Nobody in finance is doing accounting for most of the month. They are assembling evidence.
The invoice hunted for in Finder and re-attached for the fifth time. The three amounts on screen that do not tie. The reconciliation that waits a week because it crossed a desk. The AI licences on your own cost line that you cannot prove anyone opened. Below is what that work looks like today, what it looks like once Silow has mapped it, and what it gives back — measured, or not claimed at all.
How the work looks today. How it looks after.
Seven processes, in the order we would take them. The left-hand column is not a caricature of a finance team — it is what two recorded teams were observed doing while invoices, documents and approvals moved through their day. Where a row carries a number, the tag says it was measured in real recorded work. Where a row is observed without a count, the tag says that too.
Three figures sat in the reconstructed procedure — an invoice for 35,000, a figure of 35,700 EUR, and a conversion of 122,000 EUR — and they do not reconcile. Someone was looking at them in turn, across windows, holding the difference in their head. That is the actual shape of reconciliation work: not arithmetic, but the retrieval of the four documents the arithmetic needs.
The evidence is assembled around the number before a human opens it: the invoice, the bank line, the contract, the prior period, and an explicit list of what is missing. What our report did with those three amounts is the point — it refused to invent a reconciliation and wrote only that the trace shows them being looked at in turn. A system that will not fabricate a tie is the only kind a finance team can use.
One PDF was manually located in Finder and re-attached across two messengers on five separate occasions in three weeks. No shared link for it existed. Every request for a missing invoice, receipt or contract is retyped from scratch by someone who knows exactly what it will say and exactly who is going to ignore it.
Documents have one retrievable place, so the file is never hunted for again. The recurring chase — the supplier who has not sent the invoice, the manager who has not sent the receipt — is drafted from the language your team already uses, with the right attachment already attached. A human reads it and sends it.
The same pull from the same systems into the same workbook into the same deck, every month, performed by the person senior enough to know which of the numbers is wrong. Fixed sources, fixed shape, fixed output — and a week of the month spent producing it rather than reading it.
The pull, the assembly and the format run themselves; what lands on the controller’s desk is the part that needs judgement — what moved, what broke, what has to be explained. It is the highest-frequency, lowest-judgement ritual in the function, and it is the definition of an agent specification.
In the operations team we recorded, a messenger took about 24% of screen time — more than every internal system combined — with the browser and a spreadsheet taking most of the rest. The approval, the exception, the promise to pay, the "can you just check this" all live in threads. Your ERP is not where the work happens. It is where the work is finally typed up.
What actually moved, when, and on whose desk becomes a by-product of the work rather than something a person has to remember to write down. The thread stops being the system of record by accident.
A task is moving 3.4% of the time it exists — the other 96% it is sitting in someone’s queue. Work that needs a second human waits a median of 7 days; work one person can finish alone waits 1.1. Two thirds of all waiting sits across a handoff. And it is not time zones: the least-overlapping pair still shares a quarter of the working day. In a finance team that queue has a name — it is the approval chain, and the answer you are waiting on from the business.
Every handoff gets a defined shape: what is attached, what "done" means, who is waiting and since when. The unclosed queue becomes visible and owned. You are not making the accountant faster — you are attacking the 96%, which is where the close actually goes.
You have signed off AI licences and you report adoption from seats. Measured that way, the go-to-market team we recorded used AI 11.7% of the time. Measured from the actual work: 14.9%. One person was undercounted sevenfold — 0.9% against a real 6.2% — because his assistant runs inside another application, so the telemetry logs the outer application and never sees the assistant at all.
The share of real work that actually ran through a model, measured from the work itself rather than from logins. Every CFO in 2026 has bought AI licences and cannot prove what they bought. This is the one number on the page that is aimed squarely at the person signing the invoice for it.
In the operations team we recorded, the company had built its own internal AI assistant and recommended it to the team. Recorded usage across eight people: seven at zero percent, one at one percent. The licence is paid every month. It is on your cost line right now, and nothing in your reporting will ever tell you.
Adoption measured from the work, and — more usefully — the reason. An assistant nobody opens is almost never a model problem. It is that nobody mapped it onto a procedure anybody actually performs. That is the same mapping that makes the rest of this page possible.
Every figure on this page comes from anonymized recorded work. One capture was a 9-person go-to-market team — 5 weeks, ~325 hours, 748 tasks, ~88,000 scenes — whose work included licensing, regulator documents, client onboarding and invoicing. The other was an 8-person cross-border operations team — 8 days, ~95 hours, 461 tasks. Rows tagged "Measured" carry numbers from that recorded work; rows tagged "Observed in the trace" describe shapes we saw without inventing a count. Every figure was computed before a model was allowed to write a sentence about it.
What an agent may touch, and what it may not.
In finance the line between "assembled the evidence" and "made the entry" is not a detail — it is the product. Everything on the left is preparation. Everything on the right has a signature behind it, and it stays where it is.
- Assembling the reconciliation evidence: the invoice, the bank line, the contract, the prior period, and what is missing.
- Chasing the missing invoice or receipt — the same request, drafted for the fortieth time this month.
- Pre-filling the month-end pack from the sources it always comes from, in the shape it always takes.
- Drafting the recurring supplier and AP chase in the language your team already uses.
- Producing the audit trail — what moved, when, in what order, on whose desk — as a by-product, not a task.
- Watching the queue: what is unclosed, on whom it is waiting, and for how long.
- The journal entry and the accounting judgement behind it. Always.
- Anything a signature stands behind, and anything an auditor expects a named person to have done.
- The exception and the write-off.
- The decision to accept an estimate.
- Every draft an agent produces is reviewed before it is sent externally. Nothing auto-sends to a supplier, a client or an auditor.
Silow does not make an accounting judgement and it does not make a decision about a person. It removes the assembly work in front of the entry, and leaves the entry — and the named human who stands behind it — exactly where an auditor expects to find them.
Time, money, and the people you already have.
The rule for this section, and you are the buyer who will hold us to it: not one invented percentage. Every figure below is either measured in a recording or supplied by you — hours × your loaded cost, your rate, not a benchmark we made up. A saving you cannot defend in a board meeting is not a saving. It is a slide.
We do not quote you an industry percentage, and we will not quote you another vendor’s case study, because we cannot verify it. Every candidate above is scored in hours against your own recording and ranked by payback — so the first thing you build is the one that pays back fastest, not the one that demos best. The anchors we can already point at: a task moving 3.4% of the time it exists. A median seven-day wait each time work crosses a desk. A document re-found and re-attached five times in three weeks.
The first is hours × your loaded cost. The second is the one nobody counts and you are best placed to care about: the AI you already pay for and nobody opens — recorded at zero to one percent across eight people in the operations team we recorded, on a licence that is billed every month. Before you buy any new AI, it is worth knowing what the last one bought you.
No new system to adopt, no data to move, no process change while we look — Silow runs on the tools the team already uses. The capacity comes back inside the same team, doing the same job with the assembly work removed. This is released capacity, not a headcount case: Silow maps work, not workers. What you then do with the hours is a decision for the people running the business, taken with better information than they have ever had.
A task is moving 3.4% of the time it exists. Making the accountant faster optimises that 3.4%. The close lives in the other 96%.
From a close nobody has ever watched to a layer you own.
The agents are the visible part. The thing that makes them possible — and that keeps paying after they ship — is the record of how the finance function actually works.
The close checklist exists, and it is wrong. The real procedure — the one performed at 09:00 on the second working day — has never been observed, which is why every automation attempt in finance starts with a workshop and a guess.
The real steps, from the recording, with the evidence attached — in one team we recorded, all 30 supporting quotes were verified against the raw scenes, not against a summary. Ranked by impact, effort, risk and payback against your own hours.
The evidence assembly, the document chase, the pack pre-fill. Each arrives as a specification an engineer can build from: the observed steps, the scenes behind them, and the line they must not cross. The entry stays human by design.
The agents are consumers of something more valuable: a private, structured record of how your finance function actually works. Onboarding that teaches from it. Search that answers from it. An audit trail that is a by-product rather than homework. On data only you own.
Where AI actually works in a finance function — and where it stalls.
Everyone selling AI to finance teams has a case study with a number in it. We are not going to quote you someone else’s, because we cannot verify it and you would be right to discount it. Here are the patterns instead, including the failures.
The clearest win in the function and the least glamorous. Reading a document into a structured record is close to solved. What is not solved is everything around it — where the document was, who was chasing it, and what happens when it does not match.
Matching is a mechanical problem and mostly a solved one. The cost sits in the exceptions and in the retrieval of the four documents an exception needs. Assemble the evidence; never let a model invent the tie.
Highest frequency, lowest judgement, and the ritual most teams believe they already solved with a brittle workbook. It looks solved. It is a senior person’s recurring week.
The chase is not a finance problem, it is a coordination problem — and it is where the calendar actually goes. Nobody buys software for it because it is not a tool category, which is precisely why it stays broken.
Where it works: the trail falls out of the work as a by-product. Where it stalls: someone is asked to reconstruct, three months later, why a thing was done — from a messenger thread that no auditor can follow.
The failure mode that ends an AI project in finance is not a slow agent, it is a confident wrong figure. Our own report met three amounts that did not reconcile and refused to reconcile them. That refusal is the feature.
Pilots do not die because the models are bad. They die because nobody could say which of forty candidate processes to do first, or what "correct" looked like when it was done. That ranking is what Silow produces; the agents are what you build on it.
Nobody automates the waiting. Two thirds of the elapsed time in the team we recorded sat in a queue between two people — and in finance that queue is the approval chain. It is the largest number on this page and the cheapest one to move.
The part other vendors leave out.
You are the buyer who most punishes a number that cannot be defended, so here are ours — the things we cannot claim and the two products our own data killed. Better from us now than from you in month three.
The recording sees activity, never outcome. Whether the close landed sooner, whether the invoice was paid — that is not in the data. Join one column of your outcomes to our trace and it becomes provable. Until then, any ROI figure quoted at you, including by us, would be a guess.
A duplicate-work detector: 44 recoverable minutes across nine people over five weeks — 0.22% of working time, median duplicate 15 seconds. There is no product there. A meeting-ROI report: three quarters of what the metric called "meeting minutes" had no room attached. Both were cut before they ever reached a customer.
In the operations team we recorded, we captured about two hours per person per working day, and the report says so on its own coverage page instead of quietly extrapolating to a full week. Numbers are computed before a model is allowed to write a sentence about them, so no figure gets invented — but a true number can still be attached to the wrong claim. We have caught exactly that twice, both times with human eyes.
Weeks, not quarters.
Silow runs on the team’s existing machines and tools. No migration, no new system to adopt, no process change while we look. You choose the deployment mode — including fully air-gapped.
Scenes become tasks; tasks become the real procedure. Numbers are computed before a model writes a sentence about them, and every quote is checked back against the raw scene it came from.
Every candidate scored by impact, effort, risk and payback against your real hours — so the first thing you build is the one that pays back fastest, not the one that demos best.
A specification, not a suggestion: the observed procedure, the evidence, the scenes behind it, and the boundary it must not cross. The evidence gets assembled. The entry stays human.
What you do with the hours is yours.
Silow maps work, not workers. No productivity ranking, no performance evaluation, no automated decision about anyone’s job — not as a policy written afterwards, but because the unit of analysis is the process. Two weeks of recording, and the first ranked roadmap lands with the hours attached and the provenance of every number on it.