The 18-month AI loop

Between conversations at AWS re:Invent and the ridiculous amount of time I spend talking to industry experts (probably more time than I spend doing my actual job, shhhh! But hey, that’s “strategy”, right?), a clear pattern of “looping” has emerged.

Well, at least I think so. :)

Phase 1: The Committee Era

• Month 1: The mandate drops: “Put AI in everything.”
• Month 3: A few isolated wins appear, but with absolutely zero connection to actual business outcomes.
• Month 4: Committees are formed to centralise the brainstorming.
• Month 5: Shadow AI takes deep root because the committees are too slow. Naturally, the committees form sub-committees to govern the shadow AI.

Phase 2: The Tactical Treadmill

• Month 6: Prompt engineering is declared the saviour. It works, but it completely lacks scale.
• Month 8: Platform engineering and context engineering enter the chat. We build standards repositories that look beautiful but deliver very little impact to the organisation.
• Month 10: Spec-driven development takes centre stage. Half a dozen tools are tested, creating massive enterprise risk. Central tools are finally chosen, yielding a tiny trickle of usable data.
• Month 12: “AI-native workflows” and “agent-of-agents” become the new buzzwords. Everyone assumes the agents will just figure it out. Spoiler: they don’t. You still have to build.

Phase 3: The Reset

• Month 14: You step back and realise the entire software development life cycle has fundamentally changed. So, naturally, you spin up another committee.
• Month 15: You create AI strategies built around value streams. Suddenly, Anthropic drops a new feature that completely invalidates everything you just designed. You retreat to managed services.
• Month 16: In a moment of sheer existential clarity, you commit to bridging the old rivalry between Product and Engineering, a century-long battle for clarity. You realise that if product discovery and engineering execution aren’t deeply aligned, even the most advanced AI build-out is just going to hallucinate solutions to the wrong problems. This is a business crisis, not a localised one.

Month 18

The fatigue sets in and clarity finally arrives. You realise we’re trying to solve the exact same information and productivity dilemma we’ve been fighting for 10 years: working smarter, not harder.

AI is just a tool. It doesn’t fix broken workflows; it simply accelerates your existing processes. If you don’t deeply understand your process and how it drives a business outcome, your AI ROI is a myth, as elusive as a whisper.

The way out: the Work Pack

That brings us to the actual path forward: organisational AI and, more recently, harness engineering. In other words, impacting the value stream with AI in product outcomes.

It feels exactly like the early, gritty days of DevOps. It’s about building through the noise, defining excellence and measuring real value. How? The mechanism isn’t another brainstorming session. It’s the Work Pack.

At its core, a Work Pack is an executable spec, a contract, containing:

• intent
• context
• constraints

It acts as the active trigger for an agent team, which is an implementation detail if anything. We’re talking about end-to-end workflows that have a measurable impact on business outcomes and analytics. This isn’t a philosophy; it’s a measurable unit of value delivery.

When you shift to this model of org AI and harness engineering, the data finally starts showing up. The processes evaluate themselves. The noise drops, the value streams clear up, and finally, everyone wins.

Final note

If you’re entering or halfway through this loop, I wish you well. Define what’s valuable, and it’ll get you there faster. Or if you’d like to chat, I’ll talk you through it.