The best AI strategy is an iterative one.

There is an interesting contradiction in how many companies are approaching AI.

The technology is changing almost weekly, yet organisations are trying to create three-year AI strategies. By the time the strategy is approved, some of the assumptions behind it may already be wrong.

Sam Altman recently made an interesting point about this from a different perspective. He defended OpenAI’s approach of deploying AI progressively, learning from how people actually use it, improving the technology and its safeguards, then deploying again.

It is partly born out of necessity. You cannot predict everything a complex AI system will do before it meets the real world.

I think the same principle applies to enterprise AI adoption.

The best AI strategy may be an iterative one.

You cannot workshop your way to AI adoption

There is nothing wrong with having an AI strategy. Companies need to understand where AI can create value, what risks they are prepared to accept and where they should invest.

The problem starts when strategy becomes a substitute for doing.

Organisations can spend months identifying use cases, scoring them, developing governance frameworks and building roadmaps before anything reaches a real user. Meanwhile, the technology changes.

There is another problem. Until AI meets a real business process, you do not actually know enough.

AI demonstrations are easy to make impressive. Real businesses are messy. Customers ask unexpected questions. Data is incomplete. People do not follow processes exactly as documented. Systems fail. Edge cases appear. Security restrictions get in the way.

You discover these things by putting AI into a real workflow.

Start narrow, but make it real

The answer is not to deploy AI everywhere and see what happens. Start with a narrow problem that has a measurable outcome.

Instead of building an AI agent that “handles customer service”, start with one type of enquiry. Instead of automating an entire claims process, start with document intake or one decision point. Instead of asking an engineering agent to autonomously deliver software, let it implement bounded tasks with automated tests and human approval.

Starting narrow does not mean choosing something inconsequential. The problem should be important enough that people care whether you solve it.

The difference is that you are reducing the surface area of uncertainty while still learning from the real world.

Deploy, measure, learn, expand

The model is simple:

Deploy → Measure → Learn → Expand

Deploy something useful into a controlled environment. Measure it against the process it is improving. Establish the baseline before you start. Otherwise almost any AI implementation can be made to look successful.

Learn from what happens. Where did people intervene? What mistakes did the AI make? What did it cost? What exceptions had you not considered? Did customers or employees actually prefer the new process?

Then expand. Give the AI more types of work. Connect another system. Reduce human intervention. Increase its autonomy. Each expansion should be supported by evidence from the previous one.

Autonomy should be earned

This becomes particularly important as we move from AI assistants that provide information to AI agents that take action.

An AI that drafts an email is different from one that sends it. An AI that recommends a code change is different from one that deploys it.

The sensible response is neither unlimited autonomy nor preventing agents from taking action. It is progressive autonomy.

Start with the AI making recommendations. Then allow it to act with human approval. Once you understand its behaviour, allow low-risk actions automatically while escalating exceptions. Increase the boundaries when the evidence supports it.

This is also where testing, evals, monitoring and governance become part of AI delivery rather than something added at the end.

Build a strategy designed to change

There is a temptation to wait for AI to mature before making serious decisions. That is unlikely to work.

Models will improve. Costs will fall. New tools will appear. Some of today’s architectural choices may look unnecessary in 12 months. A good AI strategy should assume that.

Do not bet everything on predicting which model, platform or agent framework will win. Build the organisational and technical capability to try things, measure them and change direction.

The companies that get the most value from AI may not be the ones that predict the future most accurately. They may simply be the ones that learn fastest.

That is how we think about AI adoption at Revity:

Rethink what is possible. Build something real. Measure what happens. Scale what works.

Then do it again.