AI ROI: How to Calculate It Without Using Vendor Math

Where vendor math breaks down

Vendor ROI models share a consistent structure. They start with the cost of the current manual process, usually headcount × hours × fully loaded rate. Then they apply a large efficiency multiplier. Then they project this saving forward three years and divide by the subscription fee. The result is an ROI in the high hundreds of percent, because the subscription fee is the only cost in the model.

The costs that are missing: implementation (integrating the AI system with your ERP, your identity provider, your document storage, your approval workflows), the 4–6 month period where both the manual and AI processes run in parallel, change management and training, ongoing model maintenance as document formats evolve, and the correction overhead when the model makes errors that humans have to fix.

Vendors are not being dishonest. The model is structurally optimistic in a way that serves the sale and not the decision. Here is how to fix it.

The real cost structure

For a mid-market company deploying AI-assisted invoice processing with an ERP integration, here is a planning budget. The figures are rough estimates, there to show the shape of the model. Replace them with your own quotes.

Implementation: ~$65K–$95K

This is the integration cost: connecting the AI system to the ERP, configuring the extraction models for your specific document types, building the approval workflow routing, setting up the audit trail, and testing with your actual documents. It does not include the subscription fee. A much lower quote usually means a much simpler deployment, or overages later.

Parallel-run overhead: ~$15K–$25K

For the first 4–6 months after go-live, the finance team runs both processes. They process invoices through the AI system and verify the outputs against their existing process. This is not optional. It is how you build confidence in the system and catch edge cases before they cause problems. The cost is staff time that is not in the vendor's model.

Ongoing maintenance: ~$12K–$20K per year

Models drift. Supplier invoice formats change. The ERP gets upgraded. A new entity is added. Each of these events requires maintenance work: prompt updates, retraining or fine-tuning on new document types, integration updates when the ERP API changes. This cost is ongoing and it is usually not in the vendor subscription.

Hallucination correction overhead: ~$8K–$15K per year

At an error rate of 2–4% on extraction (which is good performance), a company processing 500 invoices per week will have 10–20 invoices per week that require human correction. At 15 minutes per correction, that is 30–45 hours per month of correction time. This is real labor cost that needs to be in the model.

The honest benefit structure

The benefits are real too, and they need the same honest treatment.

Cycle time value: For a company whose suppliers offer 2/10 net-30 terms, cutting invoice processing from four days to same-day lets you capture early payment discounts you currently miss. Work out how much of your spend carries those terms and how many discounts you miss today. That figure is real cash.

Error cost reduction: Manual invoice entry produces duplicate payments, misrouted approvals and reconciliation exceptions. Price the rework and penalties from your own records. A well-implemented extraction system removes much of it, though not all.

Capacity gains: Do not count these as savings unless you have a plan for how they are realized. If the freed capacity is absorbed by growth (more volume, a new entity, additional vendor onboarding), the benefit is real. If the team size stays constant and volume does not grow, the capacity gain is latent, not financial.

An honest ROI model will come out far below the 847% in the pitch, and the project can still be a good investment. Build the model honestly and it gets approved for the right reasons.

Building a model that will hold up

Use actual numbers from your current process, not industry benchmarks. Pull three months of AP data: how many invoices, how long they take from receipt to posting, what the exception rate is, what the rework cost looks like in support tickets or overtime. These numbers are defensible in a board discussion in a way that "industry average of X hours per invoice" is not.

Put the full implementation cost on year one. Do not amortize it unless your finance team requires it for capital classification purposes. Year one looks worse and years two and three look better. That is accurate. AI implementations have front-loaded costs and back-loaded returns.

Then run the model at three accuracy scenarios: 95% extraction accuracy (optimistic), 90% (realistic for the first 12 months), and 85% (conservative, especially if your documents are unusual formats). Against the 90% case, the correction overhead rises by half at 85% and halves at 95%. Show all three to the decision maker so they understand the sensitivity.