Should This AI Project Start? A Decision Framework
Compare AI and non-AI options through purpose, people, data, evidence, guardrails, staged tests, total cost, accountability, and a documented proceed-or-stop decision.
An AI idea may need to stop, or it may need more study, a small test, or a non-AI fix instead. It may also lead you to buy a tool, to build one, or to grow what you already have in place. There is no sound proof that most AI projects should never start. The goal is a written decision that weighs purpose, other options, proof, effects, duties, and the cost of being wrong.
The NIST AI Risk Management Framework groups optional work under Govern, Map, Measure, and Manage. Its Playbook is not one single list that suits every team. The Manage guide asks whether a system meets its stated goal, and whether the work or the use should go on.
Frame the Decision Before Naming AI
Describe the task and the people who may feel the effects, along with current results, the problem or the chance, and the owner. Then note the limits and the choice that the project must aid.
Set the allowed goal and the banned uses first, before you pick a model, a vendor, a data source, a screen, or a system design.
Tell apart facts, guesses, forecasts, wishes, and promises, then log who gave each one and when to check it again.
Compare AI With Real Alternatives
Compare no action with process fixes, training, and rule changes. Then compare search, data study, task automation, bought software, and human work, alongside a small AI part and a larger system. Use the same needs, time span, full-cost base, risks, and pass proof for each option you list. An AI label is not a reason to approve or reject.
Map People, Data, and Context
Name the users, the data subjects, and the non-users. Name reviewers, staff, customers, workers, and any groups that may gain or face harm. Then map the data source, the right to use it, its quality, and its fit. Also map private traits, access, vendors, keep time, deletion, safety, ownership, and legal duties. Do not use data just because you have it.
Define Evidence and Guardrails
Set the main result and the quality measures, plus the current base, the allowed group, and the test data. Then set the error cost, the doubt, the review steps, and the pass line that the project must clear.
Add limits for safety, fairness, and privacy. Add limits for security and access as well. Give clear reasons, add human review, and add a way to complain. Cover staff load, money, and green limits as needed.
Name who can pause, overrule, fix, roll back, or retire the system, and give affected people a way to ask for a review or for help.
Use Staged Decisions
Steps may include study, a data check, a mock-up, a fair test, and a small pilot, followed by a watched release, wider use, and retirement. For each step, set the entry proof, the spend rights, the stop rules, the owner, the review date, and the rollback. A demo or a test score does not prove a system is safe for live work.
Verify Vendors and Custom Builds
Check the current model and product notes, the contract, and how data is used. Check other vendors, locations, access, logs, keep time, and deletion. Check issue duties, test proof, change notice, help, export, and exit. For a custom build, set safe coding, parts, tests, and checks, plus staff, upkeep, and handoff. Buying or building does not remove duty.
Maintain an AI Decision Record
Keep the purpose, the other options, and the proof. Keep group input, dissent, conflicts, and approvals. Keep test results, issues, changes, and real results. Keep the reasons to go on or to stop. Review the choice when the model, data, setting, law, vendor, users, results, or risk changes.
What TTGC Can Support
TTGC can help you frame a use case and compare options. TTGC can map needs and risks, make a mock-up, and test it, then build agreed parts and set up records and checks. TTGC cannot promise approval, truth, savings, use, legal status, revenue, or other AI project results.
Should this AI project move ahead?
TTGC can help frame the choice, compare options, set proof and limits, and plan staged tests. AI and business results are not promised.
Sources
- NIST — AI Risk Management Framework 1.0. https://www.nist.gov/itl/ai-risk-management-framework
- NIST — AI Risk Management Framework Playbook. https://airc.nist.gov/airmf-resources/playbook/
- NIST AI RMF Playbook — Manage. https://airc.nist.gov/airmf-resources/playbook/manage/
- Federal Trade Commission — Privacy and Security guidance. https://www.ftc.gov/business-guidance/privacy-security
