Enterprise Data & AI Do Not Fail on Technology. It Fails on People.

There's a satirical "how to guarantee your enterprise AI program fails" checklist making the rounds this month. Start with the technology. Set objectives no one can measure. Stand up a specialist Data and/or AI team and keep them well away from anyone who understands the business. Launch a hundred proofs of concept and ship none of them. Above all, wait for perfect governance before you deploy anything.

It reads as a joke. It lands because every line is a real organization, right now, with budget and a steering committee.

And it points at the one thing most leaders still refuse to say out loud: enterprise AI almost never fails on the model. It fails on the people work nobody wanted to do.

The receipts

MIT's Project NANDA spent the first half of 2025 reviewing more than 300 disclosed AI initiatives — alongside interviews with 52 organizations and 153 senior leaders — for its report The GenAI Divide. The headline finding: 95% of organizations are getting zero return on generative AI. Only 5% of integrated pilots extract real value. And the report is unusually direct about the cause. The divide, it says, "does not seem to be driven by model quality," but by approach — tools that never learn the business and never get wired into how the work actually happens.

S&P Global Market Intelligence surveyed over 1,000 enterprises. 42% abandoned most of their AI initiatives in 2025 — up from 17% a year earlier. The average company scrapped roughly 46% of its proofs of concept before they ever reached production.

HCLTech's 2026 survey of 467 executives projects that nearly 43% of major AI initiatives will fail — and is blunt about why. The risk, it says, "is not driven by lack of experimentation or access to tools, but by the difficulty of translating ambition into consistent, enterprise-wide outcomes." Execution, not technology. WRITER's 2026 report names the wound plainly: 54% of C-suite leaders say adopting AI is "tearing their company apart."

Read that back. The model is the part that works. The organization is the part that doesn't.

The people work is the work you're avoiding

Here is the uncomfortable part. Almost every failure mode in that satirical checklist is a decision to skip the people work and hope the tool covers for it.

The anti-recipe:

• Start with the technology. Months spent on vendors, models, and architecture; business outcomes deferred to "later." You bought a platform and called it a strategy.
• Keep ownership vague. If everyone is accountable, no one is. When the system does something wrong, there is no name attached to the decision — only a committee and a finger-pointing exercise.
• Stand up a separate AI team. Position them as the people who "do AI," and let everyone who actually understands the business carry on exactly as before. Now your AI is fluent in everything except how your company works. The same applies to data teams.
• Measure activity, not outcomes. Count prompts, licenses, experiments, and training hours. Manufacture the appearance of progress while carefully never measuring whether anything improved.
• Wait for perfection. Perfect data, perfect governance, perfect operating model. Ship nothing — while a competitor with worse documentation and a live product quietly takes the market. On the other hand, never working on data quality even if afterwards, will guarantee you waste everyone's time and money.

None of these are technology problems. They are the same transformation habits that sank the last data program, the last ERP rollout, and the last "digital" initiative — wearing a newer, more expensive logo.

And now the system decides before you do

This used to be a slow-motion failure. It isn't anymore.

This week Corporate Compliance Insights ran a piece titled "Meet Your New Colleague. It's Already Making Decisions." That is the shift in one sentence. The autonomous agent in your stack approves the refund, merges the change, sends the email — and then, maybe, someone reviews it. We wrote about that inversion in When AI Decides First and Asks Permission Never.

The irony is hard to miss. The same week, Accenture's stock had its worst day on record and fell to its lowest level since 2017 — partly on fears that AI is now disrupting the very consultancies that sell AI transformation. Even the people selling the deck are being governed by the technology faster than they can govern it.

Vague ownership was survivable when a human sat between every decision and its consequence. It is not survivable when the decision-maker is a process running at machine speed and no one ever agreed, in writing, what it is allowed to do.

Governance is a discipline, not a deck

When leaders finally feel this, the reflex is to reach for the old tools: another steering committee, a 100-slide framework, a sign-off chain with seventeen stakeholders. That is governance theater. It produces excellent documentation and zero enforcement... negative value, only fatigue.

The teams that get this right treat governance the way good engineers treat data quality — not as a project with a start date and an end date, but as a continuous discipline with clear business owner and technical custodians. The shift looks like this:

That right-hand column is the whole idea behind statement-first governance. You write down what an agent — or a team — is allowed to do, as an explicit, owned commitment. You bind that commitment to the place the work actually happens, so it can be enforced and audited rather than just admired on a slide. Trust stops being assumed at the moment of deployment and starts being earned continuously, on the record.

It is the same move that turns a flaky data pipeline into a reliable one: stop trusting by default, write the contract, enforce it at the source. We made the longer case for managed, accountable agents in Managed Agents vs. Loose LLM Calls, and for why people stay the core challenge in The Human Factor.

Bottom line —

Your AI program will not fail because the model or the platform are bad. It will fail because no one wrote down who is accountable, what "good" looks like, or what the agent is allowed to do — and then a machine started making those decisions for you, at scale, before anyone was watching.

Score yourself

The satirical checklist ends with a question, so we will end with a sharper one. Pick your most autonomous AI use case in production today. Then answer three things out loud, to a colleague:

1. Who owns the decisions it makes? — by name, not by committee.
2. Where is it written down what it is, and isn't, allowed to do?
3. If it does the wrong thing tomorrow, what record proves what it was supposed to do?

If any answer is "we'd have to figure that out," you don't have a technology problem. You have the only kind of problem that actually kills both data and AI programs.

The models get a new version number every few months. The organizational habits never get patched at all. That is exactly why the people work is the work — and the organizations that ship a governance discipline instead of a governance deck are the ones still standing when the next model drops.

Further reading

• When AI Decides First and Asks Permission Never
• The Human Factor: Why People Remain the Core Governance Challenge
• The Ownership Vacuum: When Nobody's in Charge Anymore
• Managed Agents vs. Loose LLM Calls: Who Pays the Governance Bill?
• AI Governance & Compliance in 2026 and the agentic governance docs

Sources

• MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 — full report (PDF)
• S&P Global Market Intelligence, Voice of the Enterprise: AI & ML — reported by CIO Dive
• HCLTech, The AI Impact Imperatives (2026) — press release
• WRITER, AI Adoption in the Enterprise (2026) — report
• Meet Your New Colleague. It's Already Making Decisions. — Corporate Compliance Insights