Connectors Or Foundations

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

… with TPG, Brookfield, Advent and Bain Capital as co-lead founding partners, and acquired Scottish AI consultancy Tomoro the same day to staff it. Anthropic struck a parallel $1.5bn vehicle with Blackstone, H&F and Goldman two weeks ago. Both labs are making the same bet: capability is no longer the constraint, deployment is.

PwC made the bet from the other side. The firm committed 30,000 staff exclusively to Claude this week and stood up a Claude-native finance practice inside the Office of the CFO. The case studies are impressive: insurance underwriting cycles down from ten weeks to ten days.

Elsewhere: Anthropic's Mythos sent JPM, Goldman, Citi, BAC and Morgan Stanley scrambling to patch hundreds of vulnerabilities. Mistral is racing to build a sovereign European alternative. Plus WSJ on "agent sprawl," Blankfein on why Goldman doesn't trust agents, and the FT on "AI brain fry."

But first, a crash course on the two kinds of AI software being sold under the same label. Confusing them is going to cost buyers real money over the next two years.

In This Week’s Issue:

From The Trenches

Connectors Or Foundations

A question we get on almost every call is some version of "how do you compare to X, Y or Z?" or "who are your competitors?" Our usual answer is that we don't really have any in the way the question expects, and it lands awkwardly. People raise an eyebrow. It sounds like the typical nonsense a founder says when they want to dodge a comparison

But we genuinely don't compete with most of them on the same axis. There are two fundamentally different kinds of AI software in the market, sold under the same label, and confusing them is going to cost buyers real money over the next two years.

Call them connectors and foundations.

The connector is the dominant pattern in the market today. You take a frontier model (Claude, GPT, Gemini or some combination of all of them), put a "for finance" label on it and stick it in the middle of your stack, and have it connect via MCP out to your existing systems: CRM, data warehouse, email, SharePoint etc. The model reaches into each system, pulls what it needs, and synthesises an answer. A connector to your various sources of data, with a chat box on top.

This is what almost every "AI for finance" platform pitched in the last six months is doing under the hood. It's the cheap, fast, vendor-friendly way to claim you have an "AI strategy," and on a one-hour demo it can look kind of impressive.

Or at least it did until Claude came along and effectively became the same thing. In 2026 you cannot be a chat-based experience sat on top of your data. That seat is taken. You have to be doing something fundamentally different. It also doesn't work for a number of reasons:

When the model in the middle opens an MCP connection to your CRM, it doesn't actually know what your CRM is. It gets back a slice of fields and has to guess what they mean. Then it opens the data warehouse: different schema, different naming, different ideas of what "revenue" or "customer" or "deal" even means. The your documents, emails and call notes.

Each round trip is the model reconstructing context from scratch, across five separately-modelled worlds that don't agree with each other. There is no coherent map of your firm anywhere in the loop, and the connective tissue holding it together (the MCP integrations) breaks every time an upstream system changes a schema, renames a field, or shifts permissions. Each of those five systems is owned by someone different and evolves at a different cadence. The model is guessing, the guesses get worse the more systems it has to stitch together, and the brittleness is baked into the shape of the architecture rather than something that improves over time.

Worth being honest about what the connectors themselves are, too. MCP is a public protocol. "We connect to Salesforce, Slack, NetSuite and Box" is not a moat: any vendor can connect to the same systems, and increasingly the model providers themselves are shipping native connectors out of the box. There's no such thing as an exclusive connection. The connectors are commoditising in real time.

A single Excel financial model can be a million tokens on its own. Claude's largest context window today is one million tokens. So the model is either pulling one file at a time and losing the cross-file reasoning, or it's sending half a deal room through on every query.

Pieces fall out. The answer that came back clean in the demo gets noisy and inconsistent the moment you run it across real data. Anyone who has spent meaningful time prompting agents across enterprise data knows this is the dominant failure mode in production. The model is doing its best with too little, in a context window that's already too small for the task.

Marc Benioff said on the All-In podcast this week that Salesforce is on pace to spend $300m on Anthropic tokens this year, and called for an "intermediary layer" that routes simple tasks to small models and reserves Claude for the heavy reasoning. He's right about the need. Worth being explicit about what that means though, because the easy misreading is that this is about giving the user a dropdown to pick which model they want.

It isn't. A dropdown in the UI is not routing. Real routing is a platform layer that dynamically and proactively selects the right model for each step of each task, in the background, without the user thinking about it.

And that's only half the problem. Even with perfect model routing, you still have to actively manage what you send to the model and what you ask it to send back. Every input token costs money. Every output token costs more. The connector architecture's instinct is to throw the whole deal room into context "just in case" and let the model figure it out, which is exactly the behaviour that turns a manageable token budget into a 10x overrun by Q3. The platforms that hold their economics are the ones doing both jobs in the background: picking the right model for the step, and being ruthless about what context that model actually needs to see, and what shape of answer it should give back. Tokens equal dollars, and the architecture that ignores that math is the architecture that loses on it.

The picture beyond Salesforce is just as stark. Axios reported that Uber's CTO has already burned through his entire 2026 AI budget on tokens. Nvidia's chief scientist Bryan Catanzaro: "For my team, the cost of compute is far beyond the costs of the employees." That's the company selling the compute saying their own compute bill outweighs their payroll.

The fairest objection to all five of those points, and the one I hear most, is that the model will keep improving and the connector architecture will eventually catch up. Context windows will get bigger, agents will get smarter, token costs will come down. Sit tight and wait.

Three problems with that.

One, the frontier has visibly plateaued. Opus 4.5 to 4.6 to 4.7 has been a thin step. Mythos is held back. The labs are signalling, with their dollars, that the bottleneck is no longer model capability. Spending billions on consulting JVs and acquiring services firms is what you do when you've decided the model is good enough and the constraint is everything around it.

Two, even if a real step-change in capability did land tomorrow, the connector architecture would need exponentially better models to overcome the structural problem. Reading slices across five disjoint systems and reconstructing a coherent answer is not 2x harder than reading from a single connected source. It's orders of magnitude harder, and a model that's a bit better doesn't get you there.

Three, even with much better models, the cost gets worse, not better. More capable models cost more per token. Bigger context windows mean more tokens per query. The arithmetic moves against you, not with you.

The other kind of AI software is what we've been calling foundations or infrastructre. The information is consolidated first, structured once, and the model is placed inside the same architecture with first-class access to a coherent map of your firm: the CRM, the documents, the financial data, the call notes, the deal pipeline, the portfolio history, all on the same connected layer.

The model stops being a frontier API with MCP plugins, and starts being part of an architecture that actually understands what you do.

It's harder to build. It's slower to deploy. The migration is the whole job. But it's the only architecture where the model actually understands your firm, the cost stays bounded, and you can swap the underlying model the moment something cheaper, faster or better lands.

This is where the market is moving. Graphon AI emerged from stealth this week with $8.3m of seed funding and a team out of Amazon, Meta, MIT and NVIDIA, pitching a "pre-model intelligence layer" aimed squarely at this problem. The architectural question is where the capital and the talent are going.

If you're evaluating AI software right now, two questions cut through almost everything.

First: does the model have a single coherent view of my firm's information, or is it stitching together five different views every time I ask it something?

Second: if I want to change the underlying model in six months, can the platform do it without rebuilding everything?

Both answers should be a straightforward "yes." If either is "well, sort of," you're looking at a connector. A connector might still be the right tool for the next twelve months, particularly if you don't yet know where you want to land. Just don't confuse it with the destination.

Foundations are what gets you anywhere near actually achieving the step-change function that AI has been promising.

News Digest

OpenAI's $4bn Deployment Company, And The Real Trade Underneath It

OpenAI announced on Monday that it is setting up a new majority-owned company, the OpenAI Deployment Company, with more than $4bn of initial investment to embed engineers inside corporate clients and drive AI deployment at scale. The capital comes from 19 firms, led by TPG with Advent, Bain Capital and Brookfield as co-lead founding partners, joined by Goldman Sachs, SoftBank, Warburg Pincus, B Capital, BBVA, Emergence, Goanna and Welsh Carson, among others. Brookfield alone wrote a $500m cheque.

The same day, OpenAI announced it is acquiring Tomoro, a Scotland-based applied-AI consultancy founded in 2023, to add 150 forward-deployed engineers to the new unit on day one. Tomoro's existing client roster includes Mattel, Red Bull, Tesco, Virgin Atlantic and Supercell.

Two days later, TPG partner David Trujillo went on Bloomberg TV and framed the venture as "not very dissimilar from what we normally do" with Intel, Humana, UnitedHealth and AT&T. The honest line from the interview was the one to print out: "All of our CEOs are clamouring to deploy these tools" but "there's a lot of experimenting going on, things that are not working, things that are. We are in that very early stage."

The details:

Why it matters: Both frontier labs have publicly committed multi-billion-dollar capital to PE-backed services arms in the space of a fortnight. Capability is no longer the constraint, deployment is, and the labs want the implementation economics directly rather than ceding them to the Big 4.

My take: Two things to take from this. First, the labs no longer think the model is the differentiator. If they did, they wouldn't be acquiring services firms and standing up consulting JVs with the largest PE houses in the world. The bottleneck has moved.

Second, the lock-in nobody at these signings is talking about loudly enough. Trujillo was careful to say "no exclusivity" on Bloomberg, but once 150 OpenAI engineers are embedded inside your CRM, ERP and reporting stack architecting workflows around GPT, you don't casually swap them for Claude or Gemini twelve months later. You're getting a real head start on deployment in exchange for a long-dated supplier-concentration risk. For most funds in these JVs, that's still the right trade today. Mistral's Arthur Mensch already put the European version of the warning on the record this week at the French National Assembly: "an irreparable dependency." Read it as the line every PE board will be quoting in 18 months.

Read more

PwC Commits 30,000 Staff To Claude

On Wednesday, PwC and Anthropic announced a major expansion of their alliance: 30,000 PwC professionals to be Claude-certified, a joint Center of Excellence, Claude Code and Cowork rolling out US-first then globally across hundreds of thousands of staff, and a new Claude-native finance business group inside PwC's Office of the CFO practice. Three explicit focus areas: agentic technology build (engineers shipping production software in weeks rather than quarters), reinventing deal execution end-to-end with agents working alongside teams, and AI-native operating models for finance, supply chain and HR.

The case studies they ran with the press release are not fluff. Insurance underwriting cycles down from ten weeks to ten days, opening lines of business that weren't economically viable before. Up to 70% delivery improvement reported across production agentic builds.

The details:

Why it matters: PwC is the largest, most organised distribution channel any frontier lab has secured to date. The Big 4 playbook for the next 18 months has been published.

My take: The interesting question isn't whether PwC can deliver. It's what this does to the Big 4 pecking order. Anthropic has handed PwC a 12-18 month lead on industrialised Claude-trained delivery capacity. Accenture, Deloitte and EY now have to choose: match the exclusivity with a competing lab and shoulder the same lock-in risk, or stay multi-vendor and watch PwC out-deliver them on price and speed. Neither option is comfortable. The Big 4 spent two decades positioning themselves as model-agnostic systems integrators. That position has just been taken off the table.

Read more

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

Companies have a new AI problem: too many agents (WSJ CIO Journal, May 15) - DaVita employees have spun up over 10,000 agents. Lyft and GitLab are now scrambling for governance around what the WSJ is calling "agent sprawl." A direct consequence of Claude Cowork making it trivial for non-technical staff to spin up bots with no central data foundation to organise what they all touch. You can already see who's going to be paying consultants to clean it up in 2027.

Lloyd Blankfein on why Goldman doesn't trust AI agents (Fortune, May 13) - Goldman has 46,000+ employees on AI assistants but has drawn a hard line: autonomous execution above certain thresholds still needs a human signature. Blankfein's frame is the one worth quoting back at any vendor pitching full autonomy: "We don't have the ability to test whether it's right or not." The frame to bring to your next IC discussion on autonomy thresholds.

Anthropic's Mythos sends US banks rushing to plug cyber holes (Reuters, May 12) - JPM named launch partner; Goldman, Citi, BAC and Morgan Stanley also have access. The tool is finding low-to-moderate vulnerabilities in the hundreds-to-thousands and chaining them into high-risk exposures the banks didn't know they had. Patch cycles compressed from weeks to days, $100m of Anthropic credits at launch. The less good news: nobody outside the top 5 banks has the tool finding them.

Mistral building its own Mythos alternative for European banks (Bloomberg, May 13) - HSBC and BNP Paribas already clients. CEO Arthur Mensch at the French National Assembly: scanning French military source code with Mythos would create "an irreparable dependency." Sovereignty has become a procurement question for European financial institutions, not a policy one.

Anthropic publishes "2028: Two Scenarios for Global AI Leadership" (May 13) - The policy team modelling geopolitical scenarios for AI compute, energy and talent over the next 30 months, with an explicit ask for tighter US export controls on Chinese chips. Reads less like research and more like a position paper, dressed up as one of the other thing.

"AI brain fry" hits white-collar workers, BCG research finds (FT, May 13) - BCG surveyed 1,488 workers and found a small epidemic of "cognitive overload" from juggling AI outputs, software engineering and marketing hit hardest. The phrase to remember is "vibe coding paralysis": one founder's description of "six worktrees open, four half-written features, two quick fixes that spawned rabbit holes, and a growing sense that I'm losing the plot entirely." Last week's Build Trap FTT, now dressed up in BCG diagrams.

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