Decision guide · AI & small teams
Evaluate an AI workflow for a small team
A useful AI experiment begins with a task you already understand well enough to judge. Measure what happens to the whole workflow, including checking the output and repairing mistakes.
Founder Reality resource · Updated
Define the task and an accepted result
Pick a bounded workflow with examples you can inspect: resolving an issue, preparing a draft or classifying a support request. State the acceptance criteria before choosing a model or agent.
Record the current time, cost and common failure modes. Include the person who checks the result. Without a baseline, a compelling demonstration can hide additional work transferred to a reviewer.
Compare completed work, including review
Run both processes on representative examples. Keep the inputs, outputs and decisions about whether the work was acceptable. Count rejected outputs and retries as part of the experiment rather than excluding them from the comparison.
Record model or API spend, review minutes and repair work alongside elapsed time. A fast first draft may still be expensive to finish. A workflow that performs well on routine cases needs a separate check for the cases where a mistake is costly.
Decide who owns the workflow after the demo
Name the person responsible for approving changes, watching failures and maintaining a fallback. Decide what the system may do on its own and what requires a human decision. Test the fallback before depending on it.
Use the result to choose a bounded next step: adopt for a narrow task, run another experiment, or stop. Keep the evidence so a new model, price change or failure can reopen the choice. The essays below are operating accounts; the assessment scores are prompts for reflection, not forecasts of replacement.
- A baseline and acceptance criteria
- Rejected outputs and review costs included
- A named owner and tested fallback
Decision worksheet
Copy these prompts into your notes. Keep the answers and the date so you can revisit what changed.
- Task
- What input produces what accepted output, and who judges it?
- Comparison
- What changed in completion time, review time, cost and failure rate?
- Authority
- Which actions require approval, and who can stop the workflow?
- Revisit
- Which failure or change in cost would make me run the comparison again?
Use your own numbers
AI-Proof Assessment
Assess whether your work is in layers AI will eat. Get your AI-Proof Score and personalized recommendations based on the Post-AI Manifesto framework.
Free tool · no email required →
Dependency Score
Measure how dependent your business is on you across 6 dimensions: revenue, decisions, clients, knowledge, execution, and relationships.
Historical tool · assumptions may be dated →
30-Year SaaS Tax
Calculate your total subscription cost over 30 years and see what you could build instead.
Free tool · no email required →
From the field
Personal accounts behind the questions. Each essay reflects when it was written.