The AI conversation in most large organisations has a specific texture. There are the enthusiasts who see opportunity everywhere and want to move yesterday. There are the sceptics who've watched enough technology cycles to be suspicious of this one. And then there's the large quiet middle — people who believe AI will matter but aren't sure what to do about it, who are waiting for the organisation to decide, running pilots that never quite reach production.
That last group is where most of the problem lives.
Organisations run proofs of concept that demonstrate something technically impressive
in a sandbox. The numbers look good. The demo goes well. And then the pilot stalls —
because production data is messier than demo data, because the live system integration
is harder than expected, because nobody budgeted to operationalise, because the sponsor
moved on. The result is a catalogue of successful pilots and no production AI.
There's a real difference between AI adoption and what looks like AI adoption. The
visible kind generates announcements, internal presentations, and vendor case studies.
The working kind looks like a change to how invoice processing runs. It looks like a new
routing rule in the service queue. It looks like a forecasting model that replaced a
spreadsheet the head of supply chain has been maintaining manually for six years. Not a
good press release. But a real, lasting change to how the business operates.
The Data Prerequisite
Getting from pilot to production requires being honest about what's missing. The first
thing is usually data. Not perfect data — that's a fantasy — but data that is consistent
enough, documented enough, and governed enough to support a model used in real
decisions. Organisations that have done the data foundation work move to production AI
faster than those who discover the data problems after the model is built. The discovery
is rarely pleasant.
Organisational Clarity About Decisions
The second thing is clarity about what decisions AI will influence and who is
accountable for them. This is not a compliance formality. It directly determines
whether AI outputs get acted on. A sales rep who doesn't understand how their lead
scoring works will ignore scores that don't match their gut. A supply chain manager who
doesn't trust a demand forecast will override it. A model whose outputs nobody acts on
delivers exactly the same value as a model that doesn't exist.
Infrastructure — Less Glamorous, More Determinative
Production ML requires model versioning, automated retraining, monitoring, drift
detection, integration with the systems where the AI output needs to land, and security
controls that satisfy enterprise governance. None of this is exotic. All of it takes
time to build properly, which is why it doesn't happen in a three-month pilot.
What the Leading Organisations Are Actually Doing
Organisations building genuine competitive advantage through AI are not doing anything
unusual. They are disciplined about the foundation and then they apply AI to problems
where that foundation is ready. The pace at which they're doing it is getting faster.
For organisations still accumulating pilots, that matters.
