Decide whether to build or buy an AI tool

The CFO asks for one recommendation on automating invoice capture for the AP team: Option A, Option B, or Option C. Check each option against the five gates, read the pilot, compare three-year cost and cost per invoice on the same basis (correcting the internal tools team's estimate with the planning rules), and score the options that may be chosen. Recommend one option with the contract conditions you would negotiate, an exit plan, a fallback, and who decides by when.

The company, its figures, the options, their terms, and every price are fictional.

Suppose you are the finance lead at a chain of 40 fitness clubs in two EU countries. Your accounts payable (AP) team of 6 people processes about 5,000 supplier invoices a month, and volume is expected to stay about flat for three years. Invoices arrive by email as PDFs, scans, and phone photos: equipment servicing, cleaning, utilities, café stock, and self-employed personal trainers and class instructors, whose invoices carry their names, home addresses, bank details, and tax numbers. Staff key in 8 fields per invoice (supplier, invoice number, invoice date, net amount, VAT, total, club cost center, purchase order number) and check for duplicates before each payment run.

The CFO wants invoice capture automated and asks you for one recommendation. Three options are on the table:

• Option A: the AI invoice-capture add-on in the finance suite the team already uses.
• Option B: a specialist invoice-capture vendor.
• Option C: an in-house build by the 3-person internal tools team on Provider M's model API (a service that runs Provider M's AI model on request and charges per use).

The company's gates. The CFO set them with IT security and counsel. An option that fails a gate cannot be chosen until a signed change fixes it. Every outside company that receives invoice data must meet G1, G2, and G3. Whichever tool the AP team uses, bought or built, must deliver G4 and G5. Anything built in-house must also pass the company's security review before go-live.

What each option offers.

Contract terms.

Own-data pilot. 400 invoices from last quarter, drawn across all 40 clubs. The sample includes 60 phone photos from trainers, 40 credit notes, and 12 known duplicates (invoices the AP team caught as resent last quarter; each original is also in the sample). Every option ran on the same 400 invoices. An AP specialist checked the 8 fields on each invoice against what was posted: 3,200 fields per option. An invoice needs a manual fix if any field is wrong or missing. An invoice with no result counts as needing a fix, with all 8 of its fields wrong.

Pilot bar, set before the pilot: at most 20% of invoices need a fix, and at least 6 of the 12 known duplicates are flagged.

Cost inputs. All prices are fictional. Three years = 36 months; 5,000 invoices a month = 180,000 invoices.

Loaded rates (the full cost of an hour, including benefits and overhead): AP staff €46 an hour; internal tools engineers €95 an hour. All internal hours in the table are engineer hours.

Cost rules. Count the same lines for every option over 36 months: fees and usage, implementation, internal running time, and AP time to fix invoices. AP fix time per month = the option's pilot fix rate × 5,000 invoices × 5 minutes, at the AP rate. The one-minute check every invoice gets is the same for every option and is left out. Training time is similar across options and is left out.

Planning rules for in-house tools (company policy; the team's estimate did not apply them):

• P1: Gate features are part of the build. Asked to size SSO, the audit trail, full export, and a duplicate check against 24 months of past invoices, the team estimated 160 more hours, at the same 40 hours a week.
• P2: Maintenance is 25% of total build hours per year.
• P3: Plan one model migration a year (Provider M retires models on its own schedule), 30 hours each, including a re-test on a fresh invoice sample.
• P4: Plan model usage at twice today's per-invoice price, to cover a successor model priced differently and re-runs.

Scoring (weights fixed before the pilot): three-year total cost 50%, duplicate flag rate 30%, time to go live 20%. On each criterion the best option scores 10; the others score 10 × lowest cost ÷ their cost, 10 × their duplicate rate ÷ best rate, and 10 × fastest weeks ÷ their weeks. Score only options that may be chosen. Show criterion scores to two decimals and the weighted total to one decimal.