Turning an AI wishlist into funding decisions
An automotive client's AI program had more ideas than budget and no way to compare them. I rebuilt intake around the three inputs its SVPs actually decided on, and the first review cycle reached decisions in about two weeks instead of five.
- Organization
- EY, for an automotive OEM
- My role
- Senior Technology Consultant, AI Strategy
- When
- 2024 to 2025
- Result
- Decision time from about 5 weeks to about 2 in the first cycle
The problem
Teams submitted AI requests as long written essays. The four SVPs funding the program couldn’t compare one against another, review meetings stalled, and approvals sat. I sat between the engineering teams proposing use cases and the SVPs funding them, and ran the communication between the two. The bottleneck was the format: nobody could make a funding decision from an essay.
What I did
I interviewed each SVP about what actually moved their decision. Their answers differed, but the overlap across all four came down to three inputs: cost, size of the business problem, and technical feasibility. I rebuilt the intake form around only those three and built a Power BI scoring view that put every request on one page.
The hardest input was feasibility. Teams were overconfident about their own ideas and underconfident about anything unfamiliar, so self-reported scores were noise. I walked department heads through how to size an idea, and I was frank with the SVPs about how much weight those scores could carry, so they knew where to push in review.
Result
The first review cycle on the new process reached decisions in about two weeks, against a historical average of about five. Low-value requests got filtered before they reached the SVPs. My favorite example: an engineer who wanted approval for an AI agent to clean up his own work email.
What I got wrong
Cost. The rubric took cost estimates at face value, and approved projects often ran over. I went back through past project cost data to find where estimates broke, and used it to rewrite the follow-up questions we asked teams before a request went to review.
A rubric is only as good as its least reliable input. Cost looked like the easy, objective number, and it was the one we got wrong.