"AI agents" and "workflow automation" get used as if they mean the same thing. They don't, and choosing the wrong one wastes money and erodes trust in the whole idea. Here's how to tell which your process needs.
Workflow automation: rules that run on their own
Workflow automation moves data and tasks through a process using rules you define. When a deal closes in the CRM, generate the contract. If an invoice is over £5,000, route it to a partner. Three days before a deadline, send a reminder.
It's reliable, auditable and cheap to run. It's the right answer when:
- the steps are the same every time
- inputs arrive in a known format: forms, system records, structured documents
- you need a clear audit trail
- the real problem is that your systems don't talk to each other
Most back-office admin in professional services is exactly this. A consultancy's monthly invoicing run, which used to take 8 days, now takes about 4 hours, almost entirely through well-built workflows connecting HubSpot, Xero, Signable and SharePoint (case study).
AI: for inputs that don't follow rules
Large language models are good at the things rules are bad at: reading free text, understanding what someone is asking for, pulling details out of documents that don't follow a template, and drafting something sensible from messy information.
AI is the right answer when:
- inputs vary: emails, attachments, notes, websites
- something has to be understood before it can be routed or actioned
- the work is too varied for fixed rules but too routine for your most expensive people
The answer is usually both
Real operational processes almost always combine the two. A typical pattern:
- AI reads and classifies the input. What is this email? Which client? What's being asked? Roughly what will it cost?
- Workflow automation acts on it. It creates the record, routes the request, updates the systems and sends the confirmations.
- A person approves at defined checkpoints: anything client-facing, financial or uncertain.
That's the pattern behind a fund data provider's request handling, where AI now triages and costs every incoming email and standard requests flow straight into automated fulfilment. Turnaround fell from 6 weeks to 4 days (case study).
It's also how an investment group's deal research works. AI gathers and summarises research on each target and drafts outreach, workflows update the trackers, and nothing is sent without someone on the team approving it (case study).
Common mistakes
- Using AI where a rule would do. Matching an invoice to a contract doesn't need a language model. It needs a lookup.
- Using rules where understanding is needed. A decision tree for free-text emails breaks on the first unusual message.
- Skipping the human checkpoint. Both approaches need oversight for anything that leaves the building.
- Starting with an "autonomous agent". Prove one well-defined process first, then extend.
A one-question test
Ask: "Does this step follow fixed rules, or does someone need to understand something first?"
Fixed rules point to workflow automation. Understanding first points to AI, with a person reviewing. If the process has both kinds of step, which most do, combine them.
If you'd like help mapping your own processes this way, that's what the free Operations Audit is for.