Most AI projects go wrong at the first decision: what shape the system should take. The same business goal — "handle supplier invoices faster" — can be met by a rules-based workflow, a copilot that drafts for a person, or an agent that acts on its own within limits. Each has a different cost, risk and time to value.
The short answer
- Conventional workflow automation when the steps are known and the inputs are structured. It is cheaper, faster and easier to test than anything with a model in it.
- A copilot when the output needs judgement but a person should approve every result — drafting replies, summarising cases, suggesting codes.
- An agent when the task needs judgement and action across several tools, and you can write down what it may and may not do.
A decision framework
Ask four questions about the task, in order.
- Are the steps the same every time? If yes, start with workflow automation. Add a model only for the one step that needs to read unstructured input.
- Does every output need a person's sign-off? If yes, build a copilot. The person stays the decision-maker and the AI removes the preparation work.
- Can you list the tools and the limits? An agent is only safe when its permissions, approval points and failure behaviour are explicit. If you cannot write them down, you are not ready for an agent.
- Can you measure success per task? Agents need an evaluation set of real tasks with known correct outcomes. Without one, you cannot tell whether a change made it better or worse.
A worked example: supplier invoices
| Approach | What it does | When it fits |
|---|---|---|
| Workflow | Extracts fields from a fixed template and posts them | One or two suppliers with stable formats |
| Copilot | Reads any layout, proposes the posting, a clerk approves | Many formats, low tolerance for errors |
| Agent | Reads, checks the purchase order, chases missing data by email, posts within limits | High volume, clear rules for what can be posted unreviewed |
Many teams move through all three: start with a copilot, measure where people accept the suggestion unchanged, and let an agent take over only those cases.
Prerequisites for any of the three
- Access to the systems involved, with a service identity rather than a shared login
- A baseline: how long the task takes today and how often it goes wrong
- A small set of real examples, including the awkward ones
- Someone in the business who owns the outcome
Where to go next
If the task is mostly known steps, see Business Process Automation. If it needs judgement and action, see AI Agents.