How to Automate Workflows with AI

The best first automation is repetitive, high-volume, and has a forgiving error budget. Aim for workflows that hit all three:

• Runs at least 10× per week.
• Currently costs a human 5+ minutes each time.
• A wrong answer can be caught and corrected — not pushed to a customer's bank account.

Good first picks: inbox triage, lead enrichment, meeting note distribution, support tier-1 deflection.

Bad first picks: outbound sales emails, code deploys, anything billing-related.

Get the actual humans on a call. Walk the workflow step by step. Write down:

• The trigger (what tells someone to start?).
• Every manual step in order.
• Every tool touched (CRM, inbox, sheet, etc.).
• The exception cases — "if X, then we ping Y on Slack".
• The handoff (where does the work end?).

If you can't draw it on one page, you're not ready to automate it.

Most steps don't need an LLM. Mark each step as one of:

• Deterministic: fetch, write, transform — use plain workflow nodes.
• Needs reading: understand free-text email, parse a PDF — use an LLM.
• Needs judgment: classify, prioritise, route — use an LLM with a short prompt.

Fewer LLM calls = cheaper, faster, more reliable. Use AI only where it earns its keep.

You need three layers and that's it:

• Orchestrator: n8n, Make, or Zapier.
• Brain: OpenAI, Anthropic, or Gemini.
• System of record: whatever CRM/helpdesk/database you already use.

Resist adding vector DBs, fine-tuned models, or new SaaS until the v1 has proven it's worth the complexity. Most workflows never need them.

Don't deploy autonomous on day one. Three phases:

1. Shadow: the automation runs and logs what it would have done; a person still does the work. Compare.
2. Suggest: the automation drafts; a person approves with one click.
3. Autonomous: safe paths run end-to-end; risky cases escalate to a human with full context.

Most clients are at "autonomous on safe paths" within two weeks.

Track three numbers from week one:

• Time saved per week (hours × hourly cost = dollars).
• Accuracy on the steps the agent touches.
• Escalation rate — what % of runs need a human.

If accuracy drops below your threshold, tighten the prompt, add an example, or carve off the failure cases as deterministic exceptions. Don't expect 100% — aim for "better than the human on average".