To Build or Not To Build

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

… and announced ten pre-built AI agents for financial services. The technology underneath is much smaller than the staging suggested. The "agents" are Anthropic Skills, which is to say plain-text instructions in a folder, a concept that's existed for six or seven months now. Couple that with Reuters reporting the same week that the labs' PE-backed JVs are in advanced talks to acquire AI services firms, and the read is fairly obvious. Models aren't going to be improving that dramatically for the foreseeable future. The model isn't the bottleneck, and hasn't been for a while.

Bain followed up by putting an institutional number on what comes next. $100bn of US enterprise spend, $200bn globally, sitting in what they're calling "cross-system labour": the manual work people do bridging ERPs, CRMs, billing, vendor management. Over 90% of it is uncaptured. The pricing model SaaS has run on for fifteen years is about to break, and PE's software book is going to feel it first.

Plus in Other Interesting Things: ECB defences against AI-powered attacks on payment infrastructure, S&P data on dimming software exits, Medallia becoming the first really visible PE software restructuring of the AI era, and McKinsey using AI agents to staff its own client teams.

But first, my take on what I've started calling the build trap. It's the most common conversation I've had with PE firms in the last six months, and nearly every one of them is getting it backwards.

In This Week’s Issue:

From The Trenches

The Build Trap

To build or not to build, that is the question.

I've been hearing the same conversation in different rooms for the last six months. A senior partner at a mid-market PE fund pulls me aside at a conference, or starts a call ten minutes early, and floats the same idea: we're thinking about hiring an engineer or two and just building our AI tooling internally. They've had their first really good Claude moment. Something genuinely awakened. The conclusion they've drawn is that the firm should build.

I get it. The early experience is genuinely exhilarating. It elicits this almost child-like wonder of creating something out of nothing, only this time your living room floor is a computer screen and the lego blocks are millions of tokens. But it's that high that's clouding the sensibilities. The sensibilities on why, when software has never been "easier" to build, it still doesn't make sense to build your own. And in fact, even less sense than it did before.

This isn't a new phenomenon. The desire to build one's own software has existed for a while, and it can make sense. If what you're doing is genuinely unique, or you're in a niche industry, the software just simply might not exist and therefore you didn't have a choice. But when the technology does exist, and the use case has been built for, there is zero reason to choose to build it. The vendor, in todays world of AI-native software, is almost always better,.

Look, I have something of a vested interest here, i get it. I'm talking my own book. But I know this because I saw it first hand after spending a decade at JP Morgan. I watched the bank try to build its own version of the Bloomberg terminal, its own Slack, its own ChatGPT, its own deal-sourcing platform. Every one of them crashed and burned. The running joke was that Jamie Dimon would cite JPM's $20bn-plus annual tech spend and tell anyone who'd listen that the bank was really a tech company. The punchline was that every time we needed something useful, we'd end up buying the better third-party version anyway. If the bank with the largest tech budget on Wall Street can't build one piece of compelling software, it's a pretty good indicator you might not be able to either.

The reason the build trap is so seductive is that the early experience genuinely is incredible. You write a prompt and get a working dashboard. You ask a question you literally couldn't have asked twelve months ago. The cost-per-attempt has collapsed. For the first time, building software feels accessible in a way that it never was for non-engineers.

So you do more of it. But a few weeks in, the come-down starts. The dashboards that worked beautifully for one use case fall apart on the second. The agents you wired together start hallucinating in production. You realise you've built a tower of agents to do something one Google search alert would have handled just as gracefully and far more reliably.

The high makes you feel like you can build anything. It does not give you the judgment to know what's worth building.

The thing is, I do get it. Off-the-shelf software for most of the last 20+ years has just been terrible. Clunky, unintuitive, bloated, expensive "things" that, in their attempt to cater to everyone, end up serving no one. The vendors were charging too much for too little, the products didn't fit any specific firm's workflow, and if you were a senior partner two years into a frustrating procurement cycle, building yourself didn't sound crazy.

That world has changed faster than the conversation has. Software is actually getting good. Like really good. The genius of AI is that it can act as a translation layer. The need to understand how software works drops because the LLMs can help you work through it in natural language.

Software isn't just getting better. It's getting much, much more advanced. The good software today is solving much bigger, deeper problems: amongst those, converting and making your firm's data actually usable, as well as embedding actual workflows that are comparable to a human. This is effectively the whole premise DealSage was built on. Your firm's data lives across email, CRM, the data room, banker calls, portfolio reports, Excel models, and Sharepoint folders nobody has opened since 2019. None of it talks to anything else. There is no source of truth for which banker said what about which deal in March.

A model cannot reason over what it cannot see. Solving the data layer is a year of architectural work, not a weekend sprint. And the frontier labs themselves just confirmed this with their dollars: as I cover below, both Anthropic and OpenAI are now spending billions to acquire AI services firms because they've realised the binding constraint is no longer model capability. It's getting the model into the right place, with the right data, doing the right job.

The high is hiding the fact that the work that actually moves the needle is invisible from the dashboard you just built.

Before getting to the practical reasons not to build, the harder problem is conceptual. The teams building the best software right now have spent years thinking about what software even is in a world where LLMs can reason, where data can be unified, where workflows can run themselves. That isn't a question you answer over a few weekends. It's the product of thousands of conversations, dozens of dead ends, and a particular kind of taste that you only develop by being immersed in the problem daily.

The senior partner deciding to build is, almost by definition, not immersed in that problem. They're immersed in deal flow, LP relationships, portfolio governance. Brilliant operators, but pattern-matching against a frame of "software" that's a decade out of date. And the gap between that frame and where the market is actually heading is widening, not closing.

So let's number the reasons. Three are conceptual. Four are practical.

1. The idea of software has changed. I keep saying the tech stack is dead, and I genuinely mean it. The future isn't twenty different SaaS products stitched together with brittle integrations. It's one connected layer of structured information your LLM can reason over. Simple in concept, savagely hard in practice. If you're not building with that end state in mind, you're spending money to recreate what the rest of the market is already moving past.

2. Building your own CRM is rebuilding the wrong thing. This is the part most people miss. The teams currently building agentic software can see what the end state actually looks like, and it doesn't look like the categories we've been working with for twenty years. The whole frame of "we need a CRM, let's build a CRM" is stuck in what I'd call Work 1.0: software as a place for humans to click. Work 2.0 is information coming together, getting reasoned over, and getting presented to the right person at the right time, with humans there to direct and approve. AI is an amplifier. If you double down on a half-baked idea of what your work looks like, the amplifier just takes you further in the wrong direction, faster.

3. Software development itself has changed too. The same efficiency gains that made you think you could build this yourself apply, ten times over, to the people doing it day in, day out. They're constantly building, constantly updating, constantly iterating with feedback from thousands of users. AI amplifies up and down the value chain. The specialist teams using these tools to compound their lead are pulling away from the rest of the market every quarter, not converging with it. You're not catching them by hiring two engineers.

Park the conceptual problem for a second. Even if you knew exactly what to build, the practical reasons stack quickly.

4. Talent. Every very good agentic engineer in the world is currently being paid seven figures by a frontier lab. There's no shortage of developers. There's a brutal shortage of the ones who can build agentic systems that hold up in production. The funds I know who tried to hire ended up with mediocre engineers building a mediocre version of what already existed, because the market for the people who could actually do this well is closed to anyone not paying lab comp.

5. It ends up being a distraction from what you're supposed to be doing. Your day job is buying, selling, or raising capital for businesses. Building an AI platform is also a full-time job, with its own surface area: model selection, retrieval architecture, evaluations, security, compliance, all of it. The teams doing this well are doing it daily, on the receiving end of feedback from thousands of deals across dozens of clients. That feedback loop isn't replicable by an internal team of two engineers, however good. And every hour your partners spend in steering-committee meetings on the build is an hour not spent on the thing your LPs are actually paying you for.

6. The problem is already solved. The argument for building used to be that nobody understood your specific workflow. That's no longer true. Domain-specific platforms now exist, built by people who came out of the industry and partnered with people who came out of the labs. They have head starts measured in years.

7. Cost and maintenance never stop. The initial build is the cheap part. What follows is the expensive part: ongoing infrastructure, model upgrades every three months, retrieval pipelines that need re-tuning every time the data shape changes, security patches, compliance reviews. It ends up being incredibly expensive, requires constant maintenance and upheaval, and the moment your two engineers walk out the door, you're absolutely screwed. There's no team to fall back on. There's no documentation worth reading. The institutional knowledge leaves with them and you're back to square one, only now you've got a system in production that nobody understands.

8. Security - arguably the most important and yet most ignored currently. The fastest way to leak sensitive deal documents 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-priced risk in deal tech today.

Look at who Blackstone, Hellman & Friedman, Goldman Sachs and KKR partnered with to deploy AI across their portfolios. They partnered with Anthropic, OpenAI and Google. They didn't build.

These are firms with hundreds of engineers and balance sheets that can hire anyone alive. They felt the same high you're feeling. They had the resources to act on it. They looked at the question soberly and concluded that the right answer was to commit nine-figure cheques to bring the labs themselves in. If the firms with literally unlimited resources couldn't justify the in-house build, you can't either.

The real question is whether you can stay disciplined when the high tells you not to.

If you're thinking "we need a CRM, let's build a CRM," you're still solving for yesterday's workflow with today's high. The actual question is what your work looks like in an agentic world, and that's not a question an internal team of a couple of engineers can answer. It requires conversations across dozens of firms, hundreds of users, daily iteration. No amount of headcount fixes that.

The firms building the structured data layer in 2026 are creating the foundation everything else gets built on top of. The firms still debating it in 2027 will buy it from someone who started in 2024.

The high will pass either way. The question is what's actually working in your firm by the time it does.

News Digest

Anthropic's Financial Services Theatre

On Tuesday, Anthropic livestreamed a keynote announcing ten pre-built AI agents for financial services workflows: pitch decks, financial-statement review, compliance escalation, fraud detection, regulatory reporting. Microsoft and Moody's were on the partner list. Jamie Dimon was the marquee guest, with Fortune covering it the same day as Anthropic's biggest move yet to win Wall Street business.

The technology underneath is much smaller than the keynote suggested. The "agents" are Anthropic Skills, which is to say .md files. Plain-text instructions in a folder. The model and the tool-use loop have been generally available for over a year. What changed is that somebody wrote good instructions, packaged them up, and put them in front of a CNBC anchor.

That's not to downplay it. The skills are good. It's just to highlight that there's no new technology here. This could have been announced six months ago.

The details:

Why it matters: The lab playbook has clarified. Models are good enough; the gap to capture is implementation. Both labs are moving from selling capability to deploying it directly, and they're now buying the consultancies that do the work.

My take: Two things to read into this, neither of which is what the keynote was selling.

First, the labs have decided their models are good enough for now. Anthropic's next-generation model, Mythos, is reportedly being held back on safety grounds, which is the polite version of "capacity constraints and uncertain release timing." The firepower is going into financial-services skills and JV-led implementation services, not into shipping the next frontier capability. If you genuinely believed AGI was twelve months out, you'd ship the model and let it figure the rest out itself. They're doing the opposite.

Second, acquiring services firms is the loudest possible signal that the labs themselves don't think capability gains continue at the same pace. The bear case on AI services consultancies has always been that AGI would render them irrelevant in eighteen months. If Anthropic and OpenAI believed that bear case, they wouldn't be the ones doing the acquiring. Their dollars are telling you the model isn't the binding constraint any more. Implementation is.

The technology read is also worth discussing. The fact that the headline product is a folder of .md files matters, because it tells you what the actual moat is at this layer. It isn't the model. It isn't the prompts, which any sufficiently motivated firm can write or reverse-engineer in an afternoon. It's the data each agent sits on top of, and the workflow context the agent operates within.

Read more

The $100bn Hiding Between Your Systems

Bain published the latest instalment of its software-industry series on May 7, and put an institutional number on the agentic opportunity that the labs have been hinting at for months. $100bn of US enterprise spend, $200bn globally, currently sits in what they're calling "cross-system labour": the manual work people do bridging ERPs, CRMs, ticketing systems, billing, vendor management. Over 90% of it is uncaptured by current automation, and the rules-based RPA wave was never going to capture it because the work is too ambiguous and too context-dependent for anything other than agentic systems.

The breakdown by function is the bit worth printing out. Customer support and R&D sit at 40-60% automation potential. Finance and HR are at 35-45%, with pockets running much higher inside accounts payable and payroll. Sales is the largest single dollar slice at $20bn; COGS and operations contribute another $26bn. The aggregate is the most credible institutional sizing yet of where agentic AI actually lands first, and what it means for the SaaS businesses sitting on top of those workflows.

Bain's structural call is the bit PE should pay closest attention to: SaaS pricing is moving from seat-based to outcome-based. ARR per seat goes away as the metric, replaced by something closer to "share of customer outcome." Incumbent vendors who can't make that transition lose the multiple they've enjoyed for the last decade.

The details:

Why it matters: Bain is putting an institutional number on what the labs have been telling you in dollars (see story above). $100bn US, $200bn globally, in a single category, mostly uncaptured. The size is large enough to reset what enterprise software is worth, who deploys it, and how PE should be valuing both the existing portfolio and the new opportunities.

My take: The Bain analysis maps almost one-to-one onto how we've been describing the opportunity at DealSage for the last two years. "Cross-system labour" is exactly the problem. Your firm's information lives in twenty places, none of which talk to each other, and the actual work of dealmaking is the work of bridging those systems: reading the CIM, updating the CRM, pulling the call notes, building the model, drafting the memo, syncing back to the data room. That bridging is what gets automated. Not the thinking, the connective tissue.

Three reads underneath the headline number.

First, the $100bn isn't an expansion of enterprise software spend. It comes out of headcount. The dollars Bain is counting are currently being paid to people doing manual integration work between systems. That's a different P&L line being attacked, with different unit economics, and it's the line that matters most for PE software theses, because reducing the customer's adjacent labour cost is a value-creation lever inside every portfolio company you own.

Second, the pricing shift Bain calls out is the bigger story than the headline number. Seat-based pricing has anchored SaaS valuation for fifteen years. If outcome-based pricing genuinely takes over, every SaaS comp set needs new multiples and most existing PE software DCFs are wrong by a meaningful amount. The S&P data on dimming software exit multiples (see Quick Hits) is the early evidence of that recalibration starting.

Third, the AI-native entrants Bain flags are the interesting acquisition targets for the next 18 months. Cursor at $2bn ARR in 14 months is not a normal scaling curve. ServiceNow buying Moveworks is the template. Incumbent SaaS will roll up agentic capability via M&A at increasingly high prices, and the funds that recognise this earliest will rotate accordingly.

Read more

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

ECB studying defences against Mythos-powered attacks, Lagarde says (May 8) - The European Central Bank is now publicly thinking about how to defend payment infrastructure against AI-driven attacks. The list of things you want central banks to be ahead of, not behind, is short. This is one of them.

DOJ antitrust head warns dealmakers not to mislead AI (May 7) - The DOJ has clarified that misrepresenting facts to an antitrust review's AI tooling is, in fact, still misrepresentation. A new compliance line item we should probably all factor into deal-process timelines.

Wall Street Readies Data Centre IPOs as AI-Linked Debuts Surge (Bloomberg, May 6) - The IPO window is open, but only for AI-flavoured assets. Capital markets telling you, very clearly, where they think the next decade of returns sits.

S&P Global: Exit outlook dims for PE's long-held software investments (May 7) - Median software M&A multiples down materially in 2026, hold periods extending, AI uncertainty the cited driver. The cleanest data set yet on PE's largest single asset class starting to mark to reality. Pairs directly with the Bain story above; supply side meet demand side.

Blackstone-led group set to inject $100m into Medallia (Bloomberg, May 7) - $2.8bn of debt converted to equity to take control after Thoma Bravo refused fresh capital. Took private in 2021 at $6.4bn; equity now gone, lenders own the building. The first really visible PE software restructuring of the AI era. Unlikely to be the last.

Axios: AI creates a mess for private equity (May 6) - Plain-English summary of what every PE software portfolio is now quietly modelling in the background. Worth five minutes if only to confirm you're not the only one feeling it.

Carlyle Acquires Knack RCM and EqualizeRCM to Build an AI-Native Healthcare RCM Platform (May 4) - What an "AI-native" PE roll-up actually looks like in practice: two specialty RCM businesses combined into one platform purpose-built for the agentic era. Watch this template get repeated in every business-services subsector over the next 18 months.

Five Eyes agencies sound alarm over risky agentic AI deployments (IT Pro, May 8) - Five-country government warning on agentic AI security risks, urging restrictions on sensitive tasks and stronger governance before deployment. Validates the FTT point above on hand-rolled agentic systems being the largest under-priced risk in deal tech right now.

The FT's annual lament about junior bankers (FT) - The reliable spring-cleaning piece on how the latest analyst class isn't quite as sharp as the last. They write a version of this article every year. The catch is the cohort that genuinely grew up with the strongest version of these tools, the ones who hit campus when Claude was already useful, hasn't even graduated yet. Come back to me on this in 2028.

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