Drawing on several years of large-scale applied work, these eight working rules lay out an approach to implementing AI that focuses on practices that will create sustained use and growth.

Rule 1: you talk about AI Club

There is always the instinct to hoard information or wait until things are polished, but this always produces organizational lag. Agentic AI-based work is changing so quickly that learning has to happen laterally — between practitioners, between teams, between companies — rather than top-down through documentation or workshops. 

Mature organizations even build this into the basic operating cadence, with every formal review process explicitly covering what individuals are doing with AI.

Rule 2: you really talk about AI Club

Yes, this repetition is deliberate. AI is not a one-time directive or a campaign topic. It is a cultural setting that has to be reinforced continually — in every team, every meeting, every project. If AI usage does not come up regularly across the business, then it is not part of how the business works

Rule 3: know when to tap out

Not every use case is winnable, and recognising which kind of fight you are in matters more than persistence. There are three reasons why you may want to tap out of the fight: 

1. The models are not yet ready

Some workflows — particularly complex, multi-step ones requiring high accuracy — are beyond what current models can deliver. Setting an accuracy threshold (60% is a common one) and tabling the use case if it cannot be met avoids being on the fence about tapping out. The reality is that model capabilities are improving extremely quickly, so tapping out does not necessarily mean giving up forever — you might find you can return to the idea in a few months, when the technology has caught up with your ambition. 

2. You are in the wrong weight class.

Agentic work falls into roughly three categories of complexity. Deterministic workflows, where the same input must produce the same output every time, reasoning-based workflows that pull from documentation, and multi-agent, multi-knowledge-base workflows that span an entire function. Each is a different weight class and requires a different set-up. To attempt the third with tooling designed for the first is to lose before the fight even starts. 

3. The workflow itself is poor. 

No amount of agentic capability fixes a process that did not work in the first place. AI applied to a broken motion just makes the broken motion faster.

Rule 4: one initiative needs one owner

In a fight, there are only two parties. Committees and cross-functional working groups assembled to debate AI use cases tend to produce roadmaps; individual practitioners, however, with deep expertise, the right tools and the right data, produce shipped capabilities. 

A useful operational pattern is creating dedicated time — a structured day each week, for instance — during which practitioners across the business get access to the latest internal knowledge sources and AI tooling, with no specific assignment beyond: ‘do your job, but with these new tools.’

For example, in one engineering organisation, a single support engineer used one afternoon of unstructured time to automate the resolution of an entire category of ticket. The category had previously taken days to resolve per case, involved correlation across multiple systems, and consistently produced customer frustration. This engineer built the rules, skills, and logic for the system to reach the answer interactively with the customer, eliminating the need to file a ticket at all. After product validation, the ticket category ceased to exist. This was the result of one practitioner having the right knowledge, data, and the ability to take ownership. 

Rule 5: build from the bottom up.

Purely top-down AI strategy tends to fail. In almost all circumstances, the people closest to a problem understand it better than the people authorising the budget. Leadership’s job should be to remove blockers and enable their teams, giving them access to data, useful tools and time to experiment, rather than to prescribe use cases.

Rule 6: failures are inevitable.

Organisations that succeed at applied AI accumulate more failed experiments than successful ones. Trying many things and learning which ones work is the only way to build the operational intuition that separates organisations that ship AI from those that merely talk about it.

Rule 7: no spectators.

Watching talks, reading newsletters, listening to podcasts, and sitting in conferences is not equivalent to the practice of doing the work. It can feel equivalent, however, which is part of the danger. The only way to develop the judgment needed to lead in this environment is to be in the work daily.

Rule 8: the only way out is through. 

While it might be a tempting option, waiting will not make things become easier. In fact, as models improve and tools proliferate, the gap between organisations utilising agentic AI and those hesitating will only grow.

Applied AI inside an organisation resembles an apprenticed craft more than a transformation project. Thinking about it in terms of a ‘transformation project’ introduces ideas like committees and roadmaps. But framing it in terms of an ‘apprenticeship’ evokes ideas like ownership, daily repetition and refinement.

This article is based on a talk given at Planhat Open 2026. Certain sections have been modified for editorial clarity.