How to tell if an AI project is worth doing — before you build
Most AI projects don't fail in production. They fail on a whiteboard six weeks earlier, when a promising demo gets mistaken for a plan. The demo works because someone hand-picked the inputs. The plan fails because the real workflow is messier than the slide.
You can catch this early. Before you commit budget, run three checks. Each one is a question you can answer in an afternoon — no vendor, no proof-of-concept, no data science team required.
1. Is the workflow decomposable?
AI is good at narrow, well-defined steps and bad at fuzzy, end-to-end judgment. So the first question is whether you can break the target process into steps and point at the specific ones that are slow, repetitive, and rule-shaped.
If you can write the workflow as a numbered list and circle the three steps that eat the most time, you have something to automate. If the whole thing is "an expert looks at it and decides," you don't have an AI project yet — you have a research problem.
2. Is there a KPI you can actually move?
"Add AI" is not a goal. "Cut average handling time on intake from 40 minutes to 25" is. The second question is whether there's a number your business already tracks — or easily could — that a successful project would move.
If you can't name the metric before you build, you won't be able to tell whether it worked after. That's how pilots end up in permanent "we're still evaluating" limbo: nobody agreed what winning looked like.
3. Is the data reachable?
The best-scoped workflow with the clearest KPI still dies if the data lives in a system nobody can export from, or is locked behind a compliance review that takes two quarters. The third question is whether the inputs the model needs are somewhere you can actually get them — cleanly enough, and soon enough, to matter.
This is the check teams skip most, and it's the one that kills the most projects after the money's spent.
What "no" looks like — and why that's a win
If any of the three is a hard no, that's not a failure. That's the check doing its job. A clear "not yet, and here's the one thing that has to change first" is worth far more than a six-month pilot that arrives at the same conclusion with a budget attached.
The cheapest AI project is the one you decide not to build — after five minutes of honest questions instead of five months of sunk cost.
Where to start
If you want to run this on your own process, our AI Feasibility Scorecard walks you through the same three checks in about two minutes — no email required. And if you'd rather talk it through against a specific workflow, book a discovery call; we'll run these questions live and tell you plainly whether it's worth pursuing.