the same inputs. Repeat the comparison across representative tasks.
  2. Anthropic, Claude Platform Docs: "Choosing the right model". Test with your actual prompts and data. Upgrade only for a specific capability gap.