Notes From The Floor

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

… Acquisition Intelligence is a weekly read on AI in finance, where I pull together the latest stories on AI deployment, what's working, what's not, plus the considerations, challenges and guidance from actually building on the front lines. Recent issues have covered how sourcing is changing in an agentic world (The Un-Automatable Edge), the hidden costs of AI and how to avoid them (So What Is This Really Going To Cost?), and how agents actually work and how to get the best of them (How Agents Actually Work).

In News Digest this week: Anthropic and OpenAI both set up multi-billion-dollar joint ventures with Wall Street and PE to push their tools straight into portfolio companies. Blackstone formed a new West Coast AI division and KKR raised $10bn for an AI infrastructure operating company led by ex-AWS CEO Adam Selipsky. Citi turned on an agentic platform across all 180,000 employees the same week Accenture started rolling Microsoft Copilot out to its 743,000.

Plus in Other Interesting Things: Google's $40bn Anthropic commitment, Anthropic itself reportedly weighing a $50bn round at a $900bn valuation, Coatue moving into data centre land, junior-banker AI tool Rogo at $2bn, and Apoorva Mehta's all-AI hedge fund.

But first, my DealMax debrief. The seven most pertinent themes I heard, and what I think they actually mean.

In This Week’s Issue:

From The Trenches

What I heard at DealMax, and what it actually means

Five days. Hundreds of conversations. PE across the size spectrum, a wave of independent sponsors, a few corporate development teams, and more bankers and advisors than I could count.

The themes were remarkably consistent. The same questions came up over and over, in slightly different forms, from people running very different businesses.

Here are the seven that came up most, and my take on each.

Most people had been issued a licence in the past few weeks, and the feedback was generally great. As it should be. But the levels of use vary wildly. Some are still barely doing more than writing emails with it. The interesting bit was that even the people who'd had it for months said they hadn't seen the transformative changes the AI marketing promises.

The penny dropped in one conversation. Writing the memo was never actually the bottleneck. Doing it 30% faster doesn't move the needle. The real breakthroughs aren't in shaving 20% off a single workflow. They're in unlocking questions you couldn't ask at all before. That needs the data connected, not just the model available. The trap is treating Claude as a faster typewriter. The value sits one floor up.

Almost everyone was focused on portfolio company AI work, with very little at the fund level. At the fund itself, it was each person doing their own thing. A few power users. Some laggards. Nothing scalable.

This builds on the point above. Getting Claude access is step one. The harder question is what to actually do with it. Where are the real bottlenecks. What can your firm now do that wasn't possible last year.

Most people's instinct is to point AI at automation. Automate my job, automate this report, automate that update. Useful, but it's the smallest lens. The bigger one is that AI now lets you connect systems and unlock data that has been sitting in PDFs, emails, decks and spreadsheets for years. You can ask questions of your portfolio, your pipeline and your network that you couldn't ask before. The firms pulling ahead aren't the ones with the most automated tasks. They are the ones that have started asking new questions.

Lots of demand for agents. A lot of attempts to build them. Very few that survived past a few weeks.

This is the gap between the demos and the reality. Agents are very hard to actually deploy. You run into data access issues, consistency problems, edge cases that break the workflow, integrations that worked last month but don't today. I run a handful of agents to help operate DealSage and they need constant tweaking. I've spun some up thinking they'd be useful and killed them after a month because they weren't.

It's why we built DealSage to handle this for financial workflows specifically. Agents work properly when the data underneath them is centralised, clean and coherent. Without that, they fall apart in production no matter how clever the orchestration. The market is wildly underestimating how hard this is. The demos are real. The 9-month implementation projects on top of those demos are also real.

Lots of trepidation around AI from a data security perspective, and very little real understanding of where the actual exposure sits.

The question I heard most was "do they train on our data?" The answer is well-documented and it's no. Anthropic and OpenAI don't train on enterprise inputs by default, and any AI provider worth working with will have the same protections built in: data residency, SOC 2, HIPAA paths where relevant, no training on customer data. We've also been hosting sensitive deal documents on Google and Microsoft for a decade. The frontier-model question is, in many cases, the same one we already answered for cloud.

The real exposure sits somewhere different. The fastest way to leak sensitive data right now is to build a custom agentic system in-house without the people who know how to do it securely. Hand-rolled agentic infrastructure, built by teams that don't specialise in it, is the largest under-discussed risk in deal tech today. That's the conversation we should be having, and it's the one no one is.

This used to be vendor fatigue. It has become velocity fatigue. Everything is moving so quickly that nobody knows which solutions are legitimate, which will exist in twelve months, and which are worth signing a multi-year contract on.

Resist the allure of solving point problems with point tools. That is how you end up with twenty vendor licences, twenty integration headaches, twenty governance reviews and twenty places your data is sitting. Be wary of multi-year contracts and closed architectures. The market is moving fast enough that the right answer is to back people who can answer specific operational questions today and stay flexible enough to swap things out as the picture changes.

Underneath it all: the number one problem every firm I spoke to ran into when trying to deploy AI is the data itself, not the AI. Disorganised, scattered across systems that don't talk to each other. Until that is fixed, no model and no vendor will give you what you want. A big part of what we are building DealSage to solve.

Universal frustration with legacy software, particularly the CRMs. Inflexible. Don't integrate with anything modern. Don't expose the APIs that agents need. Even the ones now claiming MCP support do it in a limited and frustrating way that isn't really worth the integration cost.

Part of this is symptomatic of something broader. The finance industry has historically not run on much software at all. The agentic era is throwing a sharp light on how limited a lot of the existing tooling is. Old systems that worked fine in a manual era are revealing themselves as bottlenecks in an automated one.

What is new this quarter is that the frustration is finally turning into willingness to switch. People are no longer just complaining about it. They are planning migrations and asking whether they can consolidate the rest of the stack at the same time. The traditional CRM moat was switching cost. That cost is collapsing because agents can now do the migration themselves. The next twelve months are when this finally moves.

A line I heard repeated by several people: "If the tool delivers value, we will pay for it." Cost is a factor, but not a deterrent.

The interesting bit is what that implies. Software used to clear the value bar by saving someone an hour a week or organising data that was previously in a spreadsheet. The bar is now closer to "augments or replaces an associate" or "lets us answer a question we couldn't answer before." That is a much higher bar.

This is good news for buyers. Firms are ready to invest properly in tools that genuinely move the needle, but the threshold for what counts as moving the needle has shifted. A tool that does what an associate would have done is worth more than seat-based pricing of a year ago. A tool that lets you do something an associate couldn't have done is worth more again. Cost becomes the secondary question. Whether it actually clears the new bar is the primary one.

AI access is now table stakes. The breakthroughs aren't in personal productivity, they are in connecting your data and asking questions you couldn't ask before. Building those systems in-house is harder and riskier than it looks. Buying them needs to clear a higher value bar than it used to. And the legacy software that has been holding firms back for years is finally being looked at with the question it deserves: do we still need this.

Some firms have stopped treating AI as a faster typewriter and started asking what becomes possible now that wasn't possible eighteen months ago. Most haven't. The gap between those two groups will start showing up in fund performance soon enough.

News Digest

The Labs Want To Be Your GP

The Wall Street Journal broke on Sunday night that Anthropic is finalising a $1.5bn joint venture with Blackstone, Hellman & Friedman, Goldman Sachs and General Atlantic. The announcement is expected as soon as Monday. Blackstone, H&F and Anthropic itself are each putting in roughly $300m. Goldman is a founding investor at around $150m. The vehicle will operate as a consulting arm for Anthropic, with an explicit mandate to teach private equity portfolio companies how to incorporate AI across their operations.

The Financial Times reported on April 22 that OpenAI is in parallel talks to commit up to $1.5bn to its own private-equity-backed joint venture, internally called DeployCo, with the same target audience: PE portfolio companies that need engineering capacity to deploy AI.

Two of the three frontier labs, the same fortnight, the same number, the same playbook.

The details:

Why it matters: The PE value chain just acquired a new disintermediator. PE has been treated as the customer of AI tools for the past two years. The two leading labs just decided they would rather be the consulting partner than the supplier, and they've recruited the largest sponsors to write the cheques.

My take: This is a consulting business owned by the lab and its anchor LPs, not a fund. Anthropic gets a captive distribution channel that Big 4 and McKinsey can't match on speed or specificity, and Blackstone, H&F and Goldman get a multi-year head start on operational AI deployment inside their portfolios at sub-retail pricing. For everyone else, the asymmetry is uncomfortable: the firms that struck a deal first get the head start, while the firms that wait will end up paying for the same capability at retail later, or watching their portcos lose ground to portcos owned by the anchors. That gap will start showing up in fund-level performance some time in 2027.

But the more interesting read sits one layer up. Three of the most sophisticated capital allocators in the world looked at "how do we deploy AI properly across our portfolios" and concluded that the answer was not to build it themselves and not to wait for the next ChatGPT release. The answer was to commit nine-figure cheques to bring the lab itself in to architect, build and deploy the systems for them. That is what taking AI seriously looks like at the top of the market.

Read more

Private Equity Restructures Itself Around AI

While the labs were setting up their distribution arms, the largest private equity managers were reorganising their own firms around AI as a primary investment theme.

On April 29, Bloomberg reported that Blackstone is folding its growth business into a new West Coast division called Blackstone N1, focused entirely on its AI portfolio. The unit will be led by veteran executive Jas Khaira and will serve as the centralised AI investing resource for the rest of the firm. It will sit on top of stakes in OpenAI, Anthropic, CoreWeave and SpaceX, run alongside Blackstone Private Equity Strategies and Tactical Opportunities, and replace Jon Korngold's previous Blackstone Growth (BXG) leadership after what Bloomberg politely described as "an uneven run."

Two readouts from Schwarzman and Gray's internal memo stood out. They say AI is now "reshaping every business at the firm." And they note that eight of Blackstone's ten best-performing investments last quarter were in the AI ecosystem.

The next day, Bloomberg reported that KKR has secured more than $10bn to launch Helix Digital Infrastructure, an operating company that will design, build, own and run AI compute infrastructure. Helix is led by Adam Selipsky, the former AWS CEO who took the cloud business past $100bn in annual revenue. He'll target hyperscale data centre, power and connectivity capacity tied to AI buildouts.

The two moves are mechanically very different. Blackstone N1 is an investment division managing equity stakes. Helix is an operating company that owns and runs physical assets. But the strategic signal is the same.

The details:

Why it matters: Two of the largest alternative asset managers in the world standing up dedicated organisational units around the same theme in the same week tells you AI has been promoted from "interesting sector" to firm-level strategic priority.

My take: The interesting bit isn't that PE is "investing in AI." Everyone has been doing that for two years. The interesting bit is that the largest firms are now reshaping their internal architecture for it. Blackstone is centralising oversight. KKR is building an operating company so it can directly own infrastructure rather than buy it. Both moves accept that AI capex is going to be a multi-decade allocation, not a thematic trade.

The deeper signal sits underneath both stories. Every business is an AI business now. Optionality is gone. You are either implementing AI as best you can, or you are investing with AI as the core lens. Anything else is a bet that the next decade looks like the last one, which it won't.

The biggest firms in the world are showing the playbook. Standing units. Dedicated leadership. Direct ownership of infrastructure. Direct partnerships with the labs (per the Anthropic JV story above). They are not waiting for the picture to clarify. They have decided the picture is clear enough.

The rest of the buy-side should read it the same way. The argument that "we'll build the AI capability when the deal flow demands it" is rapidly becoming the argument of a firm that won't see the deal flow.

Read more

Wall Street Just Shipped Agentic AI At Scale

Two deployment stories in the same week, both at scale.

The more interesting of the two: Citi introduced Arc on April 30, an internal platform that lets the firm build, deploy, monitor and govern AI agents across the organisation. Per Axios's exclusive on the same day, Arc packages multiple frontier models into one centrally audited environment, with the ability to halt agent tasks mid-flight. Citi's framing is "operating system for agents," not "AI for productivity," and that distinction matters for how regulated firms approach this.

Three days earlier, Reuters reported that Accenture is rolling Microsoft Copilot out to all 743,000 of its employees, the largest single enterprise rollout of agentic AI to date.

The details:

Why it matters: Two of the largest financial services and professional services organisations in the world just crossed from pilot to production-scale agentic deployment. The implications for how regulated industries deploy AI are now visible at scale.

My take: The interesting question is why Citi built rather than bought when the enterprise products from Anthropic and OpenAI cover most of the ground off the shelf. The answer is governance, specificity, and the need to stay flexible at the model layer. A regulated bank running 180,000 employees needs audit and shutdown capability built around its own risk framework, agents that handle bank-specific workflows rather than generic ones, and the option to swap the underlying model when the next frontier release shifts the picture. That's a different shopping list to "give us your enterprise seats."

The read-across for PE is the same shopping list. Off-the-shelf seats are fine for individuals running their own workflows. Anything you want to scale across an organisation needs governance, specificity to your processes, and the ability to swap the model later as the field moves. Whether you build that yourself or buy it from someone who has already built it is a separate question, and most firms don't have the engineering bench to make build the right answer.

Worth noting that Arc starts with developers building agents for "specific, well-defined use cases" before any wider rollout. The headlines run ahead of the deployment. Where this actually lands across 180,000 employees in twelve months is genuinely an open question. The structural bet is right; the execution risk is real.

Read more

Other Interesting Things I’ve Read of Seen This Week:

Google plans to invest up to $40bn in Anthropic (April 24) - $10bn now, $30bn contingent on performance, plus multi-gigawatt TPU compute commitments. The "neutral cloud partner" pose has officially been retired.

Anthropic considering offers at $900bn valuation (April 29) - $50bn round on the table at a 15x re-rate from six months ago. The "OpenAI is the unassailable leader" anchor that quietly underwrote every AI valuation comp since 2023 just moved.

Coatue starts data centre land-buying venture (May 1) - The growth fund that backed Snowflake and Anthropic now wants to be your landlord. Hard to think of a clearer signal that the easy software returns aren't where the next decade of capital is going.

Rogo, the junior-banker AI tool, raises at $2bn (Bloomberg) - Two ex-bankers built the thing they wished they'd had as analysts. Now Wall Street is buying it back. A near-perfect microcosm of how AI is reshaping the cost stack of investment banking, told through a single startup's cap table.

Apoorva Mehta launches Abundance, an AI-only hedge fund (April 24) - The Instacart co-founder's new fund uses thousands of AI agents to source ideas, size positions and execute trades. Track record TBC. The business case is one PM and a server room.

Mercor at $10bn pays white-collar workers to train AI (Bloomberg, April 29) - The startup hiring lawyers, doctors and bankers to record their workflows so models can learn them. The cleanest expression yet of "your job is the dataset."

Cohere merges with Aleph Alpha at a $20bn valuation (April 24) - Two also-rans become one slightly larger also-ran with a new flag and a $600m cheque from Schwarz Group. The European industrial-policy angle is more interesting than the cap table.

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. We're building an AI-native deal intelligence platform to help professionals turn their institutional knowledge into better decisions. If you're curious what we're up to, check out dealsage.io or just reply here