Putting Claude to work: a practical playbook for AI in operations
The gap between “we tried ChatGPT” and “AI is part of how we operate” is wide, and most organisations are stuck in it. The demos are impressive; the day-to-day integration is where it gets hard. Here’s the playbook I keep coming back to for putting a model like Claude to actual work.
Start with a process, not a prompt
The common mistake is to start from the tool — “what can the AI do?” — and look for places to use it. That produces novelty, not value. Start from the other end: pick a real process that’s slow, repetitive or judgment-light, and ask where a capable assistant would remove friction.
Good early candidates tend to share a shape: high volume, well-defined inputs, and a human still in the loop for the final call. Drafting first versions, summarising long threads, triaging requests, turning messy notes into structured output. Unglamorous, frequent, and forgiving of imperfection.
The rules that keep it useful
A few principles have saved us from the typical failure modes:
- Keep a human accountable. AI drafts, a person decides. The moment the output touches money, customers or compliance, ownership stays human.
- Give it real context. A model is only as good as what it can see. Most disappointing results come from thin prompts, not a weak model. Feed it the documents, the examples, the constraints.
- Design for being wrong. Assume some output will be off and build the check into the process, rather than trusting blindly and discovering errors downstream.
- Measure the boring metric. Not “is this impressive?” but “did this save real time without adding risk?” If you can’t point to that, it’s a demo, not a process.
The capability is moving fast
What’s changed recently isn’t just that the models are smarter — it’s that they’ve become genuinely agentic. They can work through multi-step tasks, use tools, hold long context, and operate over a whole body of material rather than a single question. That shifts the design space from “answer my question” to “handle this piece of work,” which is exactly where the operational value is.
The real bottleneck
It’s rarely the technology. It’s whether your processes are clear enough to hand off in the first place. AI is a forcing function: to delegate work to a model, you have to articulate what the work actually is — and a lot of teams discover they never had. That clarity is worth as much as the automation.