The pattern has become familiar, an enterprise buys a few thousand seats of a frontier enterprise AI solution. Launch week is genuinely exciting: town hall, prompt library, internal champions, a Slack channel full of people discovering they can now summarize a contract in nine seconds. Usage spikes. Then it decays. By month four, weekly active users have settled somewhere between fifteen and thirty percent, most of them using AI as a better search box. At renewal, someone in finance asks what the return was, and nobody can answer in a currency the CFO recognizes. The seat count gets questioned.
Leadership draws the obvious conclusion: the technology isn't ready yet.
That conclusion is wrong, and it's expensive, because it usually triggers another year of waiting for a capability that already arrived. The technology was ready. What failed was the distribution and enablement model.
Access to AI is not a value strategy
One thing we must address up front is that there is still benefit to broad enterprise rollouts, and we are beginning to see meaningful change across our enterprise customers who have enabled these tools.
Organization-wide access is how a workforce becomes literate. It's how skepticism turns into familiarity. The people closest to the work surface use cases no strategy offsite can produce, and how you avoid the dynamic, where AI is something done “to” the organization by a central team. Breadth of AI usage builds the demand, and the fluency required for success down the road. Skipping this step misses the foundation setting most organizations require.
The mistake is stopping at the rollout and treating universal access as though it were a value strategy and then being surprised when value doesn't materialize on its own.
It doesn't materialize because value in an enterprise is not evenly distributed. In engagement after engagement, the addressable return concentrates in a handful of workflows: high-frequency, high-volume, language-heavy work where the current process is expensive and the quality bar is legible. Claims adjudication. Tier-one support deflection. Underwriting file review. Contract abstraction. RFP response. Those workflows are worth orders of magnitude more than the marginal effect of individuals getting help drafting a status email. This is also where the numbers get large enough for anyone outside the program to notice.
"AI for everyone" spreads effort evenly across a distribution that is anything but even. It has a quieter failure mode too as when a capability belongs to everyone, it belongs to no one. No owner or baseline to measure improvement. So the program gets measured with the only metrics available: seats, monthly actives, and prompts per week, none of which a P&L owner can utilize.
The teams that win do something structurally different. They distribute broadly and invest narrowly. They pick three to five workflows, put a named operating leader on each with a business metric that leader already owns, redesign the work rather than bolting a chat window onto the existing process, and instrument the result end to end. They use those builds as templates for the next wave. Everyone gets the tool and a few teams get the engineering, the process redesign, and the accountability to drive a measurable process change.
Context, not "data quality”
The reflex advice is to fix your data first, yes this is an aging message that gives organizations permission to defer AI while they run a data governance program, and it misdiagnoses the problem. Most enterprises are not short on data. They're short on context, the semantics, definitions, entity relationships, and business rules that let a model use data as the leaders in their organization do.
A model with database access and no context will tell you revenue is up. A model that knows which of your four revenue definitions the board actually uses will tell you something true. Closing that gap on metric definitions, ontologies, retrieval that respects permissions and lineage, and evaluation against known-good answers: that is the new integration work required to realize material value. It is unglamorous, yet this is where true advantage accumulates and it cannot be bought as a product feature.
Governance accelerates your AI outcomes
Governance is usually framed as the thing that slows AI down. In practice the opposite is true in that the absence of a clear posture is what slows it down and deprecates results.
With no standing policy, every new use case revisits the same questions about data residency, PII handling, human review, and model risk from first principles. You end up with six-week reviews for six-day builds, and the enthusiasm generated by your broad rollout dies waiting.
Teams that move fast have pre-cleared entire libraries of data for use in advance, with pre-defined classification and ontology. Guardrails, evaluation standards, and audit trails are standardized for use at the platform layer rather than renegotiated per project.
Unit economics, not "cost acceptance"
The most underrated discipline, and the one most enterprises get wrong, is that inference is a variable cost. Enterprise IT is institutionally wired for fixed-cost seat licensing. That mismatch produces two failure modes, surprise at the first bill followed by a potential freeze, or no measurement at all followed by a surprise at an early renewal.
The teams that win compute cost per unit of work, with measurement from the start to asses value per resolved ticket, per document processed, per case reviewed, assessing it against the fully loaded cost of the process it replaces or augments. Once that number exists, the conversation inverts. You stop asking whether AI is expensive and start asking where you'd like to invest more, because you can demonstrate the return. Treat inference as cost of goods sold, not as an IT line item.
The real lag
The gap between enterprise AI investment and measurable results is not a technology lag. Frontier capability has outrun most organizations' ability to absorb it, and the binding constraint has moved to the operating model: who owns the workflow, how context gets engineered, how risk is pre-cleared, how the unit is priced.
That's genuinely good news, because the operating model is the one part of this you fully control. Enterprise AI rollouts to everyone create substantial, qualitative value, but must be matched by picking select workflows to go deep and transform.
.png)
