Concede the awkward fact at the centre of the AI commercial story: enterprise AI adoption lagged the consensus expectations of 2024 and 2025. Major analyst forecasts in early 2024 projected enterprise AI spending to reach $200-300 billion annual run-rate by mid-2026. The actual run-rate as of Q1 2026 is closer to $80-100 billion globally, depending on how the category is defined. The gap is meaningful — somewhere between $100B and $200B of expected spend that did not materialise on the original timeline.
The lag is not caused by lack of capability. Frontier models in 2026 are dramatically more capable than they were in 2024. The lag is caused by the structural difficulty of embedding AI capability into enterprise workflows. Buying API access is easy. Building production workflows around the API that deliver measurable business value is hard. Most enterprise customers found out the hard way that the second part is where the actual cost and time live.
The frontier labs are now responding to this adoption gap with a specific strategy: force workflow embedding through services entities. The OpenAI Deployment Company, the Anthropic services partnership, and equivalent moves at Google are all framed by the public coverage as growth investments. The more accurate framing is that they are responses to slower-than-expected organic adoption.
The Adoption Gap Data
Three concrete data points on enterprise AI adoption progress through Q1 2026:
McKinsey 2026 State of AI survey: Enterprise AI usage has increased substantially year-over-year, with 78% of large enterprises now reporting some AI deployment versus 65% in 2025. The number that has moved less: only 23% of large enterprises report measurable revenue or cost impact from AI deployments, up from 19% in 2025. The gap between deployment and impact is the structural issue.
Gartner enterprise AI spend tracking: Year-over-year growth in enterprise AI spend reached approximately 45% in 2025-2026, against forecasts of 80-120% growth. The growth is real but materially below expectation. Notable: spend on AI infrastructure (GPUs, data centers) grew faster than spend on AI applications and implementation, suggesting buyers are still in infrastructure build-out phase rather than application deployment phase.
OpenAI / Anthropic enterprise revenue mix: Both labs have publicly disclosed that enterprise revenue is growing rapidly but from a lower-than-expected base. OpenAI's enterprise revenue run-rate is reportedly in the $4-5B range as of Q1 2026; Anthropic's is reportedly in the $2-3B range. These are large numbers but smaller than the labs would be capturing if enterprise adoption had hit 2024 forecast levels.
The picture is consistent: enterprise interest is high, infrastructure investment is heavy, deployment-to-impact translation is slower than expected.
Where The Embedding Friction Lives
The reasons enterprise workflow embedding is hard are not theoretical. We see five recurring blockers in real enterprise AI deployments.
Process clarity gap: AI is most effective when applied to well-defined, repeatable processes. Many enterprise processes are not actually as well-defined as they appear in process maps — they have exceptions, escalations, and tacit knowledge that humans handle implicitly. AI deployment forces explicit definition of these processes, which is consultative work that often takes longer than the AI implementation itself.
Integration depth: Production AI workflows have to integrate with existing enterprise systems (ERPs, CRMs, data warehouses, identity providers, ticketing systems). Each integration adds friction, technical risk, and political coordination. A typical enterprise AI deployment touches 4-8 integration points, each of which can be the blocker.
Change management: Once an AI-augmented workflow is deployed, the humans involved in the workflow have to change how they work. This is the part that consistently underestimates effort. Adoption rates within an organisation often hit 30-50% rather than the 80-90% that the business case assumed. The remaining workforce continues with the old workflow, partially defeating the deployment.
Quality measurement: Measuring whether the AI workflow actually delivers business value requires instrumentation that most enterprises do not have. Building the measurement layer is its own multi-month project.
Vendor relationship management: Enterprises buying AI capability from frontier labs face new vendor relationship dynamics. The labs ship new model versions on rapid cadences, which can disrupt deployed workflows. Pricing changes. Capability changes. Security and compliance terms change. Enterprise procurement organisations are not yet set up to handle vendor relationships that change on weekly cadence.
The blockers are not surprising in retrospect. They are the friction layer that every previous enterprise technology transition has had to traverse — ERP rollouts in the 1990s, cloud migrations in the 2010s, mobile transformations in the 2010s. AI is following the same friction curve.
The Strategic Response
The frontier labs are responding to the adoption gap by building capacity to handle the friction layer themselves rather than waiting for the friction to resolve. The services entities (OpenAI Deployment Company, Anthropic services partnership) are positioned specifically to address each blocker:
Process clarity: Implementation consultants from the services entities work with enterprise customers to define processes well enough for AI deployment. This is work that legacy services firms have done for decades, and the frontier lab services entities are entering the same market.
Integration depth: AI engineers from the services entities build the custom integrations with enterprise systems. This is more efficient than expecting the enterprise customer to build the integrations internally, both because the AI engineers have deeper integration expertise and because they can leverage standard patterns across multiple customer deployments.
Change management: Embedded engineers from the services entities support enterprise teams through the workflow transition, providing training, troubleshooting, and ongoing optimisation. This is closer to managed services than traditional consulting.
Quality measurement: The services entities build measurement infrastructure as part of every deployment, providing the customer with visibility into AI workflow performance and business impact.
Vendor relationship management: A single point of contact from the services entity handles the vendor relationship complexity, abstracting away the cadence and complexity of dealing directly with the frontier lab.
The strategy is to reduce the embedding friction by absorbing it into the services entity rather than expecting the customer to handle it. The bet is that customers will pay for this absorption, and that the resulting acceleration of enterprise adoption justifies the services entity cost structure.
The Math On Whether This Works
The unit economics of the services entity strategy depend on a specific assumption: that absorbing the friction can be done profitably, even though the friction layer is structurally less efficient than the underlying AI capability.
Consider a representative enterprise deployment. The customer pays $5M for an end-to-end AI implementation. Of that:
- API costs (model usage during deployment and ongoing): $200K-500K - Cloud infrastructure and tooling: $300K-700K - Implementation services (consulting, engineering, deployment): $3M-3.5M - Ongoing managed services (first year): $1M-1.5M
The implementation services and managed services lines are where the services entity captures revenue. The API and infrastructure lines are where the frontier lab's core business captures revenue. The total deal size of $5M is significantly larger than a pure API contract would be, but most of the additional spend is services rather than software.
The services entity revenue at $3-5M per deployment, scaling to hundreds of deployments per year, justifies a multi-thousand-person staffing investment. The math works if the services entity can:
- Maintain implementation services margins above 25% (achievable for AI-native services, harder for legacy services firms) - Maintain managed services margins above 35% (achievable for ongoing optimisation work) - Hit utilisation rates above 75% across the services workforce (challenging during build-out, achievable at scale)
These margin and utilisation targets are achievable but not guaranteed. Execution risk is real.
Why This Strategy Beats Waiting
The alternative to the services entity strategy is to wait for enterprise customers to develop internal AI capability and embedding skills. This is the path that legacy enterprise software companies have largely taken — Salesforce, SAP, Oracle, ServiceNow do not run large implementation services arms (they partner with services firms instead).
The frontier labs are choosing not to wait. Two reasons.
Competitive urgency: The enterprise AI market is currently in a land-grab phase, where customer commitments made in 2026-2027 will lock in vendor relationships for years. Frontier labs that capture deep workflow embedding in this window create switching costs that compound. Waiting for organic adoption surrenders this position to whichever vendor moves first.
Revenue acceleration: Frontier lab valuations depend on revenue growth rates. Waiting for enterprise customers to develop embedding capability internally adds 2-4 years to the path. Building services capacity accelerates the revenue curve by capturing the embedding spend directly. The lower margins on services revenue are offset by the higher total deal size and faster ramp.
The strategy is forced by competitive dynamics and capital market expectations. Whether it works at scale is the next 24 months of execution. The financial markets have already priced in expectations of strong execution — the funding rounds at Anthropic and OpenAI close at valuations that assume the services strategy materially expands revenue.
What This Means For Customers
For enterprise customers, the services entity strategy creates two practical implications.
Faster path to AI deployment: Customers who engage with frontier lab services entities can plausibly deploy AI workflows faster than customers who try to build internal capability or work with legacy services firms. The bundled offering compresses time-to-value.
Deeper vendor dependence: Working with a single frontier lab's services entity creates deeper coupling to that lab's models, tooling, and roadmap. Switching costs increase. This is a tradeoff customers should evaluate explicitly rather than accepting by default.
The honest read for enterprise customers in 2026: the services entity strategy is good news for organisations that want to deploy AI capability quickly and don't have strong views about vendor neutrality. It is bad news for organisations that want to maintain optionality across multiple AI vendors and minimise lock-in.
Most enterprise customers will probably engage with frontier lab services entities for at least pilot deployments, given the friction reduction benefits. The longer-term question of vendor lock-in will become acute over the next 24-36 months as deployments scale and switching costs compound. By 2028, the enterprise AI vendor landscape will probably look much more concentrated than it does today, with frontier lab services entities driving much of the consolidation.