Loading page
About this tool
Define the problem, users, data, success metric and fallback for an AI project before you build — with a completeness score.
An AI project scoping worksheet forces seven answers out of an idea before any modelling starts: the problem, the users, the data sources, one primary success metric, today's measured baseline, the target that makes the work worthwhile, and the fallback for when the model is wrong or offline. This tool captures those answers, scores how complete the scope is out of seven, and emits a markdown scope document you can paste into a ticket or design doc. It follows the standard machine-learning project-scoping sequence — define the metric and the baseline before choosing a model — used in Google's People + AI guidance and in Andrew Ng's ML project scoping steps.
Open AI Project Scoping Worksheet on AltFTool — it loads instantly in your browser.
Provide your input — an image, text, or data.
Let the tool analyze or generate the result.
Review, refine, and reuse the output wherever you need it.
Records today's measured value of the metric so improvement can actually be proved later.
Treats the wrong-answer and outage path as a core question, not an afterthought.
Produces markdown with explicit TBD placeholders, so gaps stay visible when the doc is shared.
At minimum seven things: the problem in someone's workflow, who uses the output and who is harmed by errors, the data sources and who owns access to them, one primary success metric, the current measured baseline for that metric, the target value, and the fallback when the system is wrong or unavailable. Scopes that skip the baseline cannot prove improvement, and scopes that skip the fallback tend to stall at the go-live review.
Because without a measured starting value there is no way to tell whether the model helped. If median first response is 46 minutes today, a pilot that lands at 30 minutes is a clear win; with no baseline the same result is just a number. Measure the baseline from an existing system report rather than estimating it from memory.
One primary metric, with any others treated as guardrails. A single primary metric makes the go/no-go decision unambiguous; guardrail metrics such as error rate, cost per request or customer satisfaction stop the team from optimising the primary metric into a worse product.
The defined path the work takes when the model is unavailable or below its confidence threshold — usually routing to the existing manual process plus a feature flag that disables the AI path entirely. Writing it during scoping is what makes a pilot reversible; if you cannot describe how to turn the system off, it is not ready to switch on.