For two years, enterprise AI strategy has run through one question: which model? GPT, Claude, or Gemini. Executives sat through the bakeoff. Somebody built a scoring matrix.
That matrix was the easiest decision in the program, wearing the costume of the hardest one. Frontier models converged faster than anyone’s roadmap allowed for. They reason, write, and summarize close enough to each other that the gap stopped deciding outcomes, and swapping one for another next quarter is a config change. The decisions that determine whether any of this survives contact with production have moved somewhere else. Most organizations have not followed them there.
Choosing a model is the easy part. Connecting AI to the systems, data, and workflows that run the business is the hard part, and it decides which initiatives scale. Orchestration frameworks and Model Context Protocol, the standard governing how models reach tools and data, set the terms for that work now.
Agentic AI gets sold as an autonomy story. The software does the job, headcount comes down, the line goes up. Companies that acted on that framing early got burned in public. The version that survives production is duller, and it is about integration and governance.
Model commoditization changes the definition of success
Every frontier model reasons, writes, and summarizes at roughly the same level. Nobody is winning on model choice. The companies pulling ahead have better wiring: data access, permissions, workflow integration, and some way to tell whether an agent is doing its job.
That shift breaks organizations that run AI through procurement. Buy model access, buy Copilot seats, wait for transformation. You get one of two outcomes. A demo that never reaches production, or a production system that underdelivers because nobody built the context layer and the feedback loop underneath it.
The real work lives in enterprise integration
Getting an agent to your data, with the right permissions, wired into a workflow, does not photograph well. It is where the effort and the risk live.
McKinsey surveyed 1,993 organizations across 105 countries for its November 2025 State of AI report. Eighty-eight percent use AI in at least one function. Thirty-nine percent report any enterprise-level EBIT impact, and most of those put it under 5%. About 6% clear McKinsey’s bar for high performers.
One finding explains the rest. Of every organizational change, McKinsey correlated with EBIT impact, and fundamental workflow redesign ranked highest. Only 21% of organizations using gen AI have redesigned any workflow at all.
Four out of five bought the tool and kept the process. Then they went looking for the return.
MCP shows the same pattern in a smaller frame. I build MCP servers regularly, and from a conference stage, the question I field more than any other arrives in the same shape: should we build one for X. It comes in miscalibrated often enough that I built a calculator to answer it. Seven dimensions, four possible verdicts, and two of the four are versions of don’t. Run it as a no-code workflow. Or skip the protocol and call the API directly.
That is not a knock on the protocol. MCP is an open, universal port and it does real work. It gets misread two ways.
The first: vendors announce MCP support and buyers hear agent-ready. A USB port on a laptop is not a reason to trust what you plug into it.
The second is worse. Every MCP server you stand up is another door into your data and another attack surface. Most companies are wiring servers up faster than security can inventory them.
AI orchestration is the new competitive edge
Single-task agents are a demo. Once agents chain steps across systems, orchestration decides whether the thing survives a Tuesday when an API returns a 503.
Orchestration governs how an agent hands work between tools and where errors get caught before they compound. Miss it and failures do not announce themselves. They propagate until somebody notices the numbers are wrong, usually a quarter later.
Value in the software stack is migrating to that layer for a second reason. If your product is a nicer interface over a database and a workflow, an agent can run the workflow directly and your interface becomes optional. Systems of record, proprietary data, and regulated workflows hold. Thin single-workflow tools do not.
AI governance can’t be an afterthought
An assistant suggests. An agent acts. Cross that line and the risk picture changes, because you have created an identity that needs credentials, a scope, and somebody accountable for what it does at 3 am.
Non-human identities are now doing real work inside systems that were never built to manage them, let alone track what each one can reach. Ask your team for a list of every agent in production and what each one has access to. The pause you get back is the finding. Over-scoped agents are the exposure, and nobody is counting them.
Governance has to move at the speed of what agents are allowed to do:
- A record of what each agent did and why, readable by someone who was not in the room
- Monitoring, because an unsupervised agent fails quietly for a long time before anyone notices
- A human checkpoint before anything expensive or irreversible
The journey from pilot to production
A pilot is rigged. Clean curated data, one use case, a champion who wants it to work, no compliance review. Production strips all four at once.
The model is not what breaks. These are.
- Your data. Real company data sits scattered across disconnected systems, much of it behind access controls that require specific approval to cross. The pilot ran on a clean extract.
- Your integrations. Connecting systems that were never built to talk to each other turns out to be most of the work, and it was not in the estimate.
- Your adoption. Handing someone a tool does not change how they work. Somebody has to redesign the day around it and earn enough trust that people rely on it. This is the line that gets cut when the timeline slips.
- Your ownership. Innovation ran the pilot. Nobody signed up to run it at 8 am every Tuesday for the next four years.
McKinsey’s scaling number is worth sitting with. About one-third of organizations have begun to scale AI across the enterprise. Seven percent call it fully scaled.
What leaders need to know
Two blind spots turn up in almost every AI budget.
The first is cost shape. Leaders fund the model and the pilot. The model is a rounding error. The money goes to integration, workflow redesign, governance, and evaluation, and none of it stops at launch.
The second one kills working systems. Leaders want returns to show up as headcount. Headcount is the slowest and most political number available and the last one to move. Value arrives first as cycle time, cognitive load, and capability that the business did not have last year. Measure only for cost reduction, ignore everything else, and you will shut down something that was working.
The answer is a value receipt: a running record of what the system did, what it would have taken a person to do, and what that is worth. Not a proxy metric. Not a satisfaction survey. The work is priced.
The next phase of enterprise AI runs on infrastructure nobody puts in a keynote. Governance and integration are built so AI can reach business systems safely and reliably. Model selection stays easy. It stays easy because it stopped mattering.
