Who survives the disintermediation — and why most won't
In the same week, PwC, Deloitte, EY, and KPMG all announced major AI partnerships. PwC with OpenAI. Deloitte with Nvidia, while quietly rolling Claude out to 470,000 employees on the side. EY with Nvidia and Microsoft, with 80,000 tax professionals already on 150 internal agents and a thousand more in build. KPMG all-in on Microsoft for Workbench, with their Clara audit platform already serving 95,000 auditors.
If you read these as four independent press releases, you miss the picture. Read them as a single announcement.
The Big Four are wiring themselves directly into the AI infrastructure layer.
Notice what isn't in the picture.
The software vendor.
For thirty years, enterprise IT had a stable value chain. At the bottom: infrastructure — compute, networking, databases. Above that: independent software vendors building applications on the infrastructure. Above them: systems integrators bending the software to fit a specific customer. At the top: the customer running their business on the result.
Each layer did something the layers below it couldn't. Infrastructure delivered raw capability. ISVs encoded reusable domain knowledge. Integrators shaped the product to the customer. The customer drove value out of the configured result.
The agentic shift is collapsing two of those layers at once. From above, and from below.
The view from above: the AI client eats the workflow
I've written about this side already. The argument is short: as the AI client becomes the universal access point — the new browser — applications stop being destinations and start being plug-ins. The user no longer cares which ERP rendered the invoice. They ask the agent, and the agent uses whatever tool gets the job done.
That moves the locus of control from the application to the agent. And it dissolves several of the old moats.
UX polish stops mattering when no human looks at the screen. The muscle memory of trained users stops mattering when their loyalty migrates to their assistant. Even integration depth stops being defensible — every MCP server is, by definition, a clean specification, and clean specifications are the easiest thing in the world for an agent to reimplement.
This is the first cut. It has been the focus of most of the agentic-software discourse so far.
The view from below: consulting plus infrastructure, direct
The second cut is what the Big Four partnerships make visible.
For decades, consulting firms were ISVs' best friends. They evangelized the software, sold the implementations, ran the change management, and lived off the multiplier between license cost and consulting hours. SAP, Oracle, the Microsoft Dynamics ecosystem — large parts of these businesses were built on the back of partners who needed those products to have something to implement.
The agentic stack changes the economics underneath that arrangement.
A Big Four partner now has three things it didn't have a decade ago:
- Direct relationships with model and infrastructure providers. OpenAI, Anthropic, Nvidia, Microsoft will all build with them, sometimes for them, and increasingly as them.
- Deep domain knowledge already encoded as practice. Tax, audit, controllership, treasury, FP&A — these are not new domains the firms need to learn. The partners already know how the work gets done, and they have decades of structured precedent to feed an agent.
- The customer relationship. Every CFO already has a Big Four firm on speed dial. The friction of selling a new system is dramatically lower when it comes from the firm doing the audit.
Put those three together, and a Big Four firm doesn't need an ISV to deliver finance software anymore. They need an LLM, an infrastructure provider, and the institutional knowledge they already have.
The bypass route is now open.
What PwC is quietly doing today — selling AI-enabled finance services to their own clients, built on OpenAI's stack, with OpenAI's own finance team as the test bed — is the leading indicator. The lagging indicator will be the first Big Four firm that decides not to sell "services" but a productized Finance-as-a-Service offering that replaces the ERP outright.
Graduate hiring at the Big Four fell 44% year-on-year. That isn't a story about junior accountants. That is the firms telling the market what they now believe they can deliver without that headcount.
The pincer
So the ISV layer is being squeezed from two directions:
- From above, the AI client absorbs the workflow. The user no longer talks to the application. They talk to the agent, which talks to whichever application is most convenient to call.
- From below, the consulting partner absorbs the implementation. The customer no longer buys the ISV's product and pays the consultant to install it. They buy a service from the consultant, who quietly built it on top of the infrastructure provider.
The middle, where the ISV lived, was held in place by three things: UX, integrations, and the inertia of trained users.
- The agent dissolves the first.
- MCP standardization dissolves the second.
- The consulting bypass dissolves the third.
Most ISVs do not survive that.
What does survive
What survives the pincer is narrower than the industry wants to admit. As far as I can see, it's three things:
- Data the agent cannot reconstruct. Proprietary, customer-specific, accumulated over years. If your only data is what the customer types into a form, you don't have this. If your data is the network effect of a hundred thousand companies running their close through your platform, you might.
- Judgment encoded across thousands of customer-specific decisions. Not "best practice" content, which any agent can synthesize. The kind of judgment that comes from being deeply embedded in a single domain for a long time, and from making bets that customers rely on without quite knowing why.
- Memory that compounds. Not chat history. Per-tenant context that gets sharper with use — the system's understanding of how a specific customer thinks, structures its books, defines its KPIs, runs its calendar. This is the moat that grows.
Everything else is a wrapper waiting to be rewritten — by an agent above, or by a Big Four partner with an OpenAI relationship below.
The honest part
I should be transparent about where I sit in this picture. I am building a software product. Specifically, an FP&A application. The layer I am building in is exactly the one being squeezed.
I am not building it because I think ISVs are safe. I am building it because I think there is a narrow slice of the ISV layer that survives — the slice that does the three things above — and because I think that slice is much smaller, much more valuable, and much harder to clone than the broader category that came before it.
The companies that get this right will not look like the SaaS companies of the 2010s. They will look more like specialized data and judgment companies that happen to expose an AI-callable surface. Their value will not be in their UI. It will be in what they remember, what they know, and what they decide.
The companies that get this wrong will spend the next three years racing to expose an MCP server, and then discover either that the agent built a better one over a weekend, or that the Big Four partner built the same thing in a billable engagement.
The question
A few weeks ago I argued that software was becoming a plug-in to AI. That was only the top half of the picture. The bottom half is the consulting + infrastructure shortcut around the ISV layer entirely.
The question for software founders in 2026 isn't whether the agents will eat us.
It's who gets there first — the agent above, or the consultant below — and whether what you've built is the kind of thing they need to keep, or the kind of thing they can finally route around.
With thanks to Håkan Ebersjö for the prompt, and to Rob Sinfield (Sage) whose breakdown of the Big Four partnerships shaped the data here.