Finance wants a three-year commitment to a hosted invoice extractor — which costs beyond the per-document rate decide whether to sign?
answer
- the rate is the computable part
- coupling accrues while nothing changes
- exit is cheapest in month one
- buying accumulates a bill, not a corpus
- where the documents are processed
basics
~20 sExit cost, the asset you never accumulate, and where the documents are processed. A long commitment locks a rate you can compute and a coupling you cannot: downstream systems written against their fields, no labelled corpus of your own, and customer invoices leaving your boundary for three years.
solid answer
~50 sThe rate is the most computable number in the decision and the least decisive. Three costs sit outside it. **Exit cost** grows every quarter: consumers written against their field names, review thresholds tuned to their confidence scale, corrections stored in their shape — all of it work somebody must redo later, and it is never cheaper than it is in month one. **The asset you forgo**: building accumulates labelled invoices of your own, which lowers your marginal cost and moves the crossover in your favour; three years of buying accumulates a bill. **Data exposure**: customer invoices carry bank details, prices and supplier terms, so where they are processed, how long they are retained and whether they train somebody else's model are questions no rate prices. If you sign, buy a shorter term, keep your own field schema and confidence semantics at the boundary, and write down the trigger that flips the decision.
go deeper
Recall that a long contract commits more than money. Once other systems are written around somebody else's output format, changing your mind later becomes an engineering project rather than a decision.
Be able to name the three costs outside the rate: what leaving would take, the labelled data you never collect, and where customer documents are processed and retained.
Show how exit cost accrues through concrete coupling — field names, confidence scales, corrections stored in their shape — and how owning the boundary keeps it flat.
Own the judgment: the multi-year forecast is the weakest input and exit cost is not a number until tested, so buy the reversible version and write the trigger that reopens it.
## Why the rate is the least interesting number A three-year commitment is usually presented as a discount: a lower price per document in exchange for a term. That framing hides the actual trade. The rate is the one quantity in the comparison you can compute exactly; the quantities that decide the outcome are the ones that have no arithmetic answer at all, which is precisely why this is a call a lead owns rather than a spreadsheet exercise. Three of them matter. ## Exit cost, and why it only grows Exit cost is not the integration work done again. It is everything downstream that has quietly taken a dependency on the shape of the bought output: - **Field names and structure.** Every consumer of extraction results — the ledger import, reconciliation, reporting — is written against somebody's schema. If that is theirs, changing providers is a change to all of them. - **Confidence semantics.** If the review queue's thresholds are tuned to their score scale, a different scale means re-tuning a queue whose staffing is already budgeted. - **Corrections in their shape.** Every reviewer correction stored in the provider's format is an asset held in a currency you cannot spend elsewhere. - **Behavioural assumptions.** Downstream rules that quietly rely on how *this* extractor fails — which fields it tends to miss, how it renders a date — are the ones nobody documents and everybody discovers during a migration. Coupling accumulates while nothing changes on your side. Exit is cheapest in month one and is never cheaper again, so a long term is a bet that you will not want to leave during exactly the period when leaving gets most expensive. ## The asset the buy option does not accumulate Building an extractor produces two things: predictions and a labelled corpus of your own invoices, grown by every reviewer correction. That corpus is what makes your marginal cost per document fall over time, which moves the volume crossover steadily in your own favour. Buying produces predictions and an invoice. Three years of it means three years in which the crossover never moves in your favour, because nothing on your side gets cheaper. The partial answer is to store extractions and corrections **in your own format from day one**, so the buying period accumulates the corpus anyway — subject to having checked that reusing the outputs that way is permitted, which is a contract question and not a technical one. ## Where the documents are processed Scanned supplier invoices are not neutral payloads. They carry bank details, negotiated prices, supplier terms and sometimes personal data. Sending them to an outside processor raises questions that a per-document rate never prices: - which jurisdiction the documents are processed and stored in; - how long they are retained after extraction, and what deletion actually means; - whether your documents may be used to improve the provider's model; - what your own customer contracts already promise about sub-processors. Any one of these can make the cheaper option unavailable, and none of them is visible in a comparison of rates. ## What each option actually buys | | hosted, three-year term | your own extractor | |---|---|---| | time to first value | weeks | quarters | | cost at low volume | lower | higher | | exit cost over time | grows quarterly | largely yours already | | labelled corpus | not accumulated | accumulates | | documents leave your boundary | yes | no | | quality changes | on their schedule | on yours | ## Signing anyway, without signing blind A lead who decides to buy can still buy the reversible version of it: 1. **Shorten the term** and pay for the privilege, treating the difference as the price of the option to leave. 2. **Own the boundary**: your own field names and your own confidence semantics in front of theirs, so downstream consumers never learn the provider's shape. 3. **Keep the head on rules.** The concentrated senders stay on a path you operate, which is both the degraded path and proof that leaving is possible. 4. **Store corrections in your format**, so the corpus accrues whichever way the decision goes. 5. **Write the trigger down** — the volume, the exposure requirement or the coverage failure that reopens the decision, with a date. The honest closing line is that the three-year volume forecast is the least reliable input in the whole comparison, and exit cost is not a number until somebody tries. That is why the discipline is to keep the decision cheap to reverse rather than to compute it more precisely.
- What makes exit cost grow while nothing changes on your side?Coupling accumulates by itself. Every consumer written against their field names, every threshold tuned to their confidence scale, every correction stored in their shape is work somebody must redo. None of it is visible as a change, and all of it is exit cost.
- The build option is more expensive at your volume for the whole term — can it still win?Yes, when what it buys is not on the cost sheet: documents stay inside your boundary, a labelled corpus accumulates and lowers your own marginal cost, and the decision stays reversible. A lead signs for those; a rate comparison alone cannot see them.
- How do you keep the choice reversible without paying for both options?Buy the capability and own the contract around it: your own field schema and confidence semantics at the boundary, per-sender rules kept alive over the concentrated head, corrections stored in your format, and a written trigger with a date for reopening the decision.
saying these in an interview costs you the question
- Reading a long commitment as nothing more than a discount on the rate
- Assuming switching later costs only the integration work again
- Ignoring that years of buying accumulate no labelled invoices of your own
- Sending customer invoices outside without asking where they are processed
- Trusting a three-year volume forecast as firmly as this month's
- Tuning the review queue directly to a provider's confidence scale