Why implementation is the new bottleneck of enterprise AI

Implementation, not model quality, is the bottleneck of enterprise AI in 2026. 95% of enterprise GenAI pilots produce no measurable P&L impact (MIT NANDA, Aug 2025) and only 22% of organizations have scaled AI across business units (Gartner, Sep 2026). The models got better every quarter. The results did not follow.

The numbers, from four angles

Four sources that share no methodology reach the same place.

• MIT NANDA, The GenAI Divide (Aug 2025): 95% of enterprise GenAI pilots have no measurable impact on the P&L. Solutions bought from a vendor succeed about 67% of the time; internal builds about 33%.
• Gartner (1 Sep 2026): only 22% of organizations have scaled AI across several business units, while 85% of functional leaders plan to increase spend.
• Menlo Ventures (2025): of $37B in enterprise GenAI spend in 2025, 76% was bought rather than built, and only 16% of deployments qualify as real agents.
• McKinsey, The State of AI 2026: 40% of large companies already scale agents, up from 27% a year earlier, and 32% decided not to buy at least one piece of software because of AI.

Read together: budgets go up, purchases go up, and the share of companies with a measured result stays flat. The gap is not in the model. It is in everything around it.

Why the model is never the reason

When a pilot dies, the post-mortem rarely says the model was wrong. It says the integration with the ERP was never finished. Or that nobody owned the KPI, so nobody could say whether the pilot worked. Or that the agent ran fine for six weeks and then drifted, because nobody was maintaining it after the project closed.

The CTO of Ode, the $1.5B deployment company Anthropic launched with Blackstone, Hellman & Friedman and Goldman Sachs in 2026, put it in one sentence: model selection matters, but it is not where most of the effort goes. The effort goes to context, integration, evals and change management. That is implementation work.

The market already priced this in

Follow the money. Palantir, which built the Forward Deployed Engineer model more than a decade ago, reported $1,935M in revenue in Q2 2026, up 93% year over year, with U.S. commercial growing 149% (Business Wire, Aug 2026). In May 2026 OpenAI created The Deployment Company with more than $4B of initial investment; Anthropic and its partners followed with Ode and $1.5B. FDE job postings grew more than 1,000% year over year at the start of 2026 (GetPerspective, 2026), and a16z opened an FDE Fellows program in July.

None of these companies sells a model. They sell the work of putting models into production inside someone else's systems. That is where the margin moved.

The counterweight

The same market that funds implementation is also cancelling it. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027 (Jun 2025), coined the term agent washing after finding that only about 130 of thousands of vendors ship real agents, and estimates that only around a quarter of AI use cases in customer service produce ROI (Aug 2026). Alvarez & Marsal documented the structural cost in April 2026: FDE-based businesses carry lower gross margins and scale linearly with headcount unless they turn client work into product with discipline.

So implementation is where the value is and also where the failures are. That is not a contradiction. It is the definition of a bottleneck.

It is happening in México, now

This is not a distant wave. Wonderful opened Latin America in March 2026, México included, hiring Deployment Strategists and FDEs. Primero raised $12M in August 2026 and operates from Ciudad de México. HappyRobot runs FDEs in eleven cities, one of them Ciudad de México. General Catalyst's memo on Primero says it plainly: enterprise AI is the first software category that can be won at a regional level, because forward deployed engineering is fundamentally local.

We agree with the memo. We have also been doing the local part for twelve years, with 250+ engineers in seven countries and 150+ enterprise clients whose ERPs, CRMs and contact centers we already know from the inside (DaCodes, Sep 2026). Whoever is looking for enterprise AI agents is really looking for the people who can make them work inside the company.

What this means if you are buying

Four things we would ask any vendor, including us.

• Which system of record does the agent write to, and who owns that integration after go-live?
• Which KPI does the pilot move, measured against what baseline, and what happens if it does not?
• Who runs the evals every week after the project team leaves?
• How many weeks from discovery to production, and how many of your last ten pilots hit that number?

Our answers: the agent lives inside your ERP, CRM or core, in your cloud; the KPI is written into the contract before we start; evals run from day one and continue under a retainer; and a pilot reaches production in 8 to 12 weeks. Ask everyone the same four questions and the market gets a lot smaller.