What an AI agent actually costs in year two

The cleanest agent business case in circulation does the maths like this: 10,000 customer service sessions a month, roughly 1,440 hours of capacity returned, a fully loaded rate of $72 an hour, and a headline of $103,680 a month in agent-assisted value. It is honest arithmetic. It is also a year-one model. The cost of running AI agents in year two is decided by five line items that model does not contain: human review, integration maintenance, token and model-tier drift, failed and re-run work, and the price of leaving.

Those five are why re-approval conversations go badly. The pilot delivered. The invoice grew anyway.

Why the cost of running AI agents in year two diverges from the pilot business case

Deloitte's 2025 Tech Value Survey of nearly 550 US cross-industry leaders found that "80% of respondents believe their organization has mature capabilities with basic automation efforts, whereas only 28% believe the same with basic automation and AI agent–related efforts". The same survey found 45% expect basic automation to return within three years, against only 12% for automation plus agents.

Read that as a cost signal rather than a sentiment score. Confidence in payback collapses by roughly three quarters the moment agents are added to the same automation programme — and the five costs below are where that confidence goes.

Cost one: human review scales with the reliability you need, not the headcount you saved

This is the item that most often breaks a year-two forecast, and there is now hard data on why. METR's published benchmark file shows Claude Opus 4.6 completing tasks with a 50% success time horizon of 718.8 minutes — about 12 hours — but an 80% success horizon of only 69.9 minutes. GPT-5.4 shows the same shape: 341.7 minutes at 50%, 53.9 minutes at 80%.

The translation for a finance business partner: a frontier model that can plausibly run a twelve-hour task can only be trusted unattended for about an hour of it. The high-reliability horizon does climb between model generations — in the same file, Claude Opus 4.5 sits at 49.4 minutes at 80% against Opus 4.6's 69.9 — but it climbs far more slowly than the headline number, so review hours fall much more slowly than capability marketing implies. They fall fastest when the reliability threshold you require drops, which in a regulated process it does not.

The instinctools guide states the threshold cleanly as P(success) > T(verify) / T(do): an agent only pays when its probability of success beats the ratio of verification time to doing time. Price the review budget off that ratio, then hold it in the model for year two rather than assuming it decays.

Cost two: token spend and model-tier drift are variables, not a fixed subscription

Raw inference is cheaper than most buyers assume. Anthropic's own worked example prices 10,000 support conversations at roughly $37 of tokens on its small model tier. That is the least interesting number in the stack — and the reason token cost is a poor thing to budget against, because the tier moves.

Anthropic's Economic Index for March 2026 found that "for every additional $10 of hourly wage for a task, the share of conversations using Opus increases by 1.5 percentage points" for Claude.ai users, rising to 2.8 points per $10 on the API. Higher-value work pulls itself onto more expensive models. On Anthropic's published rate card the mid and top tiers cost five and ten times the small tier per token respectively, so a workload that migrates upward multiplies its unit cost without anyone filing a change request.

The same rate card notes that Anthropic's newer models use a tokenizer producing "approximately 30% more tokens for the same text". Your token bill can move on a model upgrade you did not choose.

Cost three: integration and maintenance against systems that keep changing

An agent is only as stable as the CRM fields, permission models and API contracts underneath it, and those change every quarter. The structural problem is that most organisations have nothing documented to maintain against. Microsoft's 2026 Work Trend Index, surveying 20,000 knowledge workers across ten markets, found that even its most advanced cohort reports agent workflows are "documented and repeatable" at organisation level just 25% of the time, against 14% for everyone else.

Undocumented workflows are deferred cost. When the integration breaks in month fourteen, the person who wrote the prompt has moved teams and the fix becomes a rediscovery project. Budget maintenance as a standing percentage of build, not as an occasional ticket. Deloitte's separate survey of 3,235 IT and business leaders across 24 countries found only 21% have a mature governance model for agentic AI — the same shortfall wearing a compliance hat.

Cost four: failed and re-run tasks that nobody meters

Every retry costs tokens twice, occupies a reviewer, and sometimes leaves a half-finished action in a live system for a human to unwind. Almost no business case has a line for it and almost no invoice separates it. Ask for four auditable numbers: attempts versus completions, retries per completed task, escalation rate, and mean human unwind time. If a vendor cannot produce them from telemetry, the cost has not disappeared — it has moved onto your operations team's timesheet.

Cost five: lock-in, priced as the switching cost it becomes

Project NANDA's State of AI in Business 2025 report puts this plainly, warning that in the next few quarters enterprises will lock in vendor relationships "that will be nearly impossible to unwind", and quoting a CIO at a $5bn financial services firm: "Once we've invested time in training a system to understand our workflows, the switching costs become prohibitive." The same report records enterprise RFP-to-implementation cycles running two to eighteen months, so replacement is measured in quarters of internal effort, not in licence fees. It is a preliminary study built on 52 organisation interviews and 153 survey responses, so treat it as a directional warning rather than a market measurement.

Price exit at renewal, while you still have leverage: audit-log export, prompt and configuration portability, connector reconfiguration, and preservation of regulatory evidence. If none of that is contracted, year three is not a negotiation.

How to justify an agent that speeds work up without cutting headcount

This is the question finance committees actually ask, and it is best put by a VP of Procurement at a Fortune 1000 pharmaceutical company, quoted in the same State of AI in Business 2025 report: "If I buy a tool to help my team work faster, how do I quantify that impact? How do I justify it to my CEO when it won't directly move revenue or decrease measurable costs? … that's several degrees removed from bottom-line impact."

Three answers hold up in a committee room, and none of them is a productivity percentage:

• Absorbed volume. Cases, claims or tickets handled at flat headcount through a demand increase — the recruitment you did not do is a real avoided cost with a real salary attached.
• Cycle time on a revenue-blocking step. If underwriting or onboarding goes from four days to four hours, the value sits in the queue you cleared, not in the hours saved.
• Error and rework rate. A measured fall in exceptions is the one number auditors and CFOs already trust, because you were counting it before the agent arrived.

Braivex's position is simple: if none of those three moves in year one, stop rather than re-scope. Microsoft's data shows why programmes drift instead — only 13% of AI users say they are rewarded for reinventing work with AI even when results are not met, so nobody is incentivised to call it.

What outcome pricing changes about the year-two number

Per-seat and per-platform licensing puts every one of the five costs above on the buyer. You pay for logins whether or not the work completes, you fund the review, you absorb the retries, you carry the maintenance, and you pay again to leave. Outcome pricing moves reliability, retry and maintenance risk to the side of the table that controls the model — the only side that can actually reduce them.

There are public rates to price against. Zendesk charges $1.50 per automated resolution, a resolution being defined by a 72-hour quiet period in which the ticket stays closed and the customer does not reopen it. Intercom charges $0.99 per outcome for Fin, billed once per conversation however many questions it answers. Set those against your own fully loaded cost per case and you have a floor and a ceiling for the negotiation.

Deloitte projects the autonomous AI agent market at US$8.5bn by 2026 and US$35bn by 2030, and Foundation Capital sizes the service-as-software opportunity at roughly $4.6 trillion — $2.3tn of salaries plus $2.3tn of outsourced IT and business process services. That is a services budget being repriced, which is why the contract should read like a services contract with measured outcomes rather than a licence with an autorenew.

The year-two re-approval checklist

• Review hours per 100 completed tasks, at the reliability threshold your risk function actually requires.
• Token spend by model tier, trended monthly, with an alert on tier migration.
• Maintenance days per integration per quarter, and whether the workflow is documented well enough for someone else to fix.
• Retry and escalation rates, with human unwind time attached.
• Contracted exit terms: log export, configuration portability, evidence retention.
• One measured business number — absorbed volume, cycle time or error rate — that moved.

Take that list to the vendor before renewal. Any supplier who cannot answer it from telemetry is asking you to fund a second pilot at production prices.

Braivex deploys outcome-priced agents into the systems clients already run, across banking, healthcare, retail, telecom and HR — see how the solutions are structured or browse the marketplace. If you are rebuilding a year-two business case and want these five costs priced against your own volumes, get in touch.