Inside Advent's AI Playbook

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

… sat down with Goldman's Alison Mass on the Great Investors podcast for an interview that is, in my read, the most honest and clear-eyed thing a senior PE leader has said about AI on the record. The AI section is this week's deep dive.

Elsewhere: Anthropic closed a $65bn round at a $965bn post-money, KPMG put Claude in front of 276,000 staff, and EQT signed a parallel Google Cloud deal across 300+ portfolio companies. The Wall Street Journal reported corporate America is now rationing AI as token bills run away — the cost problem I flagged two months ago. Plus Forrester's State of Agentic AI 2026, JPMorgan's global IB rollout, and Polymarket opening contracts on private AI labs.

In This Week’s Issue:

From The Trenches

Advent's AI Strategy

Most of what gets written about AI in this industry comes from people with something to sell, and you can discount it accordingly. Brocklebank is selling nothing. He's a practitioner with nearly 30 years in the seat and a $100bn book. And Advent is not the firm you'd expect at the frontier of this: a pure-play buyout shop that prides itself on operational graft rather than technology, and has deliberately stayed out of credit, infrastructure and the rest. Which is exactly why how far along they are is worth studying.

What stands out is how specific he is about what they've built, how, and the order they did it in. I've pulled out the eight themes most worth discussing below. They apply to anyone thinking seriously about AI implementation, not just the investment-adjacent world. Original interview can be found here.

This is the first thing Brocklebank reaches for when Mass turns the interview to AI. Not a model. Not a tool. Not which lab to standardise on. Data.

Brocklebank names two problems in one sentence. The data lives in five places, none of them agree, and the agent that sits on top inherits the disagreement and returns confident nonsense. That's the first problem. The second is cultural, and it's the one most firms underrate: you can buy the tooling, but you cannot buy your team's willingness to change how it works around it.

There is no way around either of these. You can reach for the surface-level solution, and you'll get the surface-level result. Trying to shortcut it, or handing it to an analyst to sort out on the side of their desk, will not get you where you want to go. It's the messy, unsexy, genuinely painful work of standardising and centralising your data first, and then teaching people to change how they think about the work itself. There is no clever way to skip that part, and the firms pretending otherwise are the ones whose AI projects quietly die in year two.

If you want to go deeper on this: Connectors or Foundations (May 18) lays out the two ways firms are building this and why one of them quietly falls over in production, and To Build or Not To Build (May 11) on why you get the foundations right first and resist the urge to run an internal build sprint, because you'll be back at square one soon enough.

The line I can't stop turning over. He drops it almost in passing, describing where the operating model is heading.

Not apocalyptic. Structural. The right response to it is not to bolt agents onto the existing org chart, it's to completely re-evaluate what you're doing and ask a different question: how do we change the way we work in order to take full advantage of this? Because this is a genuinely revolutionary tool, and the temptation is to use it to do the same things slightly faster. The firms that win will be the ones that grasp how much more the AI can actually do, then reorganise around extracting the most from it rather than the least. Maximising that gap is the whole game.

If you want to go deeper on this: The Future Is Orchestration (Jan 5) on why the orchestration layer ends up as the centre of gravity rather than a sidecar.

His diagnosis of what's actually going wrong, and it's a good one.

This is the duct-tape problem. A consultant builds you a brittle stack of point solutions, the integrations rot the moment an upstream system renames a field, and a year later nobody owns it, nothing builds on it, and you're back at the start with a smaller budget and less goodwill. The distinction Brocklebank draws is the one that counts: AI transformation, singular, not AI projects, plural. Plural projects are what you commission when you don't have a thesis. Singular transformation is what you run when you do.

Same in venture, same on a banking floor. A drawer full of bolted-on tools isn't a strategy. It's a tax you pay monthly.

If you want to go deeper on this: To Build or Not To Build (May 11) on why running an internal build sprint to chase this leaves you back at square one, and Duct Tape Doesn't Scale (Jan 12) on why point-solution stacks rot.

The trade press covered this as "PE firm builds AI on its own deal history," which is the surface read. The actual detail is sharper.

The value here is not the AI. It's the proprietary data asset the AI was trained on, and the institutional knowledge that asset encodes. Thirteen years of committee papers, including the negative space: the rejected memos, the close calls, the deals brought to IC, killed, and then watched from a distance for three years. That reasoning normally walks out of the door when a partner retires or an associate leaves. Advent captured it instead, and turned it into something the firm can query.

What it does with that corpus is the other half of the point.

It challenges, it doesn't decide. A human still owns the call. That distinction is the line between a useful tool and a dangerous one, and it's why this works.

Worth noting how they got here on the model side, because it reinforces the same lesson. Advent built an early proprietary model trained on their own data, charmingly named Advent GPT, and have since moved to an intelligence engine that sits above the frontier models, is agnostic to all of them, and uses whichever fits. The model is interchangeable and getting more so by the quarter. The asset that compounds is the data it reasons against. As the models converge and everyone ends up renting the same intelligence from the same handful of labs, your proprietary corpus and the institutional knowledge inside it are the only things that are actually yours.

If you want to go deeper on this: The Model Doesn't Matter (Mar 9) on why the corpus beats the model, Show Your Workings (Feb 2) on why an AI in the room has to be a prompt and never a vote, and The System of Record Problem (Dec 15), the early piece on why your organised proprietary data is the one moat nobody can buy.

On hiring.

The org-chart point from line two, made concrete. Not a deal team plus a data team. One team, with the data fluency baked into the seat rather than bought in as a service line. The mechanical reason it matters: a single integrated team moves at the speed of one team, while two teams move at the speed of the handoff between them, and the speed at which a firm absorbs AI is fast becoming a real factor in how its deals get evaluated. He did add the honest caveat that juniors still have to understand how a model is built before they outsource it to an agent, which is the tension nobody running a desk has fully resolved.

If you want to go deeper on this: The Only Metric That Matters (Feb 16) on why implementation speed is becoming the number to watch.

For all the in-house capability - a chief data science officer, a chief digital officer, a proprietary engine - Advent did not try to do all of it alone. They're a founding partner in OpenAI's $4bn deployment venture, the arm built specifically to get this stuff implemented at scale. Read that against everything above: even the firm furthest along, the one that built its own IC robot, concluded that implementation is hard enough to warrant a dedicated partner. If Advent needs that, the firm without Advent's resources needs it more, not less.

If you want to go deeper on this: The Un-Automatable Edge (Apr 20) on what's left once AI has collapsed the cost of the easy work.

His advice to his sons, both heading into finance.

The point of building all the infrastructure above is not to remove the people. It's to clear their time for the part only people can do: the relationship, the judgement, the conviction to act when the data is ambiguous. AI hands everyone the same superpower, so the durable edge moves to the things that don't come out of a model. A 30-year veteran of a $100bn fund landing on "be human" as his closing thought is worth sitting with.

His lightning-round answer when Mass asks what he's most excited about in the world right now.

As it should be. Genuinely new technology, and the scale of change that comes with it, usually is.

When a co-chair of a $100bn firm calmly tells a Goldman podcast that the data comes first, the culture second, the corpus is the asset, the team is one team, and you rebuild around the AI rather than bolt it on, it stops being a contrarian take from a few people at the edge. It's the centre of gravity moving.

If all of this sounds right and you'd like it for yourself, but you don't have a chief data science officer, a team of data scientists, or a seat in a multi-billion-dollar deployment JV with the frontier labs, that's precisely the gap we built DealSage to fill. Give us a call.

News Digest

Anthropic At $965bn, And The Shape Of The Distribution Play

Anthropic closed a $65bn round at a reported $965bn post-money on Thursday, briefly the most valuable AI startup in the world, ahead of an IPO filing reported to be weeks away. Three things happened around the raise that are best read as one story.

KPMG confirmed it is rolling Claude to 276,000 employees and positioning itself as Anthropic's preferred consulting partner specifically for private equity work. The Anthropic-backed enterprise services firm that Blackstone and Hellman & Friedman seeded last month closed its first acquisition, Fractional AI, to staff the implementation arm. And Apollo and Blackstone are marketing a $36bn debt deal to buy Google TPUs and lease them back to Anthropic, with Broadcom backstopping the residual value.

The details:

Why it matters: Capital, distribution, implementation and compute are now being assembled around one frontier lab in parallel, with PE houses doing the syndication. This is not a series of funding stories. It is the operating system layer of the enterprise being built in public.

My take: The Anthropic march continues. The bottleneck stopped being the model some months ago, an argument I made directly in Connectors or Foundations (May 18), and the constraint that matters now is implementation. Anthropic is racing to capture that constraint. KPMG handles distribution. Fractional AI handles implementation. Apollo and Blackstone handle the compute supply. The $65bn round pays for everything else. None of it is the model.

They have a window. Model capabilities will converge over the next 18 to 24 months, and when they do there will be very little to choose between the frontier labs at the capability layer. The moat is whatever gets built around it before that convergence lands. The Big Four channel. The implementation muscle. The PE-backed services arm. The data centre supply. Anthropic clearly know this. Every move this week is about locking the distribution in before the narrative shifts back to one of the other labs, or away from frontier labs entirely.

Read more

Corporate America Starts Rationing AI. The Token Bill Arrives On Time.

The Wall Street Journal reported on Thursday that large US corporates have started actively rationing AI usage internally as enterprise token spend runs away from budget. Capping seat counts. Throttling agent runs. Routing more queries to cheaper models and reserving the frontier labs for what actually needs the reasoning. Companies in the piece include household-name enterprise IT buyers who told the WSJ their AI bill is now their second-fastest growing line item and is on track to displace cloud spend within 24 months.

This is the second WSJ piece on enterprise AI cost overruns in eight weeks. It follows the Financial Times piece from late March that I wrote up the week after in So What Is This Really Going To Cost? (April 6). The argument then was that the real cost of enterprise AI is hiding in your token bill and that the firms acting like it does not exist would be the ones running out of runway first. Eight weeks later, the WSJ is naming the firms doing exactly that.

The details:

Why it matters: The cost problem is not finance-sector specific, but the lesson is. The firms running ungoverned token bills hit a wall first. The ones with a routing layer that picks the right model for each step, and an architecture that does not throw the whole deal room into context on every query, hold their economics.

My take: Cost of model is going to be the defining theme of AI implementation over the next 12 months, not model capability. The capability race has obscured the cost race, and the cost race is the one that is about to start determining which firms hold their economics and which firms don't.

Two pieces of architecture get you through this. Orchestration systems that dynamically route each step of each task to the most appropriate cost tier, so that frontier-model spend is reserved for the queries that genuinely need it. And tight management of what you send to the model and what you ask it to send back, because every input token and every output token is paid for. I made exactly this argument in Connectors or Foundations (May 18). The connector pattern, with one frontier model in the middle reaching out into five systems via MCP, is the most expensive way to run any of this. The foundation pattern, with one connected data layer and dynamic model routing in front of it, is meaningfully cheaper at the same output quality.

Read more

EQT And Google Cloud, 300+ Portfolio Companies, And The Sister Bet

EQT announced on Wednesday a partnership with Google Cloud to push AI across more than 300 portfolio companies. They get the Gemini Enterprise Agent platform and full model range, Google's Mandiant and Wiz cybersecurity stacks, sovereign cloud for compliance, and Google engineers working alongside EQT's in-house AI transformation team. The underplayed detail: portfolio companies also get a route into Google's 330,000-strong partner consulting network, and software portcos get a Marketplace listing and co-sell access into Google's enterprise base. First pilots: Believe, Epidemic Sound, Keyword Studios, Zooplus.

This is the second sponsor-frontier mega-deal in a week, after KPMG-Anthropic above, and the mechanics are identical: a central contract, a distributed playbook, the lab handling capability and a partner handling embedment.

The details:

Why it matters: Two of the world's largest sponsors have placed parallel bets on different labs inside a week. This is what PE-scale AI procurement looks like now: central negotiation, sponsor-scale pricing, centralised rollout.

My take: The "accelerating AI adoption" framing undersells the real move, which is procurement leverage. A 300-company contract gets unit economics no single mid-cap would ever see at list price. But it also means differentiation can't come from "we use AI," because every company in the programme has the same access. It comes from the data each one sits on and where the agent gets embedded — the Brocklebank point above, at portfolio scale. The choice signal is the one to watch: Anthropic via KPMG, Google via EQT, OpenAI via the deployment line-up Advent, TPG, Bain and Brookfield co-led last month. The labs are no longer picked by companies one at a time. They're chosen by sponsors on behalf of whole portfolios.

Read more

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

Forrester: The State Of Agentic AI, 2026 (May 29) - The headline finding is that agentic AI has reached technical viability but most enterprises are still stuck between promise and payoff. Worth the click for the platform, pricing and governance sections, which Forrester reads as in active flux through the rest of 2026.

JPMorgan rolls out AI tools to its investment banking arm globally (Reuters, May 21) - JPM's senior banker tells Reuters the IB-wide rollout covers the full advisory stack: pitch prep, sector research, comparable analysis, model build, IC support. The "deal team is the data team" line from the FTT, executing in real time on the largest IB floor in the world.

Anthropic launches Opus 4.8 and queues Mythos for general release (Reuters, May 28) - Opus 4.8 ships now, Mythos goes broad in the coming weeks. The $965bn round above is being underwritten on the assumption Mythos clears its safety review and lands roughly on schedule. Worth watching whether the safety review actually permits the broad release Anthropic is now publicly committing to.

Snowflake to acquire Natoma (May 28) - Snowflake is buying an enterprise Model Context Protocol platform to extend governance from data into agent actions. Sensible defensive move from a data warehouse that did not want to wake up in 2027 watching every agent in its customer base auth into other vendors' platforms.

NTT Data to acquire WinWire (May 24) - NTT picking up 1,000+ Azure engineers and an "Agentic AI @ Scale" framework. The implementation roll-up trade we wrote about with PwC, Tomoro and Fractional now has its Asia-Pacific entrant.

SoftBank to build out AI data centres in France (Reuters, May 30) - Major capital commitment into French data centre capacity, framed as European sovereign AI infrastructure. The Apollo/Blackstone $36bn TPU deal above is the same trade in a different jurisdiction. Every PE house with a private credit arm is going to be sourcing data centre paper by Q4.

SpaceX to acquire Cursor 30 days after IPO (May 19) - Reuters reports SpaceX expects to take Cursor private essentially the moment the IPO lockup clears. As pre-baked exit structures go, this is the most pre-baked one I have ever seen.

Polymarket launches private-company prediction markets (FT, May 24) - SpaceX and Anthropic both now have tradeable contracts. Price discovery on private AI labs by way of crypto-native prediction markets. The bull and bear cases for the future of finance, both visible in the same headline.

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