Jenna O’Malley/PitchBook News
AI is getting more expensive, and most PE firms lack a framework for tracking their spending.
But the costs aren’t only financial: overreliance on AI is also dulling deal teams’ judgment and producing diligence work that reads as clearly AI-generated, testing the trust investors place in a firm.
PE firms are pushing AI use from pilot projects into daily deal work. Budgeting hasn’t caught up.
Siva Ilango, a London-based partner at JMAN Group, which advises PE funds on data and AI strategy, says most clients still do not understand the mechanics of token pricing, let alone budget for where it is headed.
AI models are priced per token, which correlates with the length and complexity of the text, with separate rates for what users send in and what the model generates. Output pricing is typically several times higher than input pricing, and running the same model on more complex, longer tasks compounds the difference quickly.
Ilango says most of his clients do not understand the mechanism and do not know how many tokens a typical week of use is actually costing them.
A well-known example is Uber. The ridehailing company burnt through its entire 2026 AI budget in four months after rolling out Anthropic‘s Claude Code to thousands of engineers in December, as agentic-coding adoption jumped from 32% in February to 84% by March, with 95% of engineers using some form of AI tool monthly by spring, according to Forbes.
Connor Kohlenberg, partner and London office lead for M&A at West Monroe, expects the same trajectory to hit the investment industry.
“Once you get hooked, it’s like a drug,” he said. “And then the drug prices always go up.”
Anthropic and OpenAI have now confidentially filed for IPOs—a shift that could push both companies away from the subsidized pricing that has defined the market so far.
When you think about deal teams in particular, you cannot make major errors, and decision fatigue is problematic because you are paid for how well your decisions are
Emily Cook, Found
Some firms are already adjusting. Ilango says the more disciplined clients are moving to per-user budget caps, typically £50 (around $67) to £100 a month, depending on their roles, rather than leaving usage unmetered.
Others are exploring a hybrid approach—routing lower-stakes, high-volume tasks to cheaper open-weight models such as DeepSeek or Qwen, and reserving frontier models for work that actually needs them.
Kohlenberg suggests firms should start running value-creation diagnostics that model AI’s cost over a five-year horizon rather than just the first year’s payoff. Then, pick a small number of genuinely high-value use cases instead of automating everything on the assumption that it is free.
“It is going to be really tough for companies to roll that back; it could have a negative effect in some areas, where instead of value creation, it is value erosion,” he said.
The cognitive toll
There are also other, more subtle ways in which improper AI use can turn investors away.
Emily Cook, founder of performance coaching firm Found, which works with PE and private credit firms on AI adoption and team performance, says the greater risk is how AI can affect teams’ cognitive capabilities and, in turn, decision-making if they are over-reliant on it.
AI is built to agree. A March study in Science found that across 11 leading AI models, the systems affirmed users’ actions 49% more often than humans did, even when the request involved deception or harm.
For an industry that depends on the ability to disagree with consensus when the data says otherwise, agreeableness becomes a structural risk when investment teams lean on AI that over-validates their thinking rather than challenges it.
The cognitive toll can also lead to decision errors. According to a BCG research paper, practitioners have started describing AI “brain fry”—mental fatigue from using, interacting with, or overseeing AI tools beyond a person’s cognitive capacity.
Among workers who reported experiencing it, decision fatigue rose by 33%, and major errors by 39%, compared with peers who didn’t.
“When you think about deal teams in particular, you cannot make major errors, and decision fatigue is problematic because you are paid for how well your decisions are,” Cook said.
Another risk is AI slop—content that is visibly AI-generated because no one has refined or iterated on it. In an investing context, it can show up in diligence responses, client updates, and management materials that read as if no person at the company actually stood behind them.
Cook cited a director in corporate finance who described a live process in which an investor withdrew after receiving diligence responses from a management team that was clearly AI-written and unrefined.
Drawing the line
To combat the overreliance, PE firms can first start by identifying which tasks belong to humans and which to AI.
Cook frames it with complexity and volume—tasks that are low-complexity and high-volume, for example, portfolio monitoring, data analysis, and presentation templates, are safe to automate outright, while more complex tasks, such as an investment memo, a financial model, or legal documents, work best as AI-assisted, where a person stays in the loop. But the high-level work, including fund strategy, negotiation, investment judgment and pitch delivery, should remain entirely human.
To ensure humans maintain their high performance over time, deal teams can also practice running some recurring tasks without AI from time to time and protect genuine deep-thinking time each week with no AI in the loop.
“How AI might be impacting their cognitive performance, how it might be impacting their wellbeing, and the actual quality of the output in some cases, I do not think that they are thinking about enough,” she said.
This article originally appeared on PitchBook News