The real cost model of agent-run back-office operations

The spreadsheet your CFO has already seen - and rejected

Every failed agent business case we have inherited looks the same. One column of loaded FTE cost, one column of software cost, a subtraction, and a payback period of four months. CFOs reject this spreadsheet on sight, and they are right to. It assumes the agent replaces a whole person, that the person's work was the only cost in the workflow, and that the agent runs itself. None of those hold in production.

When we run an operational audit before a deployment - mapping every step of, say, an invoice-matching or order-exception workflow across the client's ERP - we consistently find that direct labor is 40 to 60 percent of the true workflow cost. The rest is rework from errors, escalation time from senior staff, working capital tied up in slow cycle times, and the audit overhead of proving the work was done correctly. If your model only counts the labor, you are arguing about half the picture, and the CFO knows it.

Price the workflow, not the agent

The unit that matters is cost per completed workflow instance: one invoice matched and posted, one carrier claim filed, one renewal quote issued. Everything in the model rolls up to that number, for the current human process and for the agent-run process, measured the same way.

For the human baseline, take fully loaded labor cost per hour, multiply by measured handle time including rework, then add a per-instance share of supervision and QA. In finance ops we typically land between 9 and 30 dollars per instance for things like three-way match exceptions. Nobody believes that number until the audit produces it from their own timestamps.

For the agent side, the per-instance cost has four components, and honest models include all of them: inference and infrastructure (usually the smallest line, often under 40 cents per instance), the amortized build cost of the workflow, the human review gate, and ongoing operations - evals, monitoring, and maintenance as the client's ERP and policies change. A model that shows only the inference line is not a cost model. It is a demo.

The review gate is a feature of the cost model, not a bug

Every workflow we run in production has a human review gate, and the economics of that gate are where most of the leverage lives. On day one of a rollout we typically route 100 percent of agent outputs through review. A reviewer confirming a pre-assembled decision takes 30 to 60 seconds against a 12-minute manual handle time, so even full review is already a 5x to 10x labor reduction per instance.

Then the gate narrows on evidence. We run evals continuously against a golden set built during the audit, and as measured accuracy holds above threshold for a category - clean two-way matches, sub-500-dollar credit memos, standard carrier claims - that category graduates to sampled review, usually 5 to 10 percent. High-ambiguity and high-dollar categories stay at full review indefinitely, on purpose.

Model this explicitly: review cost per instance equals review rate times reviewer minutes times loaded rate. In a typical deployment that line falls 60 to 80 percent between month one and month six, and it falls because the eval data says it can, not because someone got comfortable. That declining curve, backed by eval evidence, is the single most persuasive exhibit you can put in front of a CFO, because it shows cost falling for a reason they can verify.

Error rates are a cost line, so measure the human baseline first

The most common mistake in agent business cases is comparing agent error rates to zero. Human back-office error rates in the workflows we audit run 2 to 5 percent, and each error carries a blast radius: a duplicate payment, a mis-keyed PO that cascades into a month-end close problem, a shipment claim filed past the carrier deadline and forfeited entirely.

So the model needs three numbers per process: error rate, average cost per error, and detection lag. A 3 percent error rate at 400 dollars average remediation on 8,000 monthly instances is 96,000 dollars a month that never appears in the labor line. Agent-plus-review-gate systems in our deployments typically run at 0.3 to 0.8 percent escaped-error rates on graduated categories, with detection lag measured in hours instead of weeks because every action is logged and eval-checked.

Put the delta in the model as a hard line item. It is frequently the second-largest source of savings after labor, and unlike labor it does not require anyone to argue about redeployment or attrition.

An agent workflow is not a license you buy. It is a production line you operate, and it should be priced like one.

Cycle time is working capital wearing a disguise

Speed sounds like a soft benefit until you translate it into cash. If agents compress invoice processing from 9 days to 1, you capture early-payment discounts you currently forfeit - 2/10 net 30 terms on relevant spend are worth an effective annualized return north of 35 percent. If they compress order-to-invoice on the revenue side by a week, you have permanently pulled a week of receivables forward, and DSO improvement is a working capital release your CFO can size in seconds.

In logistics ops the mechanism is deadlines rather than discounts: claims and disputes have filing windows, and work that sits in a queue past the window becomes a 100 percent loss. One client was silently forfeiting a mid-six-figure annual sum purely to queue latency. The agent did not do the work better than the team - it did the work before the deadline, every time.

Model cycle time as: days compressed, times daily value flow through the workflow, times the client's cost of capital, plus any deadline-driven recoveries. It is usually the line that moves a business case from defensible to obvious.

The framework on one page

Here is the structure we bring into CFO reviews, and the order matters. One: baseline unit cost per completed instance from the audit - labor, rework, supervision, error remediation - built from the client's own system timestamps, not industry benchmarks. Two: agent unit cost with all four components, shown as a curve over the rollout months as review rates graduate down, never as a single steady-state number. Three: error economics as a separate line with the escaped-error delta priced at measured remediation cost. Four: cycle time converted to working capital and recovered revenue. Five: one-time costs stated plainly - the audit, workflow build, integration into the existing ERP and CRM stack, and eval-suite construction - amortized over 24 months, not 60.

Two rules keep this credible. First, never claim more than 70 percent of the modeled savings in year one; rollouts go month by month and workflow by workflow, and a model that assumes instant full volume reads as fiction. Second, commit to remeasuring against the model quarterly using production data, and put that commitment in the contract. CFOs do not actually demand aggressive numbers. They demand numbers that survive contact with their own general ledger.

Priced this way, the deployments we operate typically show 45 to 70 percent unit cost reduction at steady state, with payback in 7 to 14 months rather than the fantasy 4. Those are smaller headlines than the FTE-replacement spreadsheet promised. They are also numbers that get signed, funded, and - because they were honest - beaten in production.