Three tests to validate an AI-built automation before you build it

Executive overview

AI planning tools produce ambitious plans that look impressive but collapse in execution. The failure starts in the planning phase, not the build phase.

Run three 30-minute tests before writing a single line of automation. They verify scope, technical feasibility, and cost before you invest days of effort.

Strip the plan to its core, prove the AI can do the hard part, then confirm the economics.

The strip test

• AI plans default to over-engineering: analytics, multi-campaign support, CRM integrations, landing pages — none of which are core.
• For each feature, ask: does the core value work without this?
• If yes, move it to version two.
• To pressure-test the plan, open a fresh AI chat and use this prompt: "Act as a ruthless product manager. Categorize every feature as core (product is useless without it) or later (adds value but not core). Be extremely aggressive in assigning things to 'later'."
• Add your own dictated description of what the automation must actually do between the prompt and the pasted plan.

The linchpin test

• Every automation has one or two tasks where, if AI fails, everything else is irrelevant.
• Identify that task before building anything else.
• Drop it into a chat app (ChatGPT, Claude, Gemini) with a basic prompt and test it manually — no code, no infrastructure.
• If it works immediately: proceed with confidence.
• If it partially works: refine the prompt and context, then retest. Repeat until it meets your quality bar or you conclude the current models can't do it.
• If it can't meet expectations after iteration: shelve the idea and revisit when a stronger model is available.
• Example: a client needed to extract structured data from invoices in varying formats. Three of five test PDFs worked with a basic prompt. Prompt refinement got it to five of five before any automation was built.

The price test

• The linchpin test is usually run on the best available model — which is also the most expensive.
• If the automation will process 10, 100, or 1,000 inputs per day, cost compounds fast.
• Swap to a cheaper, smaller model and optimize the prompt until it still meets your quality bar.
• Smaller models are also faster, which compounds the benefit.
• To estimate cost: prompt AI with the model name, API usage, daily/weekly volume, and average input size, and ask it to return an estimated monthly cost based on current pricing.
• If the cheaper model passes, downgrade and lock it in.