The Loss of Flow State

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.

AI has made the work remarkably easy to produce. I'm less sure we've thought about the working day that's left.

Goldman Sachs now expects the five largest hyperscalers to spend $1.2 trillion on AI infrastructure in 2027, roughly half as much again as this year. Jamie Dimon put next year's figure at about $1 trillion, and an FT investigation found Big Tech keeping $300 billion of AI exposure off its balance sheets through guarantees.

Closer to home, Hg says it has more than 1,600 live AI projects across its portfolio, with over $260 million of budgeted EBITDA attached. Microsoft, meanwhile, published what it learned rolling AI out across its own business, and most of it comes down to fixing the process before adding the agents.

Elsewhere: H&F's $10 billion Applied Systems sale, Ropes & Gray building diligence software with OpenAI, Databricks buying a spreadsheet and Goldman's rather lucrative hedge fund client.

But first, my take on what happens to the way we work when producing the work stops being the hard part.

In today's Acquisition Intelligence:

From The Trenches

One thing I’ve noticed as more of my work moves through AI is how much of the day becomes reviewing.

There is always something ready for a response. A draft to read, an analysis to check, another version to look through. I can get far more things moving, and getting something to a respectable first pass is becoming remarkably easy.

But I miss getting properly absorbed in a problem. Staying with it long enough for the pieces to connect, rather than repeatedly returning to inspect what has happened while I was elsewhere.

More Material Than Understanding

Slava Akhmechet described a related problem in enterprise AI adoption: executive pressure creates plenty of experimentation, but much of the resulting output is slop that eventually gets abandoned. He was optimistic about the applications that actually worked. His criticism was of all the activity surrounding them that accomplished very little.

The pattern is recognisable: the production of work accelerates, and the people involved become occupied with keeping that production moving.

It is easy to see how this happens. A lot of professional work is organised around producing materials. Reports, presentations, project plans. They are tangible evidence that something has happened, so they become a convenient way to measure progress.

If we can produce ten times as many, it doesn’t follow that we have understood ten times as much. Or made a better decision. Ten reports can still leave the same question unanswered.

As those materials become easier to produce, their polish tells us less. A well-structured document can arrive before anyone has properly worked through the argument inside it. Someone else then has to establish whether it deserves their attention.

The cost of producing it falls. The cost of reading it has been passed to a colleague.

Getting something to 80% creates its own temptation. Once the first pass costs almost nothing, you start far more things. Another analysis, another presentation, another version of the same argument. Each looks sufficiently promising to deserve finishing.

You end up with a much larger queue of almost-useful things, all waiting for the same person.

The benefits are real. There are tasks I’m delighted to hand over, and questions I can now explore that would previously have taken too long to justify. I spend my days building tools to make more of that possible.

But I’m less sure we’ve thought through the working day we’re assembling out of what remains.

Amy Gandon put it more starkly on X: “There is no joy that comes from a life without effort or challenge or mastery.”

I wouldn’t extend that into a defence of every laborious task. Nobody needs to preserve the character-building experience of reformatting a presentation at midnight. But I recognise the concern. Some of the satisfaction in work comes from wrestling with something until you understand it. Remove that process and you can lose more than the hours it occupied.

How Stuff Sticks Out

This came up in a recent conversation with a client about automating his financial models.

He welcomed having the data extracted. Pulling numbers out of documents and putting them somewhere usable was an obvious place to save time. But he also explained why working through his model mattered to him.

“That’s how stuff sticks out to me.”

Putting the numbers into a familiar structure helped him spot changes, inconsistencies and assumptions he wanted to challenge. He expected automation to help, and potentially catch things he would miss. He still wanted to participate in the thinking.

The same thing happens when you write a report. You start with a rough explanation of what happened. Then you try to support it, discover that two pieces of evidence don’t agree, and realise the explanation needs to change. The finished report is one result. Your improved understanding is another.

Some of that understanding develops while you are trying to make the thing work. You don’t always know which question matters until you get stuck.

Where The Flow Comes From

That is also where the flow comes from, at least for me. You hold the problem in your head, follow a thread, make a connection and keep going. There is satisfaction in gradually getting somewhere.

Constant review feels different. Each time something comes back, you have to reconstruct enough of the problem to judge the answer. With several things moving at once, that reconstruction can become a large part of the day.

I can almost feel this in my own head some days. I get to the end of the day having moved a great deal forward, with plenty to show for it, and still without the satisfaction that used to come with a productive day. The work got done, but I spent most of the day checking it rather than doing it.

It’s possible I’m projecting, and this says more about me than about how work is changing. I’d be curious whether anyone else has noticed the same thing.

Review can be demanding, thoughtful work. But a document that looks mostly finished makes it easy to slip into checking what is there, without asking what should have been there in the first place. What happens when we spend less time forming an explanation and more time approving one?

Choosing What To Keep

Thariq, who works on Claude Code, posted this weekend that what he fears most is “us eating the productivity gains of agents by just becoming lazier.” I think there’s a second way to lose them.

The temptation is to take all this new capacity and fill it. More reports, more analysis, more versions. But if multiplying the output barely improves the result, that should make us question the role the output was playing in the first place.

We have spent years organising work around things that are easy to count and circulate. Now we can produce those things almost at will. It becomes much harder to avoid asking what they actually achieve.

A report should earn its place by helping someone understand a problem or take action. If nobody uses it, making it in thirty seconds is a limited victory. We could also stop making it.

And where the work does matter, we should understand what makes it valuable before removing ourselves from it. Sometimes the process is how you develop a view. Sometimes it is how you stay sharp. Sometimes it is simply a part of the job you enjoy.

Those are legitimate things to protect. We don’t have to turn every capability of the technology into an obligation to use it.

I want AI to give me more room to get absorbed in the problems I care about. If I use it to create an endless queue of material that keeps me from those problems, I have made my own working day worse.

The ability to produce more gives us a choice. Some of the benefit should show up as work we no longer have to do, and some as time we can finally spend doing the work properly.

News Digest

Microsoft Reads Its Own Instructions

Microsoft has spent two years telling every company to roll out AI. On September 17, Kathleen Hogan published what happened when it did the same to itself. The admission comes early: "a tool licensed to over 200,000 people does not change how the work gets done." The gains came once teams redesigned the work around the agents.

The details:

My take: Hogan's best line is that "adding agents to a broken process still leaves a broken process." That's the order we build DealSage in, with the data and workflow first and the agents last. And if Microsoft, with its stake in OpenAI, is telling itself to stay model-independent and own its institutional knowledge, a mid-market portfolio company should probably take the hint.

Read more

The Build-Out Gets A Bigger Bill

Goldman Sachs strategists raised their forecast on September 25: the five largest US hyperscalers will spend $1.2 trillion on capex in 2027, up about 50 per cent on this year's roughly $800 billion, and $1.4 trillion in 2028. Four days earlier, Jamie Dimon told JPMorgan's India conference that spending across the AI ecosystem had gone from about $300 billion last year to around $700 billion now, and could hit $1 trillion next year.

How it gets paid for is becoming the more interesting story. The FT reported this week that Big Tech is using guarantees and special-purpose vehicles to keep around $300 billion of AI exposure off its balance sheets.

The details:

My take: The Banca d'Italia paper is the one I'd send to a credit committee. Bond markets are already pricing AI winners and losers by sector, and the software businesses that make up so much of sponsor and private credit portfolios are on the wrong side of that split. Meanwhile the biggest buyers are financing the build in ways that don't show up on their own balance sheets, which leaves someone else holding the risk if $300 billion of revenue arrives late.

Read more

Hg Puts A Number On Its Anthropic Partnership

Hg extended its partnership with Anthropic on September 22, and for once a sponsor has published figures you can hold it to. The programme, which began in early 2024, now runs more than 1,600 live AI projects across the portfolio. Hg says they carry over $260 million of budgeted EBITDA impact, five times the level when the partnership started.

The details:

My take: "Budgeted" is carrying a fair amount of that $260 million. I'd want to see how much of it turns up in the accounts. Even so, Hg has the structure right: the portfolio company keeps the system of record, which is what a buyer ends up paying for, while the model on top gets swapped whenever a better one arrives. That's how we build DealSage too. Hg is also the first sponsor I've seen tie AI work to named exits, and I expect that's how the rest of the industry will start telling the story.

Read more

Other Interesting Things I've Read or Seen This Week

H&F explores a sale of Applied Systems at up to $10 billion (September 24). H&F paid about $1.8 billion for the insurance software provider in 2014, and it now makes more than $550 million of EBITDA. So much for software sponsors being unable to exit anything.

US companies turn to open-weight models to rein in AI bills (FT, September 27). Mentions of open-weight or open-source models on earnings calls and at investor conferences rose sixfold in August and September on a year earlier, according to AlphaSense, with PNC, CH Robinson and Siemens among those discussing them. Tinder's CTO says its AI spend went from an annual rate of $1 million in January to $10 million by July, which would focus anyone's mind.

BNP Paribas signs a five-year AI deal with Google Cloud (September 24). Gemini goes into the corporate and investment bank first, where its AI chief says it will help "be faster in how we produce corporate credit memos". The bank is aiming for €750 million of value from AI by the end of this year, and someone will still need to read the memos.

Ropes & Gray is building a diligence tool with OpenAI (September 18). It's designed to follow the firm's own data-room review steps and produce an issue-level report "in hours, not weeks". I look forward to seeing that reflected in the fee.

Databricks buys spreadsheet startup Row Zero (September 24). Ali Ghodsi says every business analyst loves spreadsheets, and plans to put Databricks' Genie agent inside one. Excel outlives another prediction of its demise.

Tobi Lütke says his staff are throwing "slop grenades" (September 17). The Shopify boss, who made AI use a baseline expectation last year, told The Knowledge Project that people now pass round AI output nobody has read, leaving colleagues to review it. A BetterUp and Stanford survey puts the clean-up at 3.4 hours a month for those on the receiving end.

Accenture invests in Within (September 23). The software watches how employees really work and compiles it into a "Work Brain" for AI agents to learn from. Do let your team know before you install it.

OpenAI's agents leaked 53 ChatGPT users' images (Reuters, September 25). Two months after its agents hacked Hugging Face, OpenAI is still combing its logs for what else they got up to, and says the review will take "months". Worth remembering the next time a vendor asks for broad access to your data room.

Three in four Americans say AI firms aren't doing enough to prevent disaster (Reuters/Ipsos, September 22). 73 per cent want safety prioritised over staying ahead, and 55 per cent think slowing AI down would be a good thing. The labs asking to be paced may find the public happy to oblige.

Goldman made more than $200 million in fees from Situational Awareness (FT, September 26). Leopold Aschenbrenner's two-year-old AI hedge fund grew past $20 billion on borrowed money, then sold around $16 billion of positions to Citadel after the AI sell-off. Goldman did fine either way.

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. A good part of every client build is deciding what not to automate, because some of the work is where your team does its thinking. If you're working through that question for your own firm, reply and tell me where you'd draw the line.