The question that decides whether it makes sense
The back office is where an agent’s return is easiest to defend, because the manual process is already timed: you know how many invoices come in each month and how long it takes one person to process each one. The question isn’t whether the agent can do it, but how much of the process it can close without anyone touching it — and what happens to the rest.
An agent that processes 70% of invoices and leaves 30% in a queue doesn’t save 70% of the time. It saves less, because someone still has to work the queue and check that the automated part came out right. That nuance decides whether the project is viable.
The processes with the most potential
Supplier invoices. The agent reads the document in any format, extracts the issuer, amounts, taxes, and line items, matches them against the purchase order and delivery note, and posts the invoice to the ERP. Clean cases go through on their own; amount or quantity mismatches are held back, half-resolved, with the problem flagged.
Bank reconciliation. The agent matches transactions to invoices and payments, including grouped and partial ones, and proposes the journal entry. Anything that doesn’t match gets escalated with the reason why — which is what really saves time.
Orders and data entry. Orders that arrive by email in different formats, contracts that need to be keyed in, new supplier setups. The agent structures the data, validates it against your company’s rules, and records it.
Case files and administrative responses. The agent gathers the documentation for a case, checks it’s complete against the checklist, and drafts the response with the data already verified.
Where the human signs off
This is the most important design decision, and it isn’t a technical one. A split that works: the agent records and prepares, but doesn’t pay, sign, or communicate externally without approval. Above a set amount, everything stops. New suppliers always stop. And any action with an irreversible accounting effect is logged along with who approved it.
This is the territory of the harness: permissions, limits, and audit trails. Without it, the project shouldn’t reach production.
The data and systems you need
Programmatic access to the ERP (read access plus scoped write access), a history of already-processed documents to build the evaluation set, and written business rules. That last part is usually the surprise: much of the judgment your admin team applies isn’t documented anywhere, and it has to be drawn out in interviews before anything can be automated.
How we measure it
The share of documents processed end to end without intervention, average time per document including review, the error rate found through sampling, and the cost per processed document. We compare all four against the current process, measured the same way.
It’s also worth measuring cycle time: in purchasing, closing the invoice cycle sooner improves supplier relationships and unlocks early-payment discounts that don’t show up in an hours-saved calculation.
Where it fails
The classic mistake is to automate the easy case and declare victory: the clean 60% of invoices go through on their own, and the team is just as swamped as before, because their time was going into the messy 40%. Before building anything, measure where the time really goes. Often the answer isn’t processing at all — it’s chasing whoever has to approve.