The question most organisations ask when they start thinking about automation is: what can we automate? The better question is: what should we automate first?
The distinction matters because the wrong starting point creates the wrong impression. An automation that handles edge cases poorly, that needs constant intervention, or that creates more work in maintaining it — that sets back the broader programme.
Four characteristics of automatable tasks
In every deployment we've run, the tasks that automate well share these properties.
1. Rule-based at their core
The task follows a decision tree that can be fully articulated. "If the order is within 14 days and the item is unused, process the return" is automatable. "Use your judgement about whether this customer deserves an exception" is not — at least not initially.
This doesn't mean the task is simple. A rule-based process can have dozens of branches. What matters is that the logic can be written down completely.
2. High volume, low variance
The task happens frequently and the majority of instances are similar. A customer service team answering 200 order status queries per week is a better automation candidate than an accounts team processing 5 complex invoices per month — even if the invoices are harder.
Volume justifies the investment. Consistency determines whether the automation stays useful over time.
3. Clear inputs and outputs
The task has defined starting points and defined endpoints. "Customer submits a return request → return is initiated and customer receives confirmation" is clean. "Do something about the inbox" is not.
When inputs and outputs are clear, you can measure whether the automation is working. When they're vague, you can't.
4. Measurable
You should be able to tell, within two weeks of deploying a digital worker, whether it's working. That means the success condition needs to be defined before deployment, not after.
Resolution rate, first-contact resolution, time to close, human escalation rate — pick one primary metric and baseline it before the deployment starts.
Building your task inventory
A practical way to identify candidates: ask your team to spend one week logging every task they do that they wish someone else would handle. Not tasks they dislike — tasks they find routine enough to imagine a system doing.
Then sort by: frequency × time per instance. The top ten items on that list contain your best automation candidates.
From those ten, filter by the four characteristics above. You'll typically find three to five tasks that are ready now, two or three that could be ready within a few months, and a few that aren't worth automating.
A note on 80%
When we say 80% automation, we don't mean automating 80% of your workforce. We mean that in most task categories, 80% of the instances follow predictable patterns that a digital worker handles reliably.
The remaining 20% — the exceptions, the edge cases, the genuinely complex situations — stay with the human team. The digital worker flags them, routes them, and provides context. The human resolves them.
That's the model. Not replacement — augmentation, with a clear division of what belongs where.
Starting points by department
| Department | High-value starting tasks |
|---|---|
| Customer Service | Order status, FAQs, return initiation |
| Operations | Supplier communications, inventory alerts, shipment tracking |
| Finance | Invoice matching, expense categorisation, payment reminders |
| Marketing | Social monitoring, report compilation, campaign asset checks |
| Sales | Lead enrichment, follow-up sequences, CRM data entry |
Pick one task from one department. Get it working. Build from there.