When the KPI Becomes the Business

I round up the most relevant AI-in-finance news, the deals being done, who's rolling out what, and what's actually working on the front lines.

A target gets hundreds of people moving the same way. AI could make it far cheaper to change direction.

The companies building out AI have borrowed nearly $850 billion this year, according to The Economist, roughly four times what the same group raised in 2023. Credit markets have started to flinch. CoreWeave's bonds now yield 15 per cent, and the Nasdaq has given back everything it gained since May.

Meanwhile, a quarter of financial firms now tie bonuses or promotions to how well staff use AI, according to a Selby Jennings survey. I think that misses the point quite badly.

Elsewhere, investors are trimming their expectations for Anthropic's IPO, which will also test the unusual trust that controls its board. GTCR is taking CCC private, and a $30 billion data-centre IPO lasted about 48 hours.

But first, my take on what happens when a KPI becomes the thing a company steers by, and why AI could make it easier to change course.

In today's Acquisition Intelligence:

From The Trenches

One of our conversations this week started with a familiar problem: revenue was coming in broadly as expected, but profit kept falling short.

The business was watching how much work came in and how much got completed. Those were sensible things to monitor, particularly when customers expected jobs finished on time. But getting through the work could mean weekends and overtime, and the volume figures gave an incomplete account of what it cost to deliver them.

Working through it made me think about our attachment to KPIs. We put an enormous amount of faith in the idea that a complicated business can be understood, compared and managed through a handful of numbers.

The Limits Inside the Number

There are good reasons for that. A regional manager cannot inspect every job, understand every contract and keep the payroll calculation in their head. A partner overseeing several companies has even less chance, so we give them a compact picture of the business and a set of targets to orient everyone around.

Spend time underneath those numbers and the apparent precision starts to dissolve. “Productivity” depends on which work you count and whose hours you include. Two divisions can report the same KPI using different exclusions and adjustments that made perfect sense to the person who introduced them three years ago.

Then we normalise it, average it and compare it with last year. Each step makes the business easier to discuss, while removing some of the detail that explains how it actually works.

Turning the Tanker

But the bigger consequence is what happens once the organisation starts pursuing the number.

A KPI gets an owner, a target and a place in the weekly meeting. It finds its way into budgets and bonuses, and managers build routines around delivering it. Eventually, a considerable amount of the company's effort is directed towards improving something that began as an approximation of how well it was doing.

That alignment is valuable. Getting hundreds of people moving in roughly the same direction is difficult, and a clear target helps. The difficulty comes when the business needs to change direction and all that alignment becomes inertia.

A measure that made sense while the company was trying to fill spare capacity can become much less useful when capacity is stretched. A push for volume can outlive the attractive margins that justified it. The KPI becomes a straitjacket, keeping the business pointed towards a goal whose usefulness depends on conditions that have already changed.

Reversing it means revisiting targets, explaining new priorities and persuading people that the behaviour they were rewarded for last quarter should now change. It can feel like turning a tanker. Some of the rigidity comes from the sheer effort required to make a new direction understandable to everyone involved.

Making It Easier to Change Course

I have written before about that reporting “whisper chain”, and how AI can lower the barriers between teams by connecting the systems they work in. Operations can see billing, finance can see the underlying work, and the owner can investigate the business without waiting for another interpretation of it.

What I think follows from that is more interesting than getting everyone's numbers to agree.

We have built management systems around the amount of complexity people can absorb. We need categories we can remember, averages we can compare and a manageable number of priorities. Once the whole company has learned to operate through that framework, changing it becomes expensive.

AI gives us a reason to question how fixed the framework needs to be.

A system working across the underlying records, contracts and operating history can retain detail that a KPI compresses away. More relevant, reliable context gives it more opportunities to investigate relationships between parts of the business, including connections nobody thought to include in the reporting template.

A falling margin might connect to customer mix, a scheduling decision elsewhere or an exception buried in a contract. Each could look unremarkable in the report belonging to the team responsible for it. Bringing them together could reveal that the priority management set three months ago is now producing consequences it never intended.

Those connections need testing. But we can investigate them without turning every relationship into another permanent KPI that someone has to report each month.

We can also change which measures deserve attention as the business changes. During a seasonal peak, fulfilling commitments and protecting scarce capacity might dominate. In a quieter period, the emphasis might move towards customer economics or releasing cash tied up in the operation.

The business still needs consistent measures of its overall health. But the indicators guiding this week's decisions could reflect the conditions it is actually facing, with an explanation of why the emphasis has changed.

An AI-supported operation could make that change easier to carry through. The same underlying records can support a revised analysis, the system can identify where existing targets conflict with the new priority, and each team can see what the change means for its decisions. Managers would still have to agree the trade-offs and change incentives where necessary, but less of the transition would depend on rebuilding reports and passing explanations down a long chain.

That is one of the more interesting possibilities of an AI-native company: it could cost much less, in management time and organisational effort, to change course. The company could retain the ability to respond to the season or the market without rebuilding its way of understanding itself each time.

Understanding the Recommendation

There is a connection here to a Sam Altman remark I came across recently. In a July 2025 conversation at the Federal Reserve, he raised the possibility of a future president receiving advice from AI that they could neither improve upon nor fully understand. He presented it as a concern about people ceding decisions to systems they could no longer properly assess.

The business version is less dramatic, but it raises a similar tension. We want AI to consider more than we can hold in our heads, including relationships we would struggle to spot ourselves. Some of its conclusions may therefore look strange through the measures we have spent years teaching people to improve.

That makes the ability to interrogate an answer important. We need to see the evidence, understand the assumptions that matter and check whether the expected result materialises. We also need to remain explicit about what the business is trying to achieve, because more information cannot decide how much we value a customer relationship or which risks we are willing to accept.

Giving the Numbers Another Dimension

We will still need indicators to track performance, but I think we can give them another dimension. A productivity figure could show how a team performed given the difficulty of its work. A healthy margin could reveal that a few excellent contracts are concealing deterioration elsewhere.

The number would carry more of the circumstances behind it. It could help explain why a result was good, whether the same approach will work next quarter, and when management should reconsider the target itself.

Flexibility also needs a memory. If we change a priority, we should record why, retain the previous measures and follow what happens afterwards. Otherwise, it becomes too easy to redefine success whenever the result is disappointing.

That is where I want our work at DealSage to lead. The board pack could show the health of the business consistently, alongside the priorities management is pursuing now and the evidence that those choices are working. People would have a clearer basis for changing their decisions when the circumstances change.

A company needs enough consistency for people to work together. I think AI could give it much more room to adapt without spending another quarter getting everyone pointed in the right direction.

From The Builders

Brexton on why focus wins now (X, October 12). Cantor's co-head of AI infrastructure says coding agents make building so easy that companies lose sight of the one problem their customer has. Deal teams do the same: thirty agents, and the value sits in the three or four people open every week.

Felix Rieseberg on going deeper with models (X, October 13). Rieseberg, an engineering lead on Claude at Anthropic, says the fix for underusing models is depth: start from the rawest data, set a target that sounds absurd and have the work checked from outside. I tried it on some of our own work and was a little disappointed, but it's a useful way to think about what you ask for.

News Digest

The AI Boom's Debt Gets Harder To Read

The Economist added up the borrowing behind AI this week and got to nearly $850 billion so far in 2026, roughly four times what the same companies issued in 2023. The four big hyperscalers alone have borrowed $214 billion this year, against $83 billion in all of last year, and their balance sheets together carry more than $3 trillion of debt. Credit markets are starting to price the risk. CoreWeave's five-year bond yields 15 per cent, up from 10 in July, and its credit default swaps imply a better than 60 per cent chance of default within five years.

The details:

Why it matters: A lot of this sits with private credit and the insurers they own, and Ramp's card data shows business spending on AI has flatlined since July.

My take: The number I'd watch is Ramp's, more than CoreWeave's. The debt only works if businesses keep paying more for AI every year, and plenty of them are pausing because the spend hasn't changed how they work yet. That's a deployment problem rather than a technology problem, and it's the bit we spend our days on. For anyone with exposure to the lenders, the structures are now layered enough (chips, buildings and power each financed separately, with backstops on top) that working out who takes the loss will be its own diligence exercise.

Read more

You Can't Bonus Your Way To AI Adoption

Nearly a quarter of financial firms now link raises, bonuses or promotions to how well employees use AI, according to a Selby Jennings survey of 500 finance executives reported by Reuters on October 14. Another 30 per cent make staff disclose their AI use in performance reviews. The practice is more common at fintechs and smaller firms, and Selby Jennings expects the link between pay and AI to become more formal by next year.

The details:

Why it matters: Firms have bought the licences and usage is lower than they hoped, so pay is the lever they're reaching for.

My take: I think this misses the point entirely. You can't make people adopt AI by telling them their bonus depends on it, and if the strategy is to roll out a pile of tools and then measure who uses them, you'll mostly get people running fast in different directions. Usage is easy to inflate and tells you nothing about whether the work changed. What works is picking the few workflows that matter, getting the data underneath them right and building those properly, which is how we approach it at DealSage, and at that point people use the tools because they save time.

Read more

Other Interesting Things I've Read or Seen This Week

Ex-Palantir and Carlyle executives launch Aldaron with $885 million (WSJ, October 15). The new firm buys US businesses worth $200 million to $1 billion and puts its own engineers in to run the change. At Oak Point, its first deal, AI now surfaces 10 to 12 opportunities a week instead of one or two, and headcount is up 20 per cent, which is not the usual direction for an AI story.

GTCR and Elliott are taking CCC private for $6.1 billion (CCC, October 15). The claims software company sits on 40 years of industry data, and GTCR calls it "a business that we think is made better by AI." At $7.75 a share it is still below the $10 SPAC price of 2021, so the thesis comes at a discount.

Investors are trimming their Anthropic IPO expectations (Reuters, October 15). Some are now talking about $1.5 trillion rather than $2 trillion, after third-quarter revenue grew 45 per cent to $16.8 billion against more than doubling the quarter before. Only in AI does 45 per cent quarterly growth count as a slowdown.

Anthropic's IPO will test the trust that controls its board (WSJ, October 18). The Long-Term Benefit Trust appoints most of the board, and Reuters reports Anthropic will also commit a 5 per cent stake to a new public benefit fund. Public investors will be buying into an unusual governance structure, and the S-1 should make for a long read.

Accordion buys Stride and its forward-deployed engineers (Accordion, October 13). The private equity-focused consultancy is adding about 50 engineers who build agent systems for regulated industries. A finance consultancy buying engineers to get AI into portfolio companies is about as clear a signal as you'll get on where the work is.

Thoma Bravo's BlueMatrix is buying Aiera (Thoma Bravo, October 8). BlueMatrix distributes broker research, and Aiera lets AI systems read it without losing track of permissions and attribution. Somebody has to make sure the model knows who it's allowed to quote.

Hexagon is buying Rocscience from TA for about $535 million (TA Associates, October 8). TA became the geotechnical software company's first institutional investor in June 2023 and is fully exiting. Slope stability software, three-year hold, strategic buyer. Not every exit this year needs an AI story.

Nvidia-backed Firmus pulled its $5 billion IPO (Bloomberg, October 9). The pricing implied a valuation of around $30 billion, nearly triple its August funding round, for a company running two data centres. Investors declined, and Firmus will raise privately instead, where the valuations are friendlier.

Acquisition Intelligence is a weekly newsletter on AI in M&A for finance professionals, private equity investors, investment bankers, corp dev teams, and deal-makers.

For questions, feedback, or to share what you're seeing in the market, reply to this email.

P.S. I'm Harry, co-founder of DealSage. When one of your numbers moves, we can follow it down into the records underneath, across every team that touches it. If you'd like to see that on one of your portfolio companies, reply and I'll show you.