Enterprise AI lock-in didn’t disappear when models became interchangeable — it quietly relocated to the floor above, and the biggest AI vendors just spent billions of dollars to prove it. When OpenAI spins up a consulting arm backed by more than $4 billion and PwC commits to training 30,000 people on a single vendor’s model, that’s not a side hustle. That’s the tell. The model itself is becoming the commodity. The workflow, the governance, the human glue around it — that’s the new moat.
Why $4 Billion in Consultants Is the Loudest Signal in AI Right Now
PwC announced it will train and certify 30,000 staff on Anthropic’s Claude and build an Office of the CFO practice for banking, insurance, and healthcare. Anthropic committed $100 million to a partner network. OpenAI stood up the OpenAI Deployment Company — DeployCo — backed by more than $4 billion in initial investment to send forward-deployed engineers on-site and embed GPT models into customer workflows.
For companies that supposedly make their money selling tokens by the million, pouring capital into low-margin professional services looks strange. It isn’t strange at all. It’s an admission. The model call is getting easier to replace; the surrounding workflow machinery is not. Vendors are racing to own that machinery before someone else does.
If you’re a CIO at a regional insurer evaluating Claude versus GPT for claims triage, this changes the math. The cheaper model on a benchmark sheet might be the more expensive choice once a consulting firm has spent six months wiring the other one into your underwriting flow. Our prediction: within 18 months, “who staffs your deployment” will outrank “which model scored higher” on most enterprise AI RFPs.
Why the 95% Pilot Failure Rate Built DeployCo
MIT’s NANDA initiative reported that 95% of enterprise generative AI pilots fail to deliver measurable business impact. The methodology has been contested, but even the more generous counter-readings still put the gap between AI investment and AI value capture in painful territory. As the original analysis from Sanchit Vir Gogia of Greyhound Research notes, most failures aren’t about model capability. They’re about operational fit. The tools don’t learn the workflow, don’t sit inside the approval path, and don’t carry the right permissions.
That single statistic is the reason DeployCo exists. OpenAI didn’t copy Palantir’s playbook because it ran out of ideas. It copied it because three years of stalled pilots taught the company that customers weren’t asking for a smarter model — they wanted someone to show up on-site and do the expensive, hard-to-replace work of mapping the model into how work actually gets done.
Imagine you’re a mid-sized bank’s head of operations. You’ve already paid for two model subscriptions and watched both pilots stall at the same wall: nobody inside the bank can describe the underwriter’s decision tree well enough to encode it. That wall isn’t a model problem. It’s a process problem, and it’s why services revenue is suddenly strategic for vendors that, on paper, sell software. The practical implication for buyers thinking through AI agents vs broader AI automation is that the choice is less about the underlying tech and more about which partner will physically sit next to your team and codify how decisions get made.
Why Model Context Protocol Was Never Going to Save You
Model Context Protocol is useful. It collapses the cost of connecting models to tools and data, and anyone who has maintained half a dozen bespoke connectors to ServiceNow, Salesforce, or Jira knows the value. But a protocol is not a platform. MCP can help an agent talk to a tool. It cannot, on its own, tell an enterprise who approved that agent, what data it can touch, how its actions are logged, or how to shut it down safely when the person who launched it has left the company.
Kubernetes is the right analogy here. Standardizing the container layer didn’t eliminate cloud lock-in. It pushed the fight one floor up, into managed services, identity, networking, observability, and data gravity. MCP is doing the same thing for AI agents — making one floor of the building portable while leaving the harder enterprise problems exactly where they were. Lower integration cost, same operational trust cost.
If you’re a fintech building a compliance-heavy product, this matters in concrete ways. An AI-integrated software stack is not finished when the model can call your APIs. It’s finished when the approval chain, audit trail, identity bindings, and rollback paths are designed for the way a regulator will eventually read them. Our take: by 2027, “MCP-compatible” will be table stakes the way “REST API” became table stakes — necessary, unremarkable, and not a differentiator anyone will pay extra for.
Where the Real Lock-In Battle Is Being Fought
Three battlegrounds are emerging, and none of them are the model itself.
The orchestration layer is the first. LangChain counts Klarna, Replit, Elastic, and Ally among LangGraph’s production users. A team that has spent a year encoding agent behavior, evals, recovery logic, and observability traces inside one orchestration framework is not ripping it out because a competitor ships a faster model next quarter. Orchestration accumulates stickiness whether anyone planned for it or not.
The second is the vendor-controlled workflow surface. Anthropic’s Claude Cowork February 2026 expansion shipped private plug-in marketplaces, per-user provisioning, and prebuilt agents for HR, finance, investment banking, and design. No enterprise IT leader wants 400 random agents bolted onto contract systems, HR data, and customer records. So the administrative surface around the agents — not the agents themselves — becomes the product.
The third is the services layer, and this is where the irony is sharpest. PwC and Anthropic claim their joint work has moved cybersecurity incident response from hours to minutes and underwriting cycles from 10 weeks to 10 days. Those gains, per the partnership’s own framing, don’t come from the model. They come from tens of thousands of trained consultants who know how to redesign the surrounding process. Any vendor that wants to displace those workflows has to retrain that army. Good luck.
For teams building or buying around this stack, the same logic applies one level down — custom API and integration work is exactly where the operational trust gets built, and exactly where switching costs quietly compound.
What Enterprise Buyers Should Actually Be Deciding
Stop fixating on the next point solution and look one or two layers up. Three decisions matter more than any model bake-off:
Which orchestration framework will your engineering team commit code to? Which workflow surface will your end users actually live inside? Which services partner will be embedded deeply enough in your operations that their model recommendation is effectively binding?
Model substitution at the API layer is getting cheaper. The orchestration commitment is a multi-year code rewrite. The workflow surface is a behavior change across thousands of employees. The services relationship is a long-tail budget line item. Those decisions persist. The model choice does not.
Anthropic’s open sourcing of Agent Skills, with the explicit promise that “skills you create aren’t locked to Claude,” is the right hedge from the customer side. Keeping a second frontier model in active use is another. But the deeper move is to treat workflow integration as the thing you actually own — and the model and the partner as the substitutable layers around it.
FAQ
Q: What is AI lock-in in the enterprise context? A: Traditionally, lock-in meant being tied to a single model vendor’s API. Today it has moved upward — into orchestration frameworks, vendor-controlled workflow surfaces, and the consulting relationships that wire those tools into your business processes. Swapping models is increasingly easy; swapping the operating model around them is not.
Q: Does Model Context Protocol eliminate vendor lock-in? A: MCP reduces integration cost between agents and tools, but it doesn’t address governance, identity, approval flows, or audit trails. It standardizes one floor of the stack while the harder enterprise problems remain one floor up — similar to how Kubernetes standardized containers without eliminating cloud lock-in.
Q: Should enterprises commit to a single AI vendor’s ecosystem? A: The pragmatic answer is to commit deeply at the workflow and orchestration layer while keeping the model and services partner more loosely coupled. Anthropic’s open-sourced Agent Skills and the option to run a second frontier model in parallel are reasonable hedges against vendor concentration.
Key Takeaways
- Evaluate AI vendors on who staffs your deployment, not just which model scores higher on a benchmark — services depth is becoming the real moat.
- Treat your orchestration framework choice as a multi-year code commitment, on par with picking a database, not a tool you can swap each quarter.
- Audit which workflow surfaces your end users live inside today; whichever vendor owns that surface will quietly own your roadmap.
- Demand portability guarantees — open agent skills, exportable workflows, second-model compatibility — and price them into every enterprise contract.
- Build internal capability to map and codify your own workflows; teams that own that knowledge will keep model commoditization working in their favor instead of against it.