From RPA to agents: automating the last mile
Deterministic automation handles the predictable 80%. The last mile — the judgement, the exceptions, the unstructured input — is where AI agents finally earn their keep.
Most automation programmes plateau in the same place. The rules-based work — moving data between systems, reconciling records, filling forms — gets automated quickly and pays for itself. Then the backlog fills with the cases that "almost" fit: the invoice in an unusual format, the customer email that spans three requests, the exception that needs a human to read, decide, and act. That last mile is where value leaks, and where teams quietly give up and hire more people.
Where RPA stops
Robotic process automation is excellent at deterministic, high-volume tasks with stable inputs. It is brittle precisely where the real world is messy: unstructured documents, ambiguous intent, decisions that depend on context rather than a lookup table. Bolt enough exception-handling rules onto an RPA bot and you get a system nobody can safely change — the maintenance cost overtakes the saving.
What agents add
An AI agent brings three capabilities RPA lacks: it can read unstructured input and extract meaning, it can reason over a goal instead of following a fixed script, and it can call tools — including your existing RPA bots and APIs — to act. The useful mental model is not "replace RPA with agents" but "let agents handle the judgement, and let deterministic automation do the mechanical work underneath."
- Read and classify: turn an email, PDF or chat message into structured intent.
- Decide: apply policy and context to choose the next action, and explain why.
- Act: invoke the RPA bot, API or workflow that actually does the task.
- Escalate: recognise low confidence and hand off to a person with the context attached.
The last mile, in practice
Take invoice processing. RPA already keys the clean invoices. The exceptions — mismatched line items, missing PO numbers, non-standard vendors — pile up in a queue. An agent reads each exception, checks it against policy and the ERP, resolves the ones it can with high confidence, and routes the genuinely ambiguous ones to a human with a one-line summary and a recommendation. Throughput on the queue that used to need people goes up; the humans spend their time only on cases that truly need judgement.
A pragmatic path
You do not need to rebuild your automation estate. Start where the exception queue is most expensive, wrap the judgement step in an agent, and keep the deterministic plumbing you already trust. Measure the confident-resolution rate and the escalation quality before you widen scope.
Getting it into production safely
An agent that acts on your systems needs the same discipline as any production software, plus a few AI-specific controls: tight tool permissions so it can only do what it should, evaluation on real historical cases before go-live, confidence thresholds that fail safe to a human, and logging of every decision so the behaviour is auditable. Done this way, agents are not a science project — they are the piece that finally closes the last mile.