Enterprise AI transformation will take years, not months

There is a version of the AI future where companies discover that a large share of knowledge work can be automated, remove layers of labour almost immediately and emerge as radically smaller organisations.

I think parts of that future are plausible.

I do not think it describes how most established enterprises will transform over the next few years.

The technical capability is moving faster than the organisational system around it. A new company can design its workflows, data, tools and roles around AI from the beginning. An incumbent has to change those things while continuing to serve customers, operate legacy systems, manage risk, develop its workforce and protect a business that already works.

Enterprise AI transformation is therefore a multi-year operating-model problem, even when the underlying technology changes much faster.

New companies get to start from the new constraint

FundRobin has given me a useful view of what is possible when AI is part of the operating model from the beginning.

A small venture can use agents for research, planning, implementation, testing, content, operations and analysis. It can build shared organisational memory for machines as well as humans. It can design work queues, approvals and tool permissions with agentic execution in mind rather than adding them later.

There is no large installed workforce whose roles were designed around the previous technology. There is less legacy architecture. Processes do not have twenty years of exceptions embedded in them. The organisation can decide that a workflow should work differently and change the workflow itself.

That is a significant advantage.

It is one reason I expect new AI-native competitors to emerge in categories that currently look difficult to disrupt. They do not necessarily need a dramatically better model. They can have a structurally different cost base and a much shorter path between a decision and execution.

But that is only one side of the story.

Incumbents have assets that do not appear in a model benchmark

I have spent much of my career inside large organisations or helping them make technology and transformation decisions. The existing business contains much more than technical debt waiting to be deleted.

An incumbent often has things a new entrant would spend years trying to build:

• trusted customer relationships;
• proprietary data accumulated through real operations;
• distribution and brand recognition;
• deep integration into customer processes;
• regulatory and domain knowledge;
• contracts, partnerships and physical assets;
• employees who understand exceptions no process map fully captures.

These assets can make transformation slower because they create more dependencies. They can also become the raw material for a much stronger AI system.

A model with generic knowledge is useful. A model connected to reliable organisational context, proprietary data, existing customer workflows and the ability to act through established systems can be much more useful.

The question is whether the incumbent can expose those assets to AI safely and redesign the work around them.

The hard part is changing the operating system while it is running

I am sceptical of transformation narratives that start with “how many roles can AI remove?”

That may eventually be one outcome. It is a weak starting point for the work.

A large organisation has to ask much more operational questions.

Which knowledge can an agent rely on? Which systems can it access? Which actions are reversible? What approval is required? What happens when the model is wrong? Which process has grown around regulatory or customer requirements that cannot simply be removed? How does a team work differently when part of the execution can happen in seconds?

Then there is the human transition.

A workforce designed around one operating model does not become AI-native because employees receive licences for a new assistant. Roles, incentives, decision rights, governance and management expectations have to change too.

I saw a version of this in more conventional transformation long before generative AI. In BT workplace strategy, the economics of consolidating the estate could not be separated from where colleagues lived, travel-time impacts, expected leavers and the transition support required. The spreadsheet could identify the target footprint. The organisation still had to move from one reality to another.

AI transformation has the same problem at a much larger scale.

The target state can be technically obvious and organisationally difficult.

AI changes the shape of the transformation team

Five years ago, much of the automation I had seen was deterministic. In robotics work at BT, a process could be broken into repeatable steps driven by known data and rules. The transformation task was to identify the process, redesign it and automate the stable parts.

Agents change the boundary.

A system can now retrieve new information, interpret ambiguous input, select tools, draft a response, inspect results and continue working within a defined scope. That means much more of the execution layer can potentially move from people to software.

But the more capable the execution becomes, the more important the surrounding judgement becomes.

Someone still has to decide:

• which customer problem deserves redesign;
• what information is authoritative;
• which tools and systems the agent can use;
• what quality threshold is acceptable;
• where human accountability remains necessary;
• how the new workflow should fit with the old one during transition;
• which operating evidence should change the next decision.

I expect transformation leadership to become more integrative because of this. The job is not only to sponsor AI use cases. It is to connect business economics, process design, data, architecture, controls, adoption and continuous learning.

The first wave may increase competition before it reduces complexity

There is an uncomfortable period ahead for many incumbents.

AI-native entrants can create new products and operating models quickly. At the same time, established businesses cannot stop serving current customers while they redesign themselves.

That can create a temporary asymmetry. The challenger has less baggage. The incumbent has more assets.

If the challenger can build customer trust and distribution before the incumbent changes its operating model, the speed advantage matters enormously.

If the incumbent can combine its relationships, data and process integration with genuinely redesigned AI-enabled workflows, those same assets can become difficult for a new entrant to replicate.

The strategic question is more specific than “will AI disrupt incumbents?”

I am more interested in which incumbents can convert their existing advantages into an AI-native operating advantage before organisational drag overwhelms them.

Transformation has to happen in layers

I suspect most large organisations will move through overlapping stages rather than one dramatic switch.

The first is individual augmentation: people use AI to research, write, analyse and code faster.

The second is workflow augmentation: AI becomes a defined component inside an existing process, with clear inputs, outputs and human review.

The third is workflow redesign: the process itself changes because an agent can retrieve context, reason and act.

The fourth is operating-model redesign: roles, work queues, decision rights, organisational knowledge and management structures change around machine execution.

The fifth is business-model pressure: once the cost, speed and service model have changed enough, the organisation has to revisit pricing, structure, competitive position and possibly the shape of the workforce itself.

Different functions will reach those stages at different times. Some work will remain human-led for good reasons. Some processes will turn out not to need AI at all.

That messiness is normal transformation, not evidence that AI has failed.

The danger is treating the transition as a software deployment

Enterprise technology programmes often have a go-live date.

An AI operating model does not become real on the day the software is enabled.

The organisation has to learn which contexts the agent can trust, where people override it, where the workflow fails, what new risks appear, which skills become more valuable and which pieces of the old process no longer make sense.

That learning changes the design.

I increasingly think about AI transformation as three connected systems: a decision system that chooses where change creates value, an execution system that turns the decision into controlled work, and a learning system that feeds operating evidence back into the next decision.

The companies that treat AI as a one-off rollout may get productivity gains.

The companies that redesign all three systems have a chance to change their economics and competitive position.

Years does not mean slow

Saying enterprise AI transformation will take years is not an argument for caution or delay.

It is the opposite.

If the transition is structural, starting with isolated pilots and waiting for certainty is risky. Organisations need to begin learning now which workflows can change, what knowledge architecture agents require, how authority should work and what their people need to become good at.

But they should also resist the fantasy that an enterprise with thousands of employees, customers, systems and obligations can become AI-native through a sequence of tool deployments.

The technology may move in months.

The operating model will take longer.

The strategic advantage will come from learning to change it while the business is still running.