In many organisations, the first wave has already happened. Teams have prototypes. They have pilots. They have small wins. They have a sense that something has shifted. Then production happens. Production is where the real world starts pushing back. Not with drama. With quiet friction. With delays. With odd edge cases. With approvals that stall. With data that is missing at the moment it is needed.
In that moment, the question changes
It stops being “Can this model do the task?”
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
It becomes “Can the organisation support the action?”
That is why architecture becomes the constraint.
Capability is rarely the limiting factor
Modern AI can summarise, classify, draft, search, and recommend at a level that surprises people. It can do this at speed, and with a consistency that looks like progress. But capability is not the same as usefulness. Usefulness depends on whether the output can move through the organisation safely. It depends on whether the output can be checked, traced, corrected, and owned.
When AI moves from assisting to acting, the seams in a technology stack begin to show.
US Tariffs are shifting - will you react or anticipate?
Don’t let policy changes catch you off guard. Stay proactive with real-time data and expert analysis.
By GlobalDataThe best systems do not look clever.
They look well-prepared.
Production exposes the seams
Most technology estates were not designed for autonomous behaviour. They were designed for humans using screens, clicking buttons, and taking responsibility in small increments.
AI changes the flow. It compresses time. It removes steps. It blurs the boundary between suggestion and decision.
Production then tests the estate across four pressure points:
- Data: what the AI can see, and what it cannot
- Identity: who the AI is, and whose permission it carries
- Workflow: how work moves, and where it stops
- Oversight: how errors are noticed, and how humans step back in
These are not theoretical concerns. They are the everyday reasons why pilots do not scale.
Data is never as tidy as the demo
In a demo, data is clean.
In production, data is political.
It sits in multiple systems. It is incomplete in places. It is defined differently by different teams. It is sometimes delayed. It is sometimes wrong. And it is often owned by someone who is not in the room. AI can only be as useful as the data boundary it operates inside.
If the boundary is too narrow, the AI is weak.
If the boundary is too wide, the AI becomes risky.
Most organisations discover this boundary by accident, not design.
Identity is the forgotten hard problem
AI tools often start life as helpers. In that mode, identity does not bite. The user is the user. The permissions are the permissions. But once AI is expected to trigger actions, identity becomes central.
What exactly is an AI agent in the eyes of the enterprise?
Is it a user account?
A service account?
A delegated identity?
A shared capability?
And if it acts on behalf of a person, what does that really mean when something goes wrong? Most enterprises have strong identity models for people and systems. They rarely have a mature model for non-human actors operating across workflows. This is not a small gap. It becomes a bottleneck.
Workflow breaks where humans used to be
In many organisations, the real work is not the decision.
It is the handover. Handover is where context is lost. It is where approval lives. It is where exceptions appear. It is where accountability is distributed.
AI can draft the response.
But can it open the case, attach evidence, route it correctly, and close it in the right system?
AI can recommend the change.
But can it raise the request, pass the controls, and record the audit trail in the format the organisation needs?
This is where “intelligence” turns into plumbing. And plumbing is where technology estates show their age. Oversight is not monitoring. It is governance in motion. Most organisations think about monitoring as a technical problem. Uptime. Latency. Error rates. AI in production needs something different. It needs behavioural oversight. Not because AI is malicious. Because AI is variable.
Outputs drift. Context changes. Prompts evolve. Data shifts. Users learn how to game the system. Even small updates can change what “good” looks like.
What matters in production is simple:
- Can someone see what the AI did?
- Can someone explain why it did it?
- Can someone stop it quickly when needed?
- Can someone correct it without breaking everything else?
These are not AI questions. They are architecture questions. Architecture is the quiet determinant of speed
A common mistake is to treat AI as a bolt-on layer. Add a model. Add a chat interface. Add a workflow assistant.
That can work for low-risk, low-consequence tasks. But as AI touches operations, the environment needs to be ready.
Production readiness is not a single checklist.
It is a set of conditions that allow autonomy without loss of control.
In practice, that means the organisation has already invested in the unglamorous things:
- consistent APIs and service boundaries
- clear data contracts between systems
- identity structures that are explicit, not improvised
- escalation paths that reintroduce human judgement at the right moments
- logs that tell a coherent story, not just a technical trace
Some organisations are already doing this work because they had to. AI simply makes the need visible. Banking and payments will feel this early Banking and payments systems live close to consequence. They touch money. They touch access. They touch decisions that are hard to unwind. They operate under scrutiny. They operate at scale.
That does not make banks slower by nature. It makes the cost of failure more real. So the constraint is not imagination. It is production integrity.
In my experience, the organisations that handle production well are not the ones with the most ambitious roadmaps. They are the ones that know where accountability sits when something becomes ambiguous.
AI increases ambiguity.
Architecture is how it is contained.
A simple test that often reveals the truth
There is a quiet test that tells you whether an AI deployment is ready for production.
Ask this:
If the AI makes a wrong decision at 9:05am, who will notice first?
And what will they do at 9:10am?
If the answer is vague, the architecture is not ready.
If the answer is clear, the organisation has a chance.
The next phase will reward the boring disciplines A lot of attention is still on the model. Which one. How powerful. How fast. How cheap. That matters. But it will not be the main divider. The divider will be whether organisations can support autonomous action inside their own estates. That is why architecture becomes the constraint. Not as a slogan. As a reality that shows up in production, every day, in small ways. AI will keep moving forward. That is the direction of travel. Some platforms are already moving quickly to embed agentic capabilities more broadly.
The question is whether enterprise environments can keep pace.
The organisations that do will look calm.
They will not look noisy.
They will not be the ones talking most about AI.
They will be the ones whose systems can carry it.
Dr. Gulzar Singh, Chartered Fellow – Banking and Technology; CEO, Phoenix Empire Ltd
