The most common mistake in AI deployment economics is treating the inference price as the cost. When a model card says ten cents per response, the decision looks easy: a customer-support task that pays a human five dollars in fully-loaded labor cost looks fifty times cheaper to automate. Run the numbers honestly and the gap is closer to three-to-one, sometimes one-to-one, and on a non-trivial share of tasks the AI loses on cost alone before you even start measuring quality.
This post lays out the framework Wagecore uses to compute the real operational cost of substituting a task with AI. None of it is novel individually — every line item shows up in the post-mortems of failed AI rollouts. The contribution is putting them in one place and committing to numeric estimates so the answer isn't just “it's more than you think.”
The token line is the visible 10%
Take a concrete case. A senior support agent in a SaaS company handles roughly 30 tickets a day at an average of 600 input + 300 output tokens each, across one or two follow-up exchanges. At current frontier-model prices that's on the order of $0.04 per ticket in raw model spend. Across 30 tickets a day, 22 working days a month — about $26 per agent-month in tokens. Against a $7,500 monthly fully-loaded salary, the savings look absurd.
That number is also wrong, because the model is one of nine things that cost money when you actually deploy it. Here's what gets left out.
Oversight
Every AI-handled ticket either (a) auto-resolves with high confidence, (b) gets routed to a human for review, or (c) escalates to a human outright. On day one of a deployment most teams need 100% human review until calibration is solid; mature deployments hold review on the bottom 20–40% confidence band plus a 5% random audit. If a human reviewer takes 45 seconds per audited response and your reviewer pool costs $30/hr loaded, that's $0.38 per audited ticket. Audit 30% of tickets and you've added more cost than the model itself.
Retries
Production deployments don't make one model call per task — they make one to five. There's the initial completion, often a self-check pass, sometimes a critique-and-rewrite loop, and on tool-using agents a planning step plus tool calls plus a summarization. A well-instrumented support agent we benchmark against averages 3.4 model calls per resolved ticket and 8.7 per escalated one. Multiply token cost accordingly.
Error cost
This is the line that breaks more deployments than any other. A confidently-wrong AI answer is not equivalent to a wrong human answer; it is worse, because the customer believes it and acts on it. Refund disputes that resolve cleanly with an apology turn into chargebacks when the AI told the customer their refund was already processed. Account recovery cases where the AI hallucinates a verification step generate support tickets twice — the original case, and the cleanup. Klarna's May-2025 reversal of its 2024 AI-customer-support rollout is the most public case to date: the CEO acknowledged that quality outcomes had dropped and started hiring humans back. Klarna has not disclosed the underlying repeat-rate delta, but the qualitative pattern — cleanup work on complex tickets driving the reversal, not simple-ticket savings — is consistent with what we see in adjacent post-mortems.
We model error cost as a multiplier on the time it takes a senior human to triage the wrong-answer trail and either escalate or repair the relationship. For a customer-facing task the multiplier is typically 2–5× the base resolution time of the same case; for a back-office task with no customer in the loop it's closer to 1–2×.
Integration overhead
The AI doesn't read tickets from a Word document. It reads them from a CRM via an API, with auth, rate limits, schema versioning, and a retrieval layer over the company's knowledge base. That layer needs engineers to build and maintain. Amortized across the ticket volume of a single team, a serious integration effort runs $20–60k in initial build plus 10–30% of an engineer's ongoing time. On a 50-agent team that's roughly $1.50 per ticket in steady state, in our calibration.
Orchestration & vendor lock
Multi-model setups, fallback chains, prompt-template registries, eval infrastructure. None of this is free. We bucket it conservatively at $0.20–0.80 per resolved ticket depending on company stage. Strong eval infrastructure pays for itself, but the AI-cost line item still shows up.
Compounding the line items
With those five concrete additions and reasonable midpoint assumptions — 30% audit rate, 3.4 model calls per resolved ticket, 8.7 per escalated, 20% escalation rate, error-cost multiplier of 3× on the 12% of cases that go wrong — the support example moves from $26/agent-month in tokens to roughly $1,800/agent-month all-in. That's still cheaper than the $7,500 human, but the ratio is 4-to-1, not 290-to-1. And the math gets worse as you move up the value chain. For roles where wrong answers cause real damage — financial advice, medical triage, legal review — the error-cost line dominates and the deployment loses on cost before you even count salary.
The pattern is general: as task complexity goes up, the inference-cost line stays roughly flat (longer prompts, more context, but not 10× more) while every other line item scales superlinearly. Audit takes longer because reviewers have to actually read the case. Retries multiply because the model needs more steps to handle the case. Error cost explodes because the cases that go wrong are the ones with the most at stake. By the time you're looking at senior knowledge work, the operational cost is almost entirely human-time-around-the-AI, and the model has become the cheapest component of its own deployment.
Where AI actually wins on cost
Three task profiles consistently come out ahead under this kind of accounting:
- Bounded, repetitive, low-stakes. Categorization tasks where being wrong is cheap (e.g. routing an internal email). Audit rates can be low, error cost is minimal, integration is shallow.
- Drafting under human review. The AI produces the first version, the human takes it the last 30%. Both lines of cost (model + human-review) stay bounded because the human was going to look at it either way.
- Aggregation and search. Surfacing the relevant docs, summarizing yesterday's tickets, retrieving the right policy. The AI replaces a search interface, not a worker, and replaces it well because retrieval errors usually surface fast.
Each of these maps cleanly to a substitution-class in the Wagecore taxonomy: AI-augmented (drafting), human-led + AI-assisted (aggregation), and a narrow band of true replaceable work (the bounded-low-stakes case). Outside those, the math says hold.
What changes the answer over time
Three things move the operational cost line:
Inference price. Token cost has dropped roughly 10× every 18–24 months for comparable capability. This shifts the model line but doesn't touch audit, retries, or error cost — so for high-stakes tasks it barely changes the verdict.
Eval and orchestration tooling. Better evals shrink the audit-rate component meaningfully; this is currently the highest-leverage line to optimize. Going from 30% to 10% audit rate on a mature deployment is a real cost change.
Liability and regulatory regime. When an AI is the legal record-keeper, the error-cost multiplier goes up. When the AI is used as decision-support with a clear human in the loop, it goes down. This is the line that moves on policy, not on technology.
The bottom line
Pricing AI deployments off the model card is the equivalent of pricing a car by its sticker and ignoring fuel, insurance, depreciation, and the person you have to pay to drive it. Operational cost matters because it is what determines whether a deployment survives the first six months. The roles where AI is “3–10× cheaper than the human” in practice are the roles where the demo was honest about its scope. Most roles, especially the ones the discourse keeps targeting, look much more like 4-to-1 — real savings, real value, but not a replacement, and not a free one.
Wagecore computes the version of this calculation for individual roles, using the same operational categories laid out here. If you want to see what the math looks like for your work specifically, the wizard runs in two minutes and the methodology is published. You can also read the methodology and disagree with our line-item estimates — we update them quarterly based on what the data says.