Your AI rollout is a utility bill, not a strategy

The only work I saw put real money on a client's books this year came from a handful of narrow projects that required almost nobody in the company to change how they work.

I know that's the opposite of what you heard at whatever summit you attended last quarter. The pitch there is that AI is a change management problem: get seats in every hand, stand up a center of excellence, train everybody, measure hours saved, and the compounding productivity gains take care of themselves. I've been in enough client rooms this year to say plainly that the first half of that is true and the second half is a story we tell CFOs because the real answer is less flattering.

Here's the real answer, from two clients in the same month.

A client had a large contract they renegotiated every three years by hiring a consultant. His fee was half the upside. Save them a million, pocket $500K. This year we skipped him. We fed the 30-page contract and a pile of invoices into AI, built a negotiation playbook, and put the incumbent against their competitor. They ended up switching to the competitor and saving a couple million dollars over three years.

Number of employees who had to change their daily behavior for that to work: two.

Meanwhile, a different client discovered their employees had built 1,200 AI agents. That looks like the adoption slide everyone wants. Then we looked closer. It was roughly ten things, built fifty different times each. Fifty meeting summarizers. Fifty report drafters. Each slightly different, none shared, most abandoned by whoever built them.

Number of dollars that shows up on a P&L: hard to say, and I'd guess close to zero.

The distinction I've started forcing on every engagement is this: your AI spend belongs in one of two ledgers, and the rules for each are opposite.

Ledger A is utility. Licenses, seats, general employee productivity, agent-building sandboxes. This is Microsoft Word. Nobody ever built a business case for Word. At some point it just became the cost of doing knowledge work, and the alternative (taking it away from people) is not a side of history you want to be on. Ledger A gets budgeted like electricity. You do not ask electricity for its ROI.

Ledger B is projects. One narrow problem, one number attached, one owner. The complex vendor contract we renegotiated. The once a year website audit we replaced with a weekly analysis engine. The complicated legal workflow we automated. Ledger B lives or dies by the number.

The mistake I see almost universally is running Ledger A and reporting it as Ledger B. An exec asked me directly: we're paying for licenses for the whole company, so where's the savings? My honest answer was that in 2026, probably nowhere. When someone saves two hours, what happens to those two hours? Most people backfill them with email. One of his own employees said it better than I did in a training session: the hard part isn't saving the time, it's not filling it with busy work.

And in high-stakes work, the saved time doesn't even make it that far. A litigation associate told me Harvey turned a six-hour research project into 30 minutes. Then she told me it has saved her no time at all. She checks the output, checks it again, has the Word version check the browser version, runs the cases through Westlaw's AI, downloads the sources and feeds them back in to ask what it got wrong. Five or six rounds. Her estimate of raw accuracy: about 70%. Good enough to be dangerous, not good enough to file. Then the junior associates skip the checking, the partner catches the errors, and she re-verifies their work too.

She would not give the tool back. Her hours are the same and her work is better, because she spends the time on substance instead of searching. That is a genuine Ledger A win. It is also not a dollar, and any business case that pretended otherwise was lying by a full line item. Trust is the bottleneck, and trust builds slower than capability.

Once you separate them, three things follow that most AI programs get backwards.

Stop measuring Ledger A in hours saved. For the first couple of years in any high-stakes profession, hours saved is the wrong metric entirely. What you're buying is coverage and judgment: arguments caught that would have been missed, briefs stress-tested against the other side's best counters. Report that honestly and stop promising a number you won't hit.

Do enforce Ledger A anyway. A law firm pulled usage logs and found several attorneys hadn't touched the tool in six months. One said he was "still not really understanding how it can provide any benefit" to him. The committee's first instinct was the usual one, more training sessions, more lunch-and-learns, which is exactly the approach that produced six months of zero usage in the first place. Optional encouragement has that failure mode built in. What they chose instead was an internal certification: practice-specific exercises, real work, take a 25-page commitment letter and cut it to 6 pages for a client who asked for something readable, complete five and you're certified. It converts "I haven't gotten around to it" from an excuse into a gap. Utilities are mandatory. Nobody opts out of email.

Mine Ledger A for Ledger B. This is the part almost nobody is doing. Those 1,200 agents are not a mess, they're a survey. Your employees already told you what they need, and they told you fifty times each. The job of a central AI team is not to build everything or gatekeep tools. It's to look across the company, spot the ten patterns, and promote the best duplicated experiments into shared infrastructure with a number attached.

I should be clear that I get this wrong too. We showed a client everything one of our AI systems does. It aggregates intelligence, builds briefing materials, tracks risk, forecasts scenarios. We put all of it on screen and their leadership said: "Where's the action?" It took me a day to understand what happened. We presented the tasks the technology performs, which is a Ledger A framing. They wanted to see what they'd be able to do for their own clients that they couldn't do before, which is Ledger B. Those sound like the same thing. They are not close. I've built products for 25+ years and I still organize demos around where the effort went, and the effort is invisible to the buyer.

The uncomfortable conclusion: the company that skips the transformation program, pays the license bill without arguing about it, and executes four departmental workflow optimization projects will beat the company with the better AI strategy deck. Not because enablement is worthless. Because enablement is the floor, and everyone is going to have it by 2027, the same way everyone has Word.

Homework: Take your last twelve months of AI spend and sort every line into Ledger A or Ledger B. If Ledger B is empty, you don't have an AI program, you have a subscription. Then find your complex contract renegotiation. Write down three documents in your business that somebody on the other side wrote specifically to be hard for you to understand, and bring one of them to your next staff meeting.

Two things I'd genuinely like back from you. First, hit Reply and tell me which part of this matches what you're seeing. Second, if there's someone in your world who should be on this list, forward it to them.

- Robbie