What is RPA?
Robotic process automation is software that executes a fixed sequence of steps in application interfaces: open this screen, copy this field, paste it there, click submit. It is a recording of what a person would click, replayed at scale, and it works well when every input looks the same and every interface stays the same.
Its weakness is variation. An unexpected field, a changed layout, or an input that needs interpretation stops the script, and often stops it silently, with the failure discovered later.
What does a Digital FTE do differently?
A Digital FTE reads its inputs, including unstructured ones such as an invoice, an email, or an application, decides what to do within its written remit, and proposes the action with the evidence attached. When a case falls outside the remit or the evidence is insufficient, it escalates through the approval gate rather than guessing or stopping.
A Finance Controller Digital FTE matching a payout to two invoices net of fees is doing something RPA cannot: interpreting a difference, finding the fee schedule that explains it, and proposing the match with all three documents attached.
How do they handle failure?
They handle failure oppositely. RPA fails when the world changes and typically fails without explanation; someone notices later that the queue stopped. A Digital FTE surfaces the case it cannot handle as an exception with the reason, in the morning brief, so the failure is a visible item rather than a silent stop.
This is also the governance difference. RPA has no concept of evidence, approval, or an audit trail of decisions, because it makes none. A Digital FTE is built on all three.
- Input: structured and identical (RPA) versus varied and unstructured (FTE)
- Change: breaks silently versus raises the exception
- Judgement: none versus within the remit, with escalation
- Evidence and approval: absent versus on every proposal
- Cost: licence plus maintenance versus metered work
Where does RPA still win?
RPA still wins for high-volume, rule-bound steps in stable systems where there is genuinely no judgement: moving a field between two systems that have no integration, at thousands of records an hour, where the interfaces do not change. Deterministic execution is cheaper and more predictable than any model for that.
It is the wrong tool the moment interpretation appears, and the maintenance cost of keeping scripts aligned with changing interfaces is the cost most RPA programmes underestimate.
How are they used together?
They are used together by giving the mechanical step to RPA and the surrounding work to a Digital FTE. The FTE reads and decides; where a decision results in a fixed action in a system with no API, an RPA script can execute it. The FTE's proposal and the manager's approval are in the audit trail; the script does the clicking.
Where a system does have an API, which most modern tools do, the Digital FTE connects directly and RPA is unnecessary. Most teams find that the RPA estate shrinks to the few legacy systems that need it.
How do you migrate a process from RPA to a Digital FTE safely?
Migrate one process at a time, and keep the RPA script running until the record justifies switching. Record the script's last month as the baseline: runs, failures, the time spent fixing failures, and the interface changes that caused them. Connect the Digital FTE to the same systems through their APIs where they exist, with scoped credentials, and write its remit to cover the process plus the exceptions the script could never handle. Run it at Level 0 for two cycles alongside the script, with every proposal gated.
Compare on the same measures, then on what the script could not do: exceptions surfaced with reasons rather than silent stops, proposals with evidence rather than blind execution, and an audit trail with a named approver. Where a legacy system has no API, keep a minimal script for the mechanical click and let the FTE's approved proposal trigger it, so the decision is in the trail and the script only executes. Retire the rest of the script when the FTE's approval record for the process is long and clean, and expect the RPA estate to shrink to the few systems that genuinely need it.