Everyone’s excited that AI can write code in seconds.
The frontier has shifted to agentic coding – AI that doesn’t just suggest code, but acts on it: runs commands, edits files, iterates autonomously.
Here’s what neither conversation is addressing: coding is often 20–25% of the work.
The hard part isn’t the code. It’s everything around it.
After three years building AI and data solutions for large enterprises, here’s what actually eats the time:
→ Reaching out to the right team to get system access → Submitting firewall opening requests (and waiting) → Obtaining roles and permissions — sometimes weeks of back-and-forth → Deploying into environments the AI has zero context about
The agent can write the building block. It cannot navigate the organisation surrounding it.
Why? Because your existing systems are completely opaque to the model. No context, no API definitions, no map. The agent gets further than the copilot did — and then hits the same wall, just later in the process.
This is especially true in analytics environments, where integration density is far higher than in operational systems. Operational systems often have internal logic and clean API definitions. Analytics pipelines are a web of dependencies — and that web isn’t documented anywhere an AI can see.
What this means for AI strategy:
Agentic coding doesn’t remove the bottleneck. It exposes it faster.
If you want AI to actually deliver speed, the constraint isn’t the model or the agent. It’s whether your systems are AI-friendly — meaning exposed, documented, and automatable.
Two paths forward:
- Build new systems AI-ready from day one: API-first design, infrastructure-as-code, policy-as-code for access, programmatic authentication, machine-readable metadata – so agents can discover and use your systems without human in the loop.
- Systematically make existing systems AI-legible — starting with the areas of highest impact, not trying to solve everything at once: mapping dependency graph, wrapping legacy systems with standardized interfaces and automating the approval workflows that block every AI agent today.
The organisations that figure this out in the next 18 months will have a structural advantage. The ones that don’t will keep wondering why their AI pilots don’t scale — regardless of which agent they’re running.
