Generative AI in the Back Office: 5 Tasks That Still Need Human Oversight

Generative AI reached the SME back office through a side door. Nobody signed off on it in a budget meeting: someone started using it and the habit spread. Eurostat figures show the scale — in 2025, 20.0% of EU enterprises with 10 or more employees used AI technologies, up from 13.5% a year earlier. In Portugal the share is 11.5%, and among those companies the most common application is organising business administration and management processes. Not marketing. The back office.

The useful question is no longer whether generative AI works in the back office. It does. The question is where it can decide on its own, and where a person still has to put their name to the output.

Why 99% accuracy is not enough

A model that gets 99% of fields right sounds excellent. On an invoice with 20 fields, it is not.

The arithmetic fits on a napkin. If each field has a 99% chance of being correct, and assuming errors are independent, the chance the whole document is right is 0.99 to the power of 20 — roughly 82%. Almost one invoice in five carries at least one wrong field.

That is an estimate, and the independence assumption does not hold in practice: errors cluster in the same badly scanned documents. It fixes the order of magnitude rather than promising a rate. The point still stands, and it is what most vendor demos quietly skip: the accuracy advertised is per field, and the accuracy your accounts care about is per document.

There is a second, nastier effect. A generative AI error is plausible. A tired person types 1,250 instead of 1,520, and anyone who knows the supplier spots it. A model writes a figure that fits the context, formatted correctly, in the right field. It sails through a distracted review.

1. Posting supplier invoices and documents

This is where generative AI pays back fastest, and also where the gap between the demo and production is widest — something we covered when comparing document extraction accuracy in benchmarks against production.

What works: invoice number, date, tax number, supplier name. Short, unique fields.

What breaks: composite fields. Line items, several VAT rates on the same document, withholdings, discounts applied in the footer. Precisely the fields that determine what you pay.

Where the human sits: not re-reading everything. Re-reading what the system could not close on its own. Consistency rules the model cannot talk its way around — line items must sum to the total, the tax number must exist, the amount must match the purchase order. If a rule fails, the document stops and joins a human queue. If none fails, it goes through untouched.

2. Replying to customers and suppliers by email

Generative AI writes better emails than the average message leaving an inbox at half past six. The risk is not the prose. It is the commitment.

A back-office email commits the company. It confirms a deadline, accepts a condition, validates a price. And a generative model is optimistic by design: it was trained to be helpful, not cautious. When it does not know whether Thursday delivery is possible, it tends to say yes.

Where the human sits: the AI drafts, the person sends. A ready draft sitting in the outbox saves almost all the time a manual email costs, and keeps the decision on the right side of the line.

The reasonable exception is purely informational replies with no commitment attached — order status, opening hours, address, tracking number. Those can go out unattended, on one condition: the data comes from the system, not from the model's memory.

3. Summarising meetings, calls and long threads

Public benchmarks help calibrate expectations here. Vectara's hallucination leaderboard, which measures summaries of documents handed to the model, puts the best result near 1.8% and most leading models between 3% and 7%. And that is a test where the model has the source in front of it — it is not answering from memory.

Three per cent sounds residual until you translate it. In a meeting with 40 decisions and deadlines, one or two come out subtly wrong. And the summary is what survives: nobody goes back to the 50-minute recording to check.

Where the human sits: whoever attended the meeting signs off the action list before it becomes tasks. And the workflow always keeps the original transcript, with every point in the summary traceable to the moment it was said. A summary without a source is a well-formatted rumour.

4. Triaging and routing requests

Classifying emails, applications, support tickets or invoices by cost centre. It is the dullest of the five tasks and, for that exact reason, the most dangerous.

A routing error does not produce wrong text that someone reads and corrects. It produces silence. The request went to the wrong queue, nobody noticed, and the problem surfaces three weeks later with a customer on the phone.

Where the human sits: not approving every classification, which would defeat the purpose. Measuring. A sample reviewed every week, with the error rate tracked over time, and a mandatory "unsure" category in the model. A classifier forced to choose always chooses — including when it should not.

5. Producing documents that carry the company name

Proposals, reports, opinions, tender responses. Text that represents the company to outsiders.

The reference case is recent and does not involve an unprepared small business. In October 2025, Deloitte agreed to refund part of a contract worth around AUD 440,000 to the Australian government after a delivered report was found to contain non-existent academic references and a fabricated quote from a court judgment. The firm confirmed it had used a language model to produce the document.

If that happens at a consultancy with formal review processes, the question for a 30-person company is not whether it can happen.

Where the human sits: one simple rule. If the document carries the logo and leaves the building, a named person is attached to it before it goes. And whoever signs has checked the figures and the sources — not the style, which is the part AI already handles well.

The legal deadline moved. The arithmetic did not.

Plenty of decision-makers exhaled in July. Regulation (EU) 2026/1744, the Digital Omnibus, was published in the Official Journal on 24 July 2026 and pushed the AI Act obligations for standalone Annex III high-risk systems from 2 August 2026 to 2 December 2027.

Two notes before filing the topic away.

First: most of an SME back office was never high-risk under the regulation anyway. Posting invoices is not in Annex III. Screening job applications is. It is worth knowing which side each workflow falls on before assuming none of them count.

Second, and more useful: Article 14 of the AI Act works as a checklist even for companies with no obligation to follow it. It requires that the responsible person can correctly interpret the system's output, decide not to use it, disregard, override or reverse it, and interrupt the operation altogether. And it names a risk that rarely appears in vendor decks: automation bias, the tendency to automatically rely, or over-rely, on what the system produces.

Automation bias is the hidden cost behind everything above. A team reviewing AI-processed invoices that finds no error in the first fifty stops reviewing seriously somewhere around the sixtieth. Human oversight that exists on paper but not in practice is worse than none at all, because it is reassuring.

How to build this without stalling the workflow

Human oversight does not have to mean an approval meeting. On an automation platform like n8n, it is a node in the middle of the flow.

The Chat node documentation describes a "Send and Wait for Response" operation that suspends execution until someone answers. There are two response types — free text, or approval with buttons — plus a "Limit Wait Time" option that resumes the workflow on its own after an interval or at a set time. Equivalent operations exist on the Slack, Gmail, Teams, Telegram and WhatsApp nodes, which means the question arrives where the person already is, rather than in an app they open once a day.

The practical design looks like this: the workflow processes everything, applies the consistency rules, and only interrupts when a rule fails or a value crosses a threshold the company has set. An 80-euro invoice that reconciles goes through alone. An 8,000-euro one that does not match the purchase order waits for a click.

Setting that threshold is a management decision, not a technical one. It depends on knowing what each error costs — the same calculation you run to work out how much a manual process actually costs.

The question that decides

The question is not which back-office tasks can be automated with generative AI. That list is long, and there are obvious candidates in almost every company.

The question is a different one, and it belongs to you rather than to a vendor: what error rate can this process absorb, and how would I know I had crossed it?

Answer both and you can put generative AI behind nearly everything in this article. Fail to answer the second and the problem predates the AI — you were not measuring the process before it arrived.