AI in the back office: where the first real savings are showing up

1. Reading documents that arrive in a hundred formats

Invoices, claim forms, delivery notes and medical reports arrive as scans, photos and attachments in every layout imaginable. Before AI, a person keyed the important fields into a system. In the 22 teams that automated this, a model now reads the document, fills the fields and flags the ones it is unsure of. People check the flags, not the whole document.

44%Average reduction in handling time per document across the 22 teams that automated document reading.

2. Sending work to the right queue

Claims, supplier queries and customer emails need to reach the right team the first time. Misrouted work waits, then waits again. Teams that used a model to classify incoming work and suggest a queue cut misrouting from 14% to 3% of cases. The saving is not only the minutes of sorting, but the days a misrouted case used to lose.

3. Explaining why the numbers do not match

Reconciliation is mostly matching, which rules already do well. The expensive part is the breaks: the items that do not match and need a person to work out why. Finance teams that used a model to propose a likely explanation for each break, with the evidence, cleared breaks 2.3 times faster. The model is often wrong, but it is wrong in a way a person can see in seconds.

4. Drafting the routine reply

Every back office writes the same letters many times a day: a request for a missing document, a payment confirmation, a status update. Teams that let a model draft these replies from the case record, for a person to edit and send, saved an average of 5.5 hours a week per person. Nobody in the study let the model send without review.

What the 19 teams that saved nothing had in common

• They automated a decision, such as approving a payment or a claim, rather than the work around it. Of nine teams that tried this, none kept it running.
• They had no baseline, so they could not show a saving even where one may have existed.
• They chose the most visible task, not the most frequent one.
• The people doing the work were not involved in designing the tool, and quietly worked around it.

“We stopped asking what the model could decide and started asking what it could prepare. That is when the hours came back.”

What it costs

The median team spent the equivalent of five months of one person's salary to build its tool, and a modest monthly sum to run it. Payback came in seven months at the median. The fastest paybacks were in document reading, where volumes are high and the work is easy to measure. The slowest were in drafting, where savings are spread thinly across many people.