We build AI products for a living, and roughly once a week we tell a prospective client not to buy one yet. That's not charity — shipping AI that shouldn't exist produces unhappy clients and quiet failures, and we'd rather keep the reputation. Here's the framework we actually use on those calls. Run it yourself, free of charge.
The four questions
1. Is the work language-heavy or judgment-heavy — and repetitive? AI earns its keep on work that involves reading, writing, classifying, extracting or deciding — done over and over. Answering support tickets, qualifying leads, processing documents, triaging inboxes: yes. A one-off analysis or work that's mostly physical: no.
2. Does the knowledge exist somewhere software can reach? AI can't answer from documents you don't have. If the knowledge lives in clean docs, databases or ticket history — green light. If it lives in three veterans' heads, the first project isn't AI, it's getting the knowledge out of the heads (we can help with that too, but let's be honest about the sequence).
3. Can a wrong answer be caught cheaply? Every AI system errs sometimes. The question is whether an error is a caught-in-review annoyance or a catastrophe. Drafts a human approves, answers with citations, suggestions with a confidence flag: workable. Irreversible actions with zero human checkpoint: not yet — and a vendor who says otherwise is selling you their risk.
4. Is there enough volume to pay for it? The build has a real price tag. Ten repetitive tasks a day usually doesn't justify it; hundreds does. Do the arithmetic: hours saved × loaded cost × 12 months, versus the ranges in our cost breakdown. If the math is embarrassing, don't buy — yet.
Red flags — from our own call notes
- "The board wants an AI strategy" (a goal in search of a problem)
- "We'll figure out the data later" (the data is the project)
- "It must be perfect from day one" (nothing is; plan the review loop)
- "Rip everything out and start over" (integration beats revolution)
The unsexy middle option
Half the "AI projects" we scope turn out to need plain workflow automation first — moving data between systems, triggering actions on events, killing copy-paste. Tools like n8n and Make.com do this brilliantly at a fraction of the cost, and we build those pipelines too (one of our production automations runs entirely on Make.com). Start there, collect the data exhaust, add intelligence where it proves worth it. Boring advice, great outcomes.
If you scored four yeses
Don't sign a six-month contract. Run the smallest real test: our AI Sprint exists for exactly this — 2–4 weeks, fixed price, ending in a working proof of concept on your actual data, which you own outright. If the idea doesn't survive a sprint, better to learn it for the price of a prototype than a product.
Want the straight answer for your case? Bring it to a free 20-minute call with a senior engineer — or ask our homepage AI right now. If the honest answer is "you don't need us yet," you'll get that too.
