AI deployment starts with the workflow, not the model.

AI adoption is accelerating.

Organisations are buying licences, experimenting with copilots, building agents, automating tasks and training employees. All of that can be useful.

But none of it answers the harder question: what part of the organisation now works materially better because AI was introduced?

That question marks the difference between using AI and deploying AI. Using AI changes an activity. Deployment changes how work gets done. It builds on the progression set out in Field Note 01, on AI literacy and deployment capability.

And that is why the starting point should not be the model. It should be the workflow.

The technology-first question is usually the wrong one

A common starting point is: where can we use AI?

That sounds reasonable, but it immediately pulls the conversation towards tools. Which model? Which agent platform? Which automation product? Which copilot?

A stronger starting point is: which important workflow is currently too slow, fragmented, inconsistent, expensive or difficult to operate well?

That question forces the organisation to begin with work rather than technology. Instead of asking what an AI system can do, leaders begin asking what the organisation needs to improve. That distinction matters.

A workflow exposes the real problem

Consider customer support.

“Build an AI chatbot” is a technology project.

“Reduce routine resolution time while ensuring complex and sensitive cases reach the right human operator” is a workflow problem.

The second formulation immediately creates better questions:

• Where does the relevant information live?
• Which requests are repetitive?
• Which require judgement?
• When should the AI answer?
• When should it escalate?
• Who owns the knowledge base?
• How are errors detected?
• What should be logged?
• What happens when information changes?
• What outcome will determine whether the system is actually better?

The model becomes one component of the solution. It does not define the solution.

The same logic applies to procurement, finance operations, recruitment, sales, compliance, knowledge management, project delivery, customer service and many other areas of work.

The model is only one layer

A reliable AI-enabled workflow usually requires more than intelligence generation. It requires some combination of:

• people
• data
• tools
• business rules
• workflows
• integrations
• permissions
• governance
• measurement
• AI models and agents

The quality of the model matters. But operational value depends on how these elements work together. A highly capable model placed inside a poorly designed workflow can simply produce faster confusion. A more modest model inside a well-designed operating system may create significantly more value.

This is why access to powerful AI will not automatically create organisational advantage. Access is increasingly becoming common. Orchestration is not.

Workflow redesign is emerging as an important value lever

Research increasingly points towards the operating system around AI, rather than model access alone, as an important determinant of value.

McKinsey’s 2025 global survey examined 25 organisational attributes associated with reported financial impact from generative AI. Among those attributes, workflow redesign showed the strongest association with respondents reporting EBIT impact from generative AI.

Yet only 21% of respondents whose organisations reported using generative AI said that their organisations had fundamentally redesigned at least some workflows.

This should not be interpreted as proof that workflow redesign by itself causes financial performance. It does, however, reinforce an important operating hypothesis: AI creates more value when organisations redesign how work happens around it rather than simply adding new tools to existing processes.

Automation is not the same as redesign

Organisations sometimes take an inefficient process and automate parts of it without reconsidering the process itself. That can make a bad workflow faster.

The better question is not: which steps can AI automate?

It is: what should this workflow look like now that new forms of intelligence are available?

• Some steps may disappear.
• Some may be automated.
• Some may require stronger human review.
• Some decisions may need better data.
• Some hand-offs may no longer be necessary.
• New controls may have to be introduced.

The objective is not maximum automation. The objective is better work.

Human responsibility does not disappear

AI deployment also creates an accountability question. When an AI system participates in real work:

• Who owns the outcome?
• Who reviews uncertain cases?
• Who can override the system?
• What information is the system allowed to use?
• What actions can it take?
• What happens when it fails?
• Who investigates?
• Who decides whether it remains in operation?

These questions cannot be delegated to the model. The more consequential the workflow, the more explicit these responsibilities need to become.

That means governance is not something added after deployment. Governance is part of the system design.

NIST’s AI Risk Management Framework provides one useful reference point. Its core functions organise AI risk management around Govern, Map, Measure and Manage, with governance intended to operate across the AI lifecycle rather than as a final compliance step.

The framework is voluntary and sector-neutral. NIST is currently revising AI RMF 1.0, so it should be treated as a useful current reference rather than a permanent or exhaustive operating standard.

The workflow should have an owner

One of the clearest signs that an AI initiative is still experimental is the absence of operational ownership.

A prototype may have a creator. A pilot may have a sponsor. A deployed workflow needs an owner.

Someone must be accountable for:

• its purpose
• its users
• its data
• its controls
• its performance
• its exceptions
• its improvement
• and eventually its retirement

Without ownership, AI initiatives tend to remain demonstrations rather than operating capabilities.

Measurement must begin before deployment

Another common mistake is deciding what to measure after the system is already live. By then, the organisation may have no meaningful baseline.

If the workflow is important enough to redesign, the desired improvement should be defined at the beginning. Depending on the workflow, that might include:

• cycle time
• resolution time
• error rate
• rework
• cost
• throughput
• quality
• user adoption
• escalation rate
• customer experience
• employee effort
• risk events

The correct measure depends on the work. But the principle is universal:

If the organisation cannot define what improvement would look like, it is not yet ready to claim that AI created value.

A practical deployment test

At ORCHALYS, we use five questions to keep the workflow rather than the technology at the centre of deployment.

1. MAP

   What happens now, and what result matters?

2. BUILD

   What should be created and connected?

3. DEPLOY

   How will the system enter real use?

4. EMBED

   How will reliable operation become normal?

5. PROVE

   What changed, and what should improve next?

The discipline matters because an AI prototype can demonstrate technical possibility without demonstrating operational value. A deployment has to survive contact with real work.

From AI projects to operating capability

This distinction will become increasingly important as AI becomes easier to access. When many organisations can access capable models, model access becomes less differentiating.

The harder competence becomes:

• knowing where intelligence belongs
• redesigning the surrounding workflow
• connecting the required systems
• assigning ownership
• embedding governance
• driving adoption
• and measuring whether anything actually improved

That is where AI moves from experimentation into operational capability, and it is the work ORCHALYS Deployment exists to do. And that is where orchestration becomes valuable.

The leadership question is changing

The early AI question was: which tools should we adopt?

Then it became: how do we get our people using AI?

The next question is harder: which parts of our organisation now operate better because intelligence has been deliberately embedded into the way work happens?

That is a much higher standard. It moves the conversation:

• from models to systems
• from activity to workflows
• from usage to outcomes
• from pilots to operations
• from experimentation to accountability
• and ultimately, from AI adoption to AI deployment

The model matters. But the workflow is where value becomes real.

Intelligence, orchestrated.