Before any money moves on AI, there are eight questions worth answering in writing. They are practical recommendations drawn from the research, not a validated predictor of financial impact, and they cost nothing to ask. If a question has no answer yet, that is not a reason to stop. It is the work that should happen before the spend.
Where do these eight questions come from?
I read the primary research on why AI projects fail: Gartner on use-case selection and change management, McKinsey on workflow redesign, IBM on people and adoption, RAND on leadership and data (RAND is also clear that real technical limits exist and belong in the feasibility check). The figures on the enterprise-level financial (EBIT) impact gap are in why AI use doesn't turn into AI value. These eight questions are my practical recommendations, not a validated predictor of financial impact.
The eight questions
1. Which process hurts?
"AI across the business" is not a use case; it is a wish. Gartner's top failure cause is exactly this: effort spread across low-impact initiatives without prioritisation. Pick one repeated process that visibly costs people, hours or quality, and name it as an action with a beginning and an end. "Prepare a draft reply from an approved handbook for a person to check" can be examined. "Help the team with documents" cannot.
2. What number does this have to beat?
Before the pilot, write down the business metric you already track that the change should move: hours per week on the task, backlog, error rate, turnaround time. If nobody can name the number, that is the first thing to fix, not the model. A pilot with no number cannot succeed or fail; it can only linger.
3. Who owns the outcome?
Name the person, usually the process owner, who is on the hook for the business result. Not the vendor, not IT, not a central innovation team. When a cost or risk question appears mid-project, and one always appears, the project needs a defender with a stake in the outcome. MIT's research on enterprise AI found that decentralising implementation authority while keeping accountability is what the successful organisations did.
4. What changes in the workflow?
Deloitte counts 84% of organisations that have not redesigned jobs around AI. That is a workflow finding, not proof of what causes the enterprise-level EBIT gap. Decide before the pilot what the new sequence is: who prepares, who checks, who approves, where exceptions go, what the tool may do on its own and what it may not. Inserting a tool at the edge of an unchanged process is the standard death pattern: individuals get faster, the process stays the same, the number does not move.
5. Is the data for that one process good enough?
Gartner's February 2025 forecast is that, through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data. That is not a forecast for all AI projects. In its July 2024 survey of 1,203 data management leaders, 63% reported their organisations lacked or were unsure they had the right data practices for AI. The practical answer is not an enterprise data programme. It is fixing the data for the one process you chose: the sources, the versions, the missing fields, the access rules. One process is a tractable problem.
6. What are the people doing the work told?
BCG found only 36% of employees feel adequately trained on AI, and only 25% of frontline workers say their leaders give enough guidance. The people doing the work decide whether any of this is absorbed. Tell them what changes for them, what the tool is for, what it may not do, and train them as part of the project, not as a rollout afterthought.
7. How will the pilot be judged?
On the number from question 2, at the end, by the owner from question 3. Not on demo quality, not on usage counts, not on how it felt. Count the full cost: implementation, model or API usage, hosting, human review, corrections, maintenance, and failed attempts too. I measure that as cost per accepted result: the full cost, including every failed attempt, divided by the number of results that actually pass the acceptance checks. Putting only the successful attempts in the numerator, and leaving the failed ones uncounted, is how spreadsheets lie.
8. What do we honestly expect for headcount?
In the Census Bureau supplement collected November 17, 2025-February 8, 2026, 2.0% of AI-using US firms reported AI-driven employment decreases in the preceding six months. In that supplement, 44% of AI-using firms reported task augmentation, which can coexist with substitution or task creation. Those observations do not forecast headcount changes for your project. Build the case on capacity, quality and speed of completed work instead, and it will keep the change management from question 6 on your side rather than against it.
The short version for the budget meeting
| Question | A weak answer | A practical recommended answer |
|---|---|---|
| Which process hurts? | "AI across the business." | One named, repeated process with a start and an end. |
| What number does this have to beat? | "Efficiency." | A metric already tracked, with today's value written down. |
| Who owns the outcome? | "IT, or the vendor." | The process owner, named. |
| What changes in the workflow? | "The tool will help." | Who prepares, who checks, who approves, where exceptions go. |
| Is the data ready? | "Probably." | The sources, versions and access rules for that process, checked. |
| What are people told? | "A launch email." | Training and an honest account of what changes for them. |
| How is it judged? | "A successful demo." | The business metric, the full cost, failed attempts included. |
| What about headcount? | "We will need fewer people." | Capacity, quality and speed; no assumed job cuts. |
What if some questions have no answer yet?
That is normal, and it is what a scoping conversation is for. The next step after these eight is the trial itself: inputs, data boundary, acceptance checks, reviewer and a written stop or go decision, which I cover in how to scope an AI trial before buying tools. The free task questionnaire works through the first four questions on paper.
Frequently asked questions
Where do these eight questions come from?
These are my practical recommendations drawn from research on use-case selection, workflow redesign, adoption, leadership and data. They are not a validated predictor of financial impact. Technical limits matter too: RAND discusses feasibility, and MIT identifies output-quality, memory and adaptability limitations.
How are these questions different from scoping an AI trial?
These gate the spending decision: they test whether the project deserves a budget at all. The trial scope, meaning inputs, data boundary, acceptance checks, reviewer and stop or go criteria, is the next step. I cover it in how to scope an AI trial before buying tools.
Should we judge an AI pilot on usage or adoption metrics?
Usage and adoption metrics show whether people use the tool; they are not themselves measures of business outcomes. I recommend tracking a business metric too, and counting the full cost including human review, corrections and failed attempts, divided by the number of results that pass acceptance checks. I measure that as cost per accepted result.
Should the business case assume AI cuts headcount?
I would not assume job cuts. In the US Census Bureau supplement collected November 17, 2025-February 8, 2026, 2.0% of AI-using US firms reported AI-driven employment decreases in the preceding six months. In that supplement, 44% of AI-using firms reported task augmentation, which can coexist with substitution or task creation. These are observations, not a headcount forecast. Build the case on capacity, quality and speed of completed work instead.
Bring me the task that keeps coming back
The free 45-minute remote call works through the first four of these questions with you, on one real task, and ends in a one-page written note on fit and next steps. If the honest answer is an existing tool, a process change or no build, the note will say so. Book the call.