Setting a confidence threshold: how much should AI book?

What is a confidence threshold?

A confidence threshold is the minimum an AI proposal has to reach to be processed without human intervention. Every booking proposal in Autoboeker gets a confidence score. If that score is above the threshold you have set, the proposal is posted directly to the accounting software. If it is below, the proposal waits for review by a team member.

The threshold is therefore the main tool for balancing automation and control. A high threshold means only proposals the AI is very sure of are booked automatically, and your team reviews more. A lower threshold means more automatic processing and less review. There is no single right value for every firm: the right threshold depends on the administration, its history and how much risk you are willing to accept.

What is the difference between a confidence score and a threshold?

The confidence score is a property of one proposal; the confidence threshold is a setting of your firm. The score indicates how sure the AI is about a specific booking or match, based on things like previous bookings in the administration, corrections and how well the document fits known patterns. With every proposal you see the score together with the explanation of what it is based on.

The threshold decides what happens with that score. Two firms can see the same score for the same invoice and still choose a different outcome: one books automatically, the other wants to see the proposal first. That makes the threshold a policy decision rather than a technical setting. So record who in the firm may change the threshold, and why an administration has a particular threshold.

Which levels of automation can you choose?

In Autoboeker you choose per firm or per administration between three levels of automation. They range from approving everything by hand to automatically booking everything above the threshold. The levels are meant to be worked through step by step, so you build confidence per administration before letting more be processed automatically.

The difference lies in what your team still sees. With suggestions you review every proposal, even when it is entirely correct. With smart confirmation you only see what deviates. With fully automatic booking, high-confidence proposals are processed directly and you only see what falls below the threshold or is flagged, such as a high amount or an unknown supplier. The three levels are:

• Suggestions: the AI fills in the booking, you click to approve.
• Smart confirmation: you only review the deviations.
• Fully automatic: high-confidence proposals are booked directly.

How do you choose the right threshold per administration?

You choose the right threshold per administration based on how much history the AI has, how predictable the administration is and how serious the consequences of an error are. An administration with many recurring suppliers and a stable chart of accounts lends itself to automatic booking sooner than a new client or an administration with many one-off, large or international transactions.

Also consider what happens after booking. If an administration is reviewed periodically for a compilation engagement, you may choose a different threshold than for a client whose figures are used directly for decisions. With approval before or after booking you can combine the threshold with a sign-off step. Factors to weigh are:

• The number of historical bookings in the administration.
• The share of recurring suppliers and cost items.
• Complex VAT situations, such as reverse charge, EU transactions or an exemption.
• The size of individual amounts and the impact of an error.
• Agreements with the client about review and approval.

How do you set the confidence threshold step by step?

You set the confidence threshold best in phases: start cautiously, look at what the AI proposes and increase automation per administration once the proposals are consistently right. Every correction you make in the early phase improves the next proposals, because the AI learns from corrections and historical bookings. That first period of reviewing is not lost time, but an investment in the quality of later automatic bookings.

Team statistics help you adjust. You see per team and team member the number of bookings, the automatic booking percentage and open client questions, and per administration how far automation has progressed. That shows where you can loosen the threshold and where an administration actually needs more attention. A practical order:

1. 1Connect the accounting software and choose the administrations you want to automate.
2. 2Let the AI learn from the history and master data of each administration.
3. 3Start with suggestions and review the proposals and their confidence scores.
4. 4Record recurring corrections as an administration rule or AI instruction.
5. 5Move to smart confirmation once the proposals are consistently right.
6. 6Set stable administrations to fully automatic and keep an eye on team statistics.

Where else does a confidence threshold apply?

In Autoboeker a confidence threshold applies not only to booking invoices and receipts, but also to bank matching and client questions. In bank matching every match gets a confidence score and only matches above your threshold are reconciled automatically; below it a proposal waits and deviations are highlighted. Bank matching is available for Exact Online, Twinfield and SnelStart.

Client questions can be sent fully automatically or from a confidence threshold. For firms from 25 FTE, Atlas, the AI agent in the Groot kantoor plan, also works within the confidence thresholds and rules the firm sets, and escalates whatever needs human judgement. In every case, each action by a person or the AI is traceable in the audit log, so afterwards you can see what happened above and below the threshold.