How we measure bookkeeping without a human: the 99.5% method
Manoj Tutika · Published 16 November 2026 · Last checked 16 November 2026
Working on this yourself? Talk to Mark, our co-founder. We’re here to solve it with you.
What does the 99.5% figure say?
On 19 September 2026 we counted every work unit Thred Ledger had processed in production: 36,963. Of those, 196 had been changed by a person. The other 36,767 were completed by the agents with nobody altering the result. That is 99.47%, rounded to 99.5%. Together they booked EUR 19.6 million in debits, from bank charges of a few euros to invoices of tens of thousands.
The figure is about Thred Ledger, the ledger Thred runs for small businesses, mostly Dutch B.V.s. Bank lines, payments from Stripe, Shopify, PayPal and Mollie, and emailed documents arrive there and are booked automatically, with the source document linked to every entry.
What counts as a work unit?
A work unit is a piece of work a bookkeeper would otherwise do, with its own outcome. We counted five kinds.
| Kind of work | Count | What the agent decides |
|---|---|---|
| Bank transaction classification | 12,901 | Account, counterparty and VAT for a bank line |
| Journal entry | 16,915 | The entry itself, with debit, credit and source |
| Purchase or sales document | 3,007 | Supplier or customer, amounts, VAT and due date from the document |
| Reconciliation link | 2,124 | Which payment belongs to which invoice |
| VAT determination | 2,016 | The VAT treatment of a line, such as reverse charge or intra-EU |
| Total | 36,963 | |
| Changed by a person | 196 | |
| Not changed by a person | 36,767 (99.47%) |
When does a unit count as no-human?
A unit counts as no-human if nobody changed its result after the agent. We look at the change history of each unit. These count as a human change: choosing a different account, editing or reversing an entry, breaking a reconciliation link, changing a VAT treatment and manually editing a document.
There were 11 manual journal entries, all opening balances for new customers. They are included in the 196.
How did we measure?
- Fix the period. All units in production up to and including 19 September 2026, from the first customer onwards.
- Count units by kind. Read directly from the production database, read-only, with one definition per kind.
- Look for changes. For each unit, check the history for a change by a user after the agent.
- Production only. The test environment and demo data are excluded.
- Divide and round. 196 changed units out of 36,963 is 0.53%. The remainder, 99.47%, rounds to 99.5%.
Which other measures belong with it?
One figure never tells the whole story, so we recorded these in the same measurement.
| Measure | Result | What it tells you |
|---|---|---|
| Bank transactions no person touched | 97.4% | Most bank lines are booked without anyone opening them |
| Corrections to bank classifications, July to September 2026 | 0.32% | How often a person fixed a result afterwards |
| Reversed reconciliation links | 0.24% | How often a payment and invoice link was wrong |
| Time from document to entry, median September 2026 | 6.6 minutes | How fast an emailed invoice reaches the books |
| Manual journal entries | 11 | All opening balances for new customers |
Time to entry depends on the channel. Through WhatsApp the median was about a minute, through Gmail and the upload address about ten minutes. Books that keep up this fast do not need a monthly catch-up, and the month-end close becomes a review instead of a rush.
What are the limits of this measurement?
An honest figure has limits, and we would rather state them ourselves.
- Not changed is not the same as reviewed. Not every unit nobody changed was looked at by a person, so the correction rate is a floor.
- It covers Thred Ledger. Not the agents on Exact Online, Xero or Moneybird, where each write comes as a proposal for approval or under a rule you set.
- Small businesses, mostly in the Netherlands. A group with thirty entities or many currencies may see different results.
- No external audit. Thred measured this itself. We publish the definitions so you can test them.
- A snapshot. The figure covers the period up to 19 September 2026. New measurements will appear here, dated.
Why publish the method?
Every AI bookkeeping product quotes a percentage. Without a definition, the number means little. Is a unit an invoice or a line? Does a question to the client count as human work? Is reviewed the same as unchanged? Four questions help with any figure, including ours.
- What is a unit, and how many were there?
- Over which period and which customers was it measured?
- Does any human change count, or only an explicit rejection?
- What share did a person review, and what did they find?
A vendor without answers to these is giving you a marketing number, not a measurement. For how to test an agent on your own books, read evaluating accounting agents on real books.
What does this mean for Thred's agents?
Thred is a team of AI agents that takes the recurring finance work off your team, on the accounting software you already use. The Bookkeeper keeps the record up to date and is the agent behind most units in this measurement. The Connector brings in documents and payments from email, WhatsApp, bank and payment providers. The Accountant sets the VAT treatment per line and prepares the return.
On Exact Online, Xero and Moneybird the same agents work within your rules. Each write arrives as a proposal you approve, until you release a type of entry. That is what this figure is about: work nobody has to fix. Read about AI agents for finance teams or see Thred Ledger.
Frequently asked questions
What does 99.5% without a human mean?
Of the 36,963 work units Thred Ledger processed in production up to 19 September 2026, 36,767 were not changed by any person. That is 99.47%, rounded to 99.5%. It is work the agents completed on their own, with nobody altering the result.
Does 99.5% mean the books are 99.5% correct?
No. It measures that nobody changed a unit, not that someone reviewed every unit. An error nobody spots counts as no-human. That is why we also publish the correction rate on bank classifications and the limits of the measurement.
What counts as a work unit?
Five kinds of work: a bank transaction classification (12,901), a journal entry (16,915), a purchase or sales document (3,007), a reconciliation link between a payment and an invoice (2,124) and a VAT determination (2,016).
Does the figure apply to Thred on Xero or Exact Online?
No. The measurement covers Thred Ledger, Thred's own ledger for small businesses. On Exact Online, Xero and Moneybird the agents work with approval per write or with rules you set; we will publish a separate measurement for those.
How often do you re-measure?
We aim to repeat the measurement every quarter with the same definitions, so the figures stay comparable. New results will appear on this page with their date and counts.
Sources
- Thred Ledger, Thred's own ledger (source of the 19 September 2026 measurement)
- Thred Layer, AI agents on your existing accounting software
Read next
- AI agents for finance teams: the recurring work they take over
- How an agent learns your booking rules without you writing them
- Evaluating accounting AI: how to test an agent on your real books
- Is your company ready for AI in its books? A readiness check for founders and finance leads
- AI bookkeeping for UK SMEs: what it does today, and what still needs a person
Where Thred comes in
Thred’s own general ledger, built for agents. Every transaction is booked as it happens, under your rules, and traceable back to its source.
