Where the administrative hours actually go
Asana's Anatomy of Work Index, based on a survey of more than 13,000 professionals, landed on an uncomfortable figure: 60% of working time goes to work about work.
Not the job itself. Chasing information, updating statuses, copying data between systems, confirming things already confirmed.
The same study breaks down the average year: 103 hours in unnecessary meetings, 209 hours of duplicated work, 352 hours talking about work rather than doing it.
In a ten-person company, that is more than a full-time role spent on internal friction.
What AI genuinely handles well
AI is good at three things that dominate administrative work: reading unstructured documents, classifying them, and drafting repetitive text.
An invoice arriving as a PDF attachment no longer needs a human to read it. Fields are extracted and posted to the ERP. A support request is classified and routed without manual triage. A proposal is drafted from data already sitting in the CRM.
None of this is new as an idea. What changed is the cost of making it work. Integration projects that took months now take weeks, a shift we covered in how SMEs increase productivity without growing the team.
The finding that undercuts the hype
Here is the number that rarely makes it into a sales deck.
Generative AI at Work, by Brynjolfsson, Li and Raymond, tracked 5,179 customer support staff given access to an AI assistant. Average productivity gain: 14%.
But the average hides the point. The least experienced staff gained 34%. The most experienced gained close to nothing.
The correct reading is not "AI lifts productivity by 14%". It is that AI transfers what experienced people know to people who do not know it yet. It shortens the learning curve.
For an SME this has a blunt consequence. If your admin team is senior and knows the processes cold, AI will return less than promised. If you have high turnover or people still learning, it will return considerably more.
Where accuracy breaks, and why it matters
Vendors quote document extraction rates of 97% to 99%. Those numbers are real, but measured under controlled conditions.
In production, documents arrive scanned at an angle, with stamps over the text, in formats that differ by supplier. Research on document extraction shows a consistent gap between benchmark and field — in one reviewed clinical study, the best model dropped from 99% on synthetic data to 87.7% on real data.
Simple fields, such as an invoice total, stay near 99%. Line items and split VAT rates land around 95% to 97%.
A 95% rate sounds excellent. Across 200 invoices a month, that is 10 errors. If nobody catches them, they reach your accounts.
The calculation that changes the maths
It is worth doing the sums with supervision included rather than assumed away.
A company processing 200 invoices a month, at 6 minutes each, spends 20 hours monthly on that process alone.
With automated extraction at a realistic 90% accuracy — not the brochure figure — 180 invoices pass untouched. The remaining 20 need human review at 2 minutes each: roughly 40 minutes. Add spot-checking the ones that passed, and you land near 3 hours.
Savings come to about 17 hours a month. Roughly 200 hours a year, from one process.
That is less than "eliminates manual work". It is more than enough to justify the investment. And it holds up when somebody challenges it.
Human review is not a system failure
The conclusion is not that AI falls short. It is that designing the process as though it never errs is what makes these projects fail.
Mature platforms already assume this. The n8n AI agent node can require human approval before executing sensitive actions — sending a message, modifying a record, deleting data. Automation runs unattended right up to the point where a mistake costs money, then waits for a person.
For anyone handling customer data in Europe, there is also the question of where the models run. n8n can be self-hosted, keeping data inside your own infrastructure — something we went into when explaining what n8n is and how its licence works.
Where to start without burning months
Pick a process that is frequent, dull, and rule-bound. Frequent so the return shows up quickly. Dull because nobody will defend keeping it manual. Rule-bound because that is where AI errs least.
Invoice processing, email triage and CRM updates are the usual candidates. For a wider starting point, we have a list of the 20 business processes most companies should automate.
Measure the process before you touch it. Without that baseline, any later gain is a story rather than a result.
The question that decides it
Before buying any tool, answer this: how many hours a month does your team spend on its most repetitive administrative process, and what would it cost to review 10% of that volume by hand?
Know both numbers and you know whether AI pays. Know neither and no tool will work it out for you.