MIT's widely cited 2025 research on enterprise AI — based on interviews with dozens of executives and analysis of hundreds of deployments — found that roughly 95% of generative AI pilots produced no measurable impact on the P&L, despite an estimated $30–40 billion in enterprise investment. That number gets quoted often. What gets skipped is the part of the same research that explains it: of the 25 factors MIT tested for their effect on financial return, workflow redesign had the largest impact of all. Companies weren't failing because the models were weak. They were failing because the work around the model never changed shape.

McKinsey's most recent State of AI survey puts a figure on how rare real redesign still is: only about 21% of organizations have redesigned even a single workflow around AI. Deloitte's enterprise AI research tells a similar story from the operating-model side — nearly half of organizations introduced AI without redesigning workflows or roles at all, and only 12% report redesign at scale with a genuinely new operating model behind it. Everyone else added a tool to an unchanged process and called it transformation.

Bolting on vs. redesigning

Bolting on looks like progress because something visibly changes: a new interface, a new button, a demo that gets applause. But if the underlying steps, approvals, and handoffs stay exactly as they were, the AI is just a faster way to produce the same inputs the old process already required. Someone still re-keys the output into the system of record. Someone still reviews every single item instead of a sampled few. The bottleneck that made the process slow in the first place never moves.

Redesign starts from a different question. Not "where can we drop AI into this process," but "if we were building this process today, knowing what AI can do, what would we remove?" That's a much more uncomfortable question, because the honest answer is usually a step, an approval layer, or a role — not a tool.

Where redesign actually pays off

MIT's research also found an uncomfortable mismatch in where the investment was going versus where the return showed up. More than half of enterprise AI budgets in 2025 went to sales and marketing pilots — the most visible, most demo-friendly category, and the one with the weakest measurable returns. The stronger returns came from less glamorous back-office work: invoice processing, contract administration, data reconciliation, compliance reporting, client onboarding. Those functions had something the flashier pilots usually lacked — well-defined steps, existing volume, and a workflow simple enough to actually redesign end to end rather than merely accessorize.

That pattern holds in our own client work. The rollouts that produce a number worth reporting are the ones where someone was willing to touch the process, not just the tool:

  • Remove a step instead of adding a review of it. If AI can draft a document reliably, the redesign question is whether every draft still needs the same three approvals — not how to route the AI's output through the old approval chain.
  • Change who does the work, not just how fast they do it. If a specialist's job was mostly first-pass triage, and AI now does triage, the specialist's role has to be redefined toward exceptions and judgment calls — or the head count savings never materialize.
  • Redraw the handoffs. Most workflow time isn't spent doing the task; it's spent waiting between people. AI that removes a task but leaves the same handoff queue behind saves less than it looks like on paper.
  • Retire the manual fallback. If staff can quietly revert to the old spreadsheet whenever the new tool feels unfamiliar, the old workflow is still alive, and it will absorb all the friction the new one was supposed to remove.
The question isn't where AI can help the current process. It's what the process would look like if you designed it today, knowing what AI can do — and what you'd be brave enough to cut.

Redesign is a change-management problem before it's a technical one

This is exactly why workflow redesign stalls at 21% instead of becoming standard practice: removing a step or a role is a harder conversation than adding a tool. It touches headcount, territory, and the comfort of a process people already know how to do. No one objects to a new dashboard. People absolutely object to having their job description rewritten.

That's the real reason most AI initiatives stay in "bolt-on" mode. It's not that leaders don't understand the technology — it's that redesigning a workflow requires the same sponsorship, sequencing, and stakeholder work as any other significant change, and most AI rollouts get treated as a software install instead. The fix isn't a better model. It's naming an owner for the redesigned process, retiring the fallback path deliberately, and measuring the same baseline metric before and after — the same discipline that makes any other change stick, applied to the process AI was supposed to improve in the first place.