Why 95% of AI projects fail (and the operating fix most companies skip)

MIT's State of AI in Business 2025 report put a number on something every operator already felt: roughly 95% of enterprise generative AI initiatives fail to deliver measurable financial value. Not 95% of bad ideas. 95% of funded, board-sponsored, vendor-backed projects. The ones with real budget and real intent. They stall in pilot, or they ship and never move the P&L.

I have sat in the CEO or COO chair nine times across PE-backed companies, and I have led or supported more than fifty acquisitions. So let me say the part most AI vendors will not: the model is almost never the problem. The 95% number is not a technology failure rate. It is an operating-model failure rate wearing an AI costume.

Here is how it looks from inside the company. The board asks about AI. A competitor announces something. A vendor runs a slick pilot. Six months later there are three pilots, two of them stalled, and a CFO who has stopped signing the checks. Everyone has a theory. The vendor blames change management. The team blames the tools. Nobody wants to say the quiet part: the workflow the model was bolted onto was already broken before the model arrived.

AI does not fix a company that cannot make a decision. I have watched it make the dysfunction faster. If three people own the AI roadmap, no one can kill a vendor or approve a deployment, so the burn continues for another two quarters. If the workflow lives in one person's head, the model has nothing to anchor on. AI amplifies whatever operating system it is deployed into. On a clean operating backbone it compounds advantage. On broken operations it compounds dysfunction.

That is the whole thesis, and it is why the order matters more than the algorithm. Operations first. AI second. The companies that win the next five years will not be the ones with the best AI strategy. They will be the ones whose operating model is strong enough to absorb it.

When I diagnose a stalled AI portfolio, I am not looking at model benchmarks. I am looking for four things. Do decisions close, or do they loop? Does every initiative have one owner who can act, or does everything escalate to the CEO? Is the workflow documented well enough that a model can plug into it, or does it live in tribal knowledge? And is there a cadence that forces the rollout to finish, or does the pilot drift between "we are refining" and "we will evaluate" forever? When those four are missing, AI is a science experiment with a board update attached.

So the work has an order of operations, and we run it the same way every time. First, audit the AI portfolio honestly. Which pilots are stalled because of model quality, and which are stalled because the workflow underneath was already broken? Cut the second pile. Most of the spend is in the second pile.

Second, fix the workflows worth fixing. Decisions close. Ownership lands on a single name. The work the model needs to plug into gets documented. This is not a six-month change-management program. It is operating discipline installed exactly where the AI is supposed to live.

Third, ship AI where the margin actually lives. Not where it is trendy, and not where a vendor has a demo. Production deployments measured against revenue or margin, not adoption metrics. A dashboard that says usage is up is not ROI. A line on the P&L is ROI.

Fourth, build the AI operating model so it holds. Governance tied to outcomes. Vendor rationalization, so you are not paying for six tools that do one job. A roadmap your CFO will defend in the next board meeting instead of apologize for. That is the difference between a portfolio of experiments and an operating capability.

We run all of this on VOOCS, our execution operating system, which is really just the operating system AI needs to run on. Vision, so every team is not building its own pilot. Outcomes, so an AI initiative has a metric, a target, and a deadline instead of a vibe. Ownership, so one person can decide. Cadence, so evaluation turns into rollout in weeks instead of quarters. Systems, so the workflow is clean enough for a model to anchor on. If your AI ROI is in single digits, one of those five is broken.

I am not speculating about this. Every AI ROI band we have shipped, 3.8x to 5.1x in six to nine months, was on an engagement where the operating foundation was already in place or installed first. We rebuilt a voice-AI sales tool that had been collecting dust and took adoption from 18% to 92%, but only after we fixed who owned ramp and installed a coaching cadence. We deployed an AI virtual agent that now resolves 55% of tickets at 91% CSAT, because we set the decision rights on what it resolves versus escalates before we built it. We shipped an AI knowledge assistant that cut internal tickets 68%, after we consolidated a source of truth that did not exist yet. In every case the AI was the headline. The operating fix was why it worked.

The most useful thing I can say to a CEO whose board wants AI is also the thing most vendors will never say: AI might not be your next move. If your top performers left tomorrow and the company would stop running, you do not have an AI problem, you have a systems problem, and AI will accelerate the dependency instead of solving it. Fix that first. Then the AI you turn on compounds, because it finally has something solid to compound.

That is the entire game. The sequence is the strategy. Operations first. AI second. AI compounds advantage, or it compounds dysfunction, and the operating model is what decides which one you get. Heroes don't scale. Systems do. Get the operating model right, then turn on AI.