The Reality Check

The Reality Check

The direction is clear. The results so far are not.

The numbers

• MIT's NANDA initiative reported in August 2025 that around 95% of enterprise generative AI pilots it studied had produced no measurable impact on profit and loss.
• Gartner predicted in June 2025 that more than 40% of agentic AI projects will be canceled by the end of 2027, because of cost, unclear value or weak risk controls — and warned about "agent washing", vendors relabeling old automation as agents.
• Industry surveys through 2025 and 2026 consistently report that most enterprises overran their AI budgets and that few can forecast AI spend with any accuracy.
• Agents are spreading faster than the ability to see and control them. Most large enterprises now run agents; a small minority manage them centrally.

The rhetoric is running ahead of the operating reality.

Why it fails

When you look at the programs that do produce a return, they tend to share four things: unified data, narrow scope, real governance and a human escalation path. None of those is a model capability. All of them are organizational.

We'd put the diagnosis more bluntly. Most AI programs fail because the AI doesn't know why the work exists.

• It can read the ticket, but not the budget behind it.
• It can summarize the project, but not the strategy the project is meant to move.
• It can draft the forecast, but not explain which decision changed it.
• It can take an action, but nobody decided in advance who owns that action or what it may touch.

So the AI either asks a person for everything — and saves nobody any time — or it guesses. And because the guesses are fluent, nobody can tell which answers were grounded and which were invented.

The cost of guessing

Guessing at scale is expensive in three ways:

• Direct cost. Tokens, model calls and agent runs spent on work that has to be redone.
• Trust cost. One confidently wrong answer in a board pack or a budget review sets adoption back months.
• Risk cost. Actions taken without a clear owner, permission or record are exactly what regulators and auditors are now asking about.

What this means for the playbook

The fix isn't a better model. It's giving every model the same, governed understanding of how the enterprise runs — and keeping people on the decisions. The rest of this playbook is about how to do that.