A team reviewing a report together in an office
ARIFTLY / FIELD NOTES

Practical thinking for difficult technical work.

All insights

PRODUCTION ENGINEERING

September 28, 2026 / 6 min read

Why enterprise AI gets stuck after the pilot

A prototype can prove that a model is capable. Production requires a system that remains useful when the data, workflow and operating conditions become real.

A software engineer working with a laptop inside a data centerARIFTLY / FIELD NOTE

The demonstration is the easy environment

A pilot is often built around selected examples, manual preparation and people who already understand its limitations. Production introduces inconsistent inputs, unavailable systems, changing permissions, ambiguous requests and users who expect the workflow to work without an engineer standing beside it.

The gap is not solved by a larger prompt. It is solved by engineering the surrounding system.

Four production gaps appear repeatedly

These are system-design problems. They require connectors, permissions, deterministic checks, durable state, observability and an explicit operating model around the model call.

  • Context: the system cannot reliably find the current, authorized source of truth.
  • Action: the prototype recommends work but cannot safely interact with operating systems.
  • Control: nobody has defined approval, escalation, audit or recovery paths.
  • Evaluation: the team measures model quality but not whether the workflow improved.

Design the failure path before the happy path expands

Production readiness becomes clearer when the team asks what happens if a source is missing, two systems disagree, the model is uncertain or the destination rejects an action. The answer may be a retry, a deterministic fallback, a human review queue or a deliberate stop. Silent improvisation is rarely acceptable for consequential work.

This is also where human approval should become specific. A reviewer needs the proposed action, its evidence, the uncertainty and an efficient way to correct the result. A generic approval button without context creates delay without meaningful control.

Graduate on operational evidence

A pilot should move forward when it produces a useful outcome across representative work, operates inside the required control boundary and shows a credible economic effect. The decision should include the ongoing cost of exceptions, evaluation and maintenance.

A small workflow that performs reliably can expand. A broad demonstration with no owner, baseline or production path remains a demonstration.

CONTINUE READING

AI adoption

A practical playbook for adopting AI in complex operations

Illustrative scenario

From scattered evidence to a review-ready technical investigation