In 2025 MIT Project NANDA published “The GenAI Divide: State of AI in Business 2025”, built on 150 executive interviews, 350 employee surveys and an analysis of 300 public AI deployments. Its headline finding has become the most-quoted number in enterprise AI: 95% of generative-AI pilots deliver no measurable profit-and-loss impact. Against tens of billions of dollars of enterprise spend, only a small minority of organisations could point to a return.
The number is usually read as an indictment of the technology. It is not. The report’s own explanation is a learning gap: most deployed systems do not retain feedback, do not adapt to the organisation’s context, and do not fit the workflows people already run. The models were good enough. The delivery was not.
Where pilots actually die
Coventa has watched enough demos die on contact with real operations to recognise the pattern. A pilot succeeds in a sandbox with clean data and a willing sponsor. Then it meets the enterprise.
- Integration. The system has to read from and write to an ERP, a CRM, an EHR, a ticketing tool or a plant control layer that was never designed for it. The pilot’s API mock becomes six months of work.
- Data. Production data is incomplete, inconsistent and governed. The pilot was built on an extract; the system has to live on the real thing.
- Approvals. Every consequential action needs an accountable human, an audit trail and a rollback. A demo has none of these and the organisation cannot run without them.
- Ownership. The vendor built the pilot; the strategy firm recommended it; nobody owns getting it into production and keeping it there. The gap between the deck and the delivery team is where the value goes.
- Learning. Nothing captures what users correct, so the system is exactly as wrong in month six as in week one, and trust drains away.
Why 2026 became the year of the forward deployed engineer
The industry’s answer to the divide has been organisational rather than algorithmic. The forward deployed engineer — an engineer embedded with the customer, accountable for a system in production rather than a recommendation about it — was Palantir’s model for years. In 2026 it became everyone’s. AWS announced a $1 billion forward-deployed engineering organisation on June 30, 2026, describing pods of engineers that compress deployment timelines from months to days and leave customers self-sufficient. OpenAI launched a deployment-services company; Anthropic launched an AI-services venture of roughly $1.5 billion; Deloitte stood up a named Forward Deployed Engineering practice. LinkedIn’s January 2026 labour-market report found forward-deployed-engineering roles grew 42-fold between 2023 and 2025.
The reason is simple. Every item on the list above is delivery work that has to happen inside the customer’s operation, by someone who can both understand the business and write the code. That is what a forward deployed engineer is.
What changes when an FDE owns the last mile
- Discovery and delivery are one motion. The people who learn the landscape and audit the workflows are the people who build. There is no handoff for the understanding to be lost in.
- Governance is implemented, not recommended. Bounded autonomy, an immutable audit trail, human approval on consequential actions and a dry-run before any live write are built into the system on day one, because the FDE has to run the system, not just describe it.
- The unit of value is the outcome. An FDE can be paid on the result — cycle time, cost taken out, incidents prevented — measured against a frozen baseline. A time-and-materials team cannot be, and its incentive runs the other way.
- The organisation keeps the capability. Under hire-and-manage or build-operate-transfer, the team, the platform and the process end up inside the organisation. The pilot does not leave with the vendor.
How Coventa applies it
Coventa built VisionX, its AI forward deployed engineer, to do the understanding first and the execution second in one continuous motion. Landscape, Audit and Blueprint compress into days the discovery a consulting scope takes months to produce; Architecture, Migrate and Build ship the system in staged, tested, audited releases; and Coventa’s engineers stay on to run and improve it. Clients typically see a working first version within the first week and production within weeks, and can hold the capability on Coventa’s payroll, hire-and-manage it, or take it over through build-operate-transfer.
Five questions before the next pilot
- Who owns production? Name the person accountable for the system running in your environment, not the demo.
- What does it integrate with on day one? If the answer is “an extract”, it is a demo.
- Where is the human gate? Which actions can the system take alone, which need approval, and where is that enforced in code?
- How does it learn? What captures corrections, and who reviews them?
- What do we own at the end? A system, a team and a process — or a licence and a slide deck?
The 95% is not a verdict on AI. It is a description of what happens when nobody owns the last mile. Change who owns it, and the number changes.