Beyond the pilot: what it actually takes to put GenAI into production.
GenAI pilots are easy to start and painfully easy to stall. The organisations getting to production are not just choosing better models; they are building release discipline, knowledge controls, and human-operating patterns that survive the first wave of edge cases.
Production GenAI is an operating challenge before it becomes a platform challenge.
The same failure modes kept appearing in the 340 deployments we analysed: uncurated knowledge bases, unclear human ownership, loose release criteria, and no reliable way to inspect why the system behaved the way it did. In most cases, the model itself was not the root problem.
Teams that scaled successfully built a compact production system around the model. They limited the workflow boundary, used retrieval and policy controls deliberately, created escalation rules users trusted, and measured intervention as part of the product rather than as a temporary safety net.
Most GenAI programmes break at the seams between knowledge, workflow design, and human supervision. The teams that scale treat all four layers as product requirements.
What separates production-ready teams from the rest.
Knowledge quality dominates user trust
When retrieval content is stale, duplicated, or poorly permissioned, no amount of prompt tuning can recover confidence for long.
Escalation design matters more than autocomplete flair
Users trust systems that can fail gracefully, show confidence, and hand work back with context more than systems that simply sound fluent.
Release criteria need to include operational metrics
Successful teams gate launches on acceptance rate, citation quality, latency, abuse cases, and supervision effort - not just benchmark scores.
Product ownership cannot be delegated to the lab
The use cases that scale have accountable business owners who decide where the workflow starts, where it ends, and what failure costs are acceptable.
The operating moves that de-risk the first six months.
Constrain the workflow boundary
Choose one high-friction task with a clear handoff and resist the urge to solve adjacent workflows in the same release.
Publish explicit supervision rules
Define when users must review, when the model may act autonomously, and what signals trigger mandatory escalation.
Version the knowledge layer
Treat prompts, retrieval sources, policy filters, and evaluation sets as governed assets with the same discipline applied to application releases.
Measure intervention as product telemetry
Track edits, overrides, rejections, and fallbacks so supervision becomes a source of learning rather than a hidden tax.
Where production GenAI is holding up best.
Service agents are using AI as a case-preparation layer
The most resilient contact-centre deployments summarise prior interactions, surface policy guidance, and draft options before the advisor speaks to the customer.
Knowledge workers are getting structured first drafts
High-value legal, procurement, and HR use cases are narrowing scope to clause comparison, policy interpretation, and document preparation with explicit review checkpoints.
Engineering copilots work best inside release controls
Teams seeing durable gains connect coding assistance to test coverage, secure defaults, and review workflows instead of measuring productivity on code volume alone.
The shift from demo to production happened when we stopped asking if the model was clever enough and started asking whether the workflow could survive real users.David Okonkwo - Global Head, AI & Data
Explore adjacent insights from the same research stream.
The next operating model: small teams, large autonomy, AI in the loop.
A framework for reorganising IT and operations around small outcome teams, shared platforms, and AI-assisted execution without recreating old handoffs.
Research noteThe cloud bill came due. Now what?
Why 62% of enterprises are returning to cloud unit economics, what they are learning about workload fit, and the playbook that is reducing spend without slowing delivery.
Industry brief - BankingCore modernisation in 24 months: a comparison of four playbooks.
A banking industry brief comparing strangler, big-bang, sidecar, and greenfield core-modernisation approaches and the conditions in which each one actually finishes.