AI SHARED SERVICES · BENCHMARKS
Exceptions are only rules nobody has written down yet.
How we measure straight-through rate, exception mix and cost per document, and the numbers we can publish today.
- PROCESSES
- AP, AR, freight billing, contract intake
- MEASURES
- straight-through rate, exception mix, cost per document
- MECHANISM
- daily database snapshot, delta, new rule
- POSTS TO
- SAP, Oracle, Dynamics, NetSuite, Coupa, or REST
- STATUS
- in production on a live freight billing pipeline
Three numbers a CFO can hold us to.
STRAIGHT-THROUGH RATE
Documents posted with no person in the loop.
Counted per process, per month. A document that a person opened, even to approve, is not straight-through.
EXCEPTION MIX
Which rule each exception failed, ranked.
Not a count of exceptions. The list of reasons, ordered by volume, so the top three tell you where the next rule goes.
COST PER DOCUMENT
Read once, plus the review it caused.
Model cost to read and match, plus the minutes a person spent on the exceptions. Never a per-query cost, because nothing is re-read.
Every shared-services benchmark you have seen quotes a straight-through rate with no denominator and no exception list. We return all three numbers on your own corpus after the first pass, with the exception mix as a ranked list, so the second month has a target and the third month has a smaller list.
What we can publish today.
Loads billed per day
A freight operation running document indexing, matching and auto-billing on Talonic, four to five documents per load.Bridgeway · the customer · 2026
2,200 to 2,400
Pages per year
The volume the engagement was scoped on.Bridgeway · the customer, re-baselined during scoping · May 2026
5 to 5.5 million
Open loads each batch is matched against
Every incoming document batch is resolved against the live open-load set before a rate is checked to the penny.Bridgeway · production configuration · 2026
8,412
One-time cost to read a document
Paid once at ingest. Every question after is a lookup.Talonic · corpus benchmark · 2026
$0.17
Model cost per query after ingest
Fast path over the registry. The full table, with the RAG arms, is on the benchmark page.Talonic · corpus benchmark · 2026
$0.0015
Each figure is dated and attributed. Matching accuracy on the freight pipeline is measured and held under the customer’s release gate; it is described on the Bridgeway case page in words, not numbers, until they release it.
How one fix becomes a rule.
We snapshot your database daily and analyse the deltas. Every fix your team makes becomes a rule, and the straight-through rate climbs without anyone writing a rule down. The example below is the one printed in the briefing.
Monday 06:00 snapshot
Tuesday 06:00 snapshot
The rule the agent derives
INV-88213 · $18,400 · open
A deduction of $412 sits unresolved against it.
INV-88213 · $17,988 · cleared
Credit note CN-1042 for $412 applied. A person worked it.
A deduction that matches an open credit note clears automatically
Every matching deduction after this one clears itself. No person.
Illustrative record identifiers, as printed in the briefing. The mechanism is the production one: the delta between two snapshots is the training signal, and the derived rule is written down, reviewed, and only then allowed to decide.
The three things a person used to do.
Read the hard documents
scans, stamps, handwriting
Match with judgment
Learn from every fix
- Read the hard documents. Visual language models read the scans, stamps and handwriting that break OCR. The worst documents still come in as clean data.
- Match with judgment. Credit notes against deductions, remittances against open items, three-way matches. The matches that always needed a person.
- Learn from every fix. Daily snapshots, delta analysis, a new rule. The automation rate climbs because the exception list shrinks, not because a threshold was lowered.
What we do not publish yet.
- A straight-through rate without a process behind it. The number is real only per process, per customer, on their documents. We return yours after the first pass. We do not print a blended one.
- Accuracy on a customer’s pipeline before they release it. The freight pipeline is measured monthly. The figures are the customer’s to publish, and the case page says so in words until they do.
- An exception mix from a demo corpus. The ranked list is only useful if it is your list. A demo list would tell you where our rules are weak, not where yours are.
Send one month of AP. Get your three numbers back.
One pass over one month of documents returns your straight-through rate, your ranked exception mix and your cost per document, with the rows behind each. Clean postings land in your ERP or over REST once you say so.
Note: Every figure on this page is already public on the Bridgeway case page or the registry benchmark, with the same date and attribution. The AR example uses illustrative record identifiers, as printed in the briefing.