Can't Answer These Five? Don't Sign the AI Project
Five questions to settle before funding an AI project: can failure be reversed, is the data in hand, can acceptance be measured, who can call stop, and who maintains it in three months.
Our position
Most AI projects don't fail in delivery. They fail at the decision to start.
What to take away
- Ask whether failure is survivable before asking how good the upside looks.
- If acceptance can't be written as a number, there is no acceptance.
- Without a named owner, launch day starts the depreciation clock.
Method breakdown
-
1. Can a failure be reversed?
The risk with AI is not that it performs badly; it is that a wrong output cannot be taken back. Once outbound replies, ledger entries or generated contracts are automated, one mistake gets copied many times. List the irreversible actions in this use case first, then decide which ones keep a human gate.
- List irreversible actions (send, post, sign, delete)
- Assign a human confirmation point to each one
- State the rollback window and who owns the rollback
-
2. Is the data actually in hand?
Most cases of "the AI is inaccurate" are cases of "the data isn't there". Before funding, you should be able to name the three data sets this use case minimally needs, which system holds each one, who can grant access, and how often it changes. If those four answers are missing, a more expensive model will not help.
- List the minimum data set, not all the data
- Name the data owner and the access path
- State the refresh cadence and what happens when data goes stale
-
3. Can acceptance be measured?
"Improve efficiency" is not an acceptance criterion. Before signing, you should be able to write down which 30 to 50 real samples will be used, who labels them, what score passes, and what score fails. The samples must come from the business, not from the vendor's demo set.
- Build the evaluation set from real business samples
- Write the pass line and the fail line as numbers
- Agree who has the authority to declare a pass
-
4. Who has the authority to stop it?
Every AI project needs one person who can call stop and one explicit stop date. Without both, the project extends itself one more iteration at a time and the budget is spent without a decision ever being made.
- Name a business owner, not an IT owner
- Agree the stop date and the conditions that trigger it
- Decide how data, models and documents are handed over after a stop
-
5. Who maintains it in three months?
Launch is not the finish line. Knowledge goes stale, models change, processes get redesigned. If nobody is still watching accuracy, adoption and cost three months later, the system fails quietly and users simply call it unusable.
- Name the maintainer and a monthly review cadence
- Track three long-run metrics: accuracy, adoption, cost per run
- Document how knowledge is updated and how versions are rolled back
Action checklist
- If this use case fails, what is the worst outcome, and can it be rolled back within 24 hours?
- Is the required data in hand, with a named owner responsible for keeping it current?
- Do the acceptance samples come from real business work, and are the pass and fail lines written as numbers?
- Who can call stop, under what conditions, and on what date?
- Who maintains it three months after launch, and which three metrics get reviewed?
Who this is for
Executives, digital leads and procurement signing off an AI budget.