AnswersAI transformation

Why do enterprise AI pilots fail to reach production?

In short

Most AI pilots stall for a handful of reasons: they are built apart from the systems where the work happens, they don't learn from use, nobody recorded a baseline, security arrives late, and each one starts from scratch on its own stack. Projects that reach production connect to real systems from the first day, carry governance from the start, and are measured against a number the business already tracks.

95%of organizations in MIT NANDA's 2025 study reported no measurable return from generative AI

40%+of agentic AI projects Gartner expects to be canceled by the end of 2027

2 in 3organizations had not begun scaling AI across the enterprise, per McKinsey's 2025 survey

Why pilots stall, and what gets them to production

Cause	What it looks like	What gets it to production
Built apart from the business	A demo on exported data or a sandbox, with no connection to the ERP, CRM or tools where the work happens.	Connect the real systems in place from the first day, with each source's permissions intact.
A system that doesn't learn	Accuracy stays where it was at launch and feedback goes nowhere. MIT NANDA calls this the learning gap.	Evaluations on real cases, feedback loops and continuous tuning after launch.
No baseline	Nobody recorded cycle time, cost or error rates before the pilot, so success is a matter of opinion.	Record the baseline first and measure against it every month.
Security arrives late	The pilot works, then waits months for a security review it was never designed to pass.	Single sign-on, scoped roles, approval gates and an audit trail from the first build.
Every pilot starts from scratch	Each use case gets its own vendor, stack and integrations, so the second costs as much as the first.	One platform where each new system reuses the connections, the model of the business and the governance.
Costs that climb	Every task runs on the most expensive model, and the bill grows faster than the value.	Route each task to the right model and train smaller models of your own for the high-volume work.

What the research shows

MIT NANDA's 2025 study of enterprise generative AI found that 95% of organizations reported no measurable return, and it traced the gap to systems that do not retain feedback, adapt to context or improve over time. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. McKinsey's 2025 survey found AI in use at 88% of organizations, with nearly two-thirds yet to scale it across the enterprise.

Our read

We think most pilots stall because they are built as experiments that sit beside the business. A pilot on exported data shows that a model can answer questions. Running the work also takes live connections, permissions, exception handling, approvals and someone who owns the number.

The pilots that reach production are built as the first system on a platform: connected to the real systems, governed from the start, measured against a baseline, and designed so the second system starts from what the first one built.

How Fig gets a pilot into production

A Fig pilot puts one workflow into production on the company's own systems in weeks. Fig's forward deployed engineers build it on CosmOS with the people who do the work, with approval gates and a full audit trail from the first release, and measure it monthly against the baseline set before the build. Fig tunes it continuously after launch, and the next workflow reuses its connections, digital twin and governance.

Transformation →Forward deployed engineering →

Frequently asked questions

What percentage of AI pilots make it to production?

Estimates vary with the definition of success. MIT NANDA's 2025 study found 95% of organizations reported no measurable return from generative AI, and McKinsey's 2025 survey found nearly two-thirds of organizations had not begun scaling AI across the enterprise.

What is the biggest reason AI pilots fail?

Pilots are built apart from the systems and people that do the work. A pilot on exported data shows a model can answer; running the work also needs live connections, permissions, exception handling and approvals.

How long should an AI pilot take?

Weeks. A Fig pilot puts one workflow into production on the company's systems, measured against a baseline, and the next system builds on it.

How do you scale AI after a successful pilot?

Build the next system on the same foundation: the same connections, model of the business, evaluations and governance. Each workflow then costs less and ships faster than the one before.

Sources

See it running on your company's real work.

Talk to us See the platform