Estimated Reading Time: 5 minutes
Hi, I am an executive with decades in the trenches, summarizing proven techniques and playbooks that are working today to help you succeed. Connect with me here: https://www.linkedin.com/in/stevegrady/
This CEO AI Pains Series insight covers Pain #1, No Measurable ROI, and includes a Claude playbook to address it before any AI agent exists.
This question usually arrives in the last ten minutes of a board meeting, after the operating review, when everyone is tired. A director asks what the company has gotten back from the AI spend.
The room goes quiet. The CEO says something about early wins and learning.
That answer buys a couple of weeks to come up with some concrete results.
The famous MIT Sloan paper reported that 95% of generative AI pilots failed to produce measurable P&L impact. I don't think that statistic means the technology failed. It means nobody underwrote the work before the money left the building.
After starting 18 months ago with AI, a typical mid-market company has fifteen or more AI initiatives running. Marketing bought a copywriting tool. Finance is testing something with their ERP for collections. Two engineers built a chatbot on a corporate card. Each one had a champion and a plausible story.
But none of them had a baseline.
That's the whole problem.
If the team never wrote down what the process cost before the pilot, you cannot prove what the pilot saved. Finance was not in the room when the hypothesis was set, so finance will not defend the result. And because there was no kill criteria, nothing ever gets shut down. The unwieldy portfolio only grows.
There's a second layer that stings more. The initiatives that genuinely worked look identical, on paper, to the ones that did nothing. You have no way to sort them, so you fund all of them or none of them. Both answers are wrong.
You don't need an AI agent at this point to fix this. You need an underwriting standard and about two weeks of disciplined work. We will use Claude Cowork with the following sequence and the files you already have.
Step 1. Build the inventory from money, not memory. Export twelve months of AI-related line items from your expense system, your SaaS management tool, and your project tracker like Jira. Paste them into Claude Cowork and ask it to consolidate duplicates, group by department, and produce one table with vendor, annual cost, sponsor, and stated purpose. Expect the count to surprise you. It usually runs 30% to 50% above what the CIO believes.
Step 2. Have Claude write a one-page underwriting standard. Six fields, no more:
Have Claude draft the template, then edit it yourself so the language matches how your finance team already talks.
Step 3. Interview each sponsor through the model. Give the sponsor fifteen minutes. Use an AI transcript generator and paste the notes into Claude with your standards skill, and have it draft the business case in that format. Then ask it to list what's missing and what claim it cannot support with the information provided. That gap list is the most valuable output of the entire exercise.
Step 4. Grade the portfolio. Three buckets. Fund, fix, or kill. Ask Claude to score each initiative on confidence in the value claim and difficulty of proving it, then have Claude place them on a simple two-by-two matrix. Your first wave is the high confidence, low difficulty corner. Nothing else.
Step 5. Set the monthly run-rate review. Same table, same prompt, refreshed with actuals. The discipline matters more than the sophistication. A CFO who sees the same six fields every month starts to trust the numbers by the third cycle.
Step 6. Produce the Board page. One page with entries for: Total AI spend, Spend with a named owner and baseline, Initiatives killed, Run-rate value captured against hypothesis, Next three AI initiatives funded. Have Claude compress your table into that page, then rewrite the narrative in your own words before it goes in the pack.
The first outcome is not financial. It's the ability to answer the director's question with a specific number and the underlying assumptions. That changes the conversation from defense to allocation.
Using this solution, your portfolio gets smaller. Every company I've watched run this exercise kills something, usually a tool with an enthusiastic champion and no measurable output. That's an uncomfortable meeting, but it’s an important part of the rationalization process.
Now you track four things:
Be honest with yourself about the hard part. Finance and the initiative sponsors will fight over the baseline because a generous baseline makes the pilot look good, and a strict one makes it look ordinary. Somebody senior has to arbitrate that, in writing, once. If you skip it, the standard becomes decorative within two cycles.
This AI ROI pain point comes up most often in client discovery meetings, so it sits first in our TPC AI Orchestrate Solutions development plan. The AI Value Underwriter agent will do I described above continuously rather than in a two-week sprint, and it tracks run-rate value after deployment instead of stopping at the funding decision.
Before we build further, I want to know whether this matches what you're dealing with. Does the baseline fight sound familiar, or is your obstacle somewhere else entirely?
Next CEO AI Pains edition: Pain #2 is the failure that shows up around 90 days after the budget and the calendar are already spent. But it gets blamed on the wrong thing every time.
Now, if you want to jumpstart conversations in these CEO AI Pains, Solutions, and Outcomes, email me: sgrady@tpcon.com
A note on the AI Agent ROI math used in the graphics
All ROI figures anchor to a single mid-market reference firm so the ten agents are directly comparable: $250M revenue, 400 FTEs, 4–7 core enterprise systems, no Chief AI Officer in place at the AI start, 12–18 months of pilot activity, and 15+ scattered AI initiatives. Loaded hourly rate of $85 blended ($120 manager, $65 IC), sourced from BLS Q1 2026 mid-market compensation data with 30% overhead.
Till next time…
