We've spent the last 18 months building production automations with Claude. Here's an honest write-up of what works, what doesn't, and the workflows we'd still tell you to do by hand.
What works exceptionally well
- Document understanding. Claude can read a 50-page contract, extract every clause, classify them, and pull the riskiest ones to the top. We've replaced two full-time legal-ops roles with this workflow.
- Drafting from templates. RFP responses, customer success follow-ups, support replies: anything where the structure is fixed and the variables come from your CRM. 5× faster, same quality.
- Code triage. Reading a bug report and pulling the likely files / functions to investigate. We use this internally and it saves engineers 30 minutes per bug.
What works with caveats
Multi-step tool use. Claude can plan a 5-step workflow, call APIs, branch on results, and finish. But you need rigorous evals, retries, and a human-in-the-loop checkpoint for any irreversible action. Don't let an agent send emails, pay invoices, or modify databases without explicit approval.
What we'd still do by hand
Anything where the cost of a single mistake is high and the work isn't repetitive. Pricing decisions, hiring evaluations, anything legally binding. The right play is augmentation, not delegation.
How we structure agent projects
Our standard pattern:
- Eval-first. Build 30-50 example inputs with known good outputs before writing prompts.
- Tools, not chains. Give the model real tools (read database, draft email, etc.) and let it plan. Don't hardcode the chain.
- Observability. Every action gets logged. Every failure gets reviewed. The agent improves through your eval set, not vibes.
- Cost ceilings. Per-task budget enforced at the orchestration layer. No "oops we spent $5k" mornings.
The biggest mistake we see: teams treating Claude as a search engine instead of a teammate. The wins come when you give it a real job to do, with the tools and the budget to do it.