ROI is the first question every CFO asks about AI agents, and it deserves a better answer than a vibe and a demo. This is the arithmetic we actually run with customers: where the return comes from, what it costs to keep an agent alive in production, the payback we put in writing, and the honest cases where the numbers don't work.
Where the return actually comes from
Every production agent we run earns its keep from some mix of four sources. First, hours returned: the repetitive handling work, reading, matching, typing, chasing, moves from people to the agent, and people keep only the exceptions. Second, error cost avoided: agents that fill set fields and cite their sources don't make the transposition and fatigue mistakes that surface weeks later as wrong payments, wrong notes, or wrong approvals.
Third, speed: when an invoice posts in seconds instead of days, suppliers stop calling, discounts stop expiring, and month-end stops being a crunch. When a clinic answers every call, no-shows fall. Cycle time is money even when nobody's salary changes.
Fourth, scale without headcount: volume can double without the queue doubling, because the agent's capacity isn't tied to hiring. That's usually the source that turns a good ROI into an obvious one.
The cost side: running cost decides the ROI
The build cost of an agent is one number. The running cost is a number every month, forever, and it's the one that quietly kills AI business cases.
Token-priced AI scales its bill with the volume of work: the busier the agent, the worse the maths. Our agents run small, job-specific models on your own infrastructure, which makes the running cost roughly ten times cheaper in production than the general-purpose approach, and keeps it nearly flat as your documents and users grow.
We also treat cost as an operations discipline, not a quarterly surprise: the running cost is watched from the first day in production and managed as volumes grow. FinOps for AI, handled for you.
The payback model we put in writing
An engagement starts with a short, fixed-fee assessment from $10K: we pick the right first workflow with you, scope it on your own data, and give you an honest go or no-go before you commit to anything bigger.
The agent build itself starts from $15K, runs on your servers or your own cloud account, and goes live in production in two to four weeks. Not a pilot: real work, every answer cited, the running cost watched from day one.
And the line that matters: most agents return their cost within six months. Fast timeline, production-grade, payback inside six months. These are written commitments, not slideware.
A worked example, anonymized
A multi-business-unit conglomerate receives thousands of supplier invoices a month. Before: every invoice read by a person, matched by hand against purchase orders and goods receipts, posted into the finance system, with errors surfacing weeks later and month-end requiring overtime.
After: the Invoice Intelligence agent reads any format, matches each invoice three ways, applies the tolerance rules, and posts a clean, cited entry. Hours of manual three-way matching become a few seconds per invoice, and people handle only the genuine exceptions the agent flags.
The return stacks all four sources at once: handling hours come back, matching errors stop leaking money, suppliers get paid on cycle, and invoice volume can grow without the AP team growing with it.
How to measure it on your own workflow
Baseline four numbers before anything is built: items per month (invoices, calls, claims, documents), minutes of handling per item, the exception or error rate, and what a single error costs you downstream.
After go-live, compare the same four. The agent's running cost sits on the other side of the ledger, and because it's nearly flat, the arithmetic gets better every month your volume grows.
We set this measurement up with you during the assessment, on your real workflow and your real volumes, so the business case is yours, not ours.
When the numbers don't work
Honesty clause: not every workflow pays back. If the volume is low, the process changes every quarter, or the data the agent needs is locked somewhere we can't responsibly reach, the ROI isn't there, and we'll say so at the assessment.
That's what the go-or-no-go is for. A no costs you a short assessment. A bad yes would cost you a year of quiet running costs. We'd rather give you the no.