.png)
Every AI automation proposal leads with a number designed to look affordable. A platform license. A per-seat price. A per-agent fee. Sometimes, with a public model, something close to free. That number is real, and it is also the smallest one you will pay. The cost that decides your budget is everything the sticker price does not mention: the team you have to hire to run it, the maintenance that never ends, and the token meter that runs on every transaction and grows with volume.
Pricing an AI deployment on its license is like pricing a car on the key fob. The interesting money is everywhere else. Here is the rest of the iceberg, layer by layer, because the option that looks cheapest on the quote is usually the most expensive one to actually run.
The first hidden cost: the team you didn’t know you were hiring
The model is nearly free to call. The people who turn it into something production-grade are not.
If you build on public AI agents in-house, the do-it-yourself path, the sticker is tiny and the staffing is enormous. To reach production you need AI and ML engineers, MLOps, data engineers, and AI governance specialists, plus the security, testing, and monitoring work that never appears in a demo. By market salary estimates, a serious in-house build runs north of 1.5 million dollars a year in people before a single token is spent [1].
Across our deployments, we’ve found the salary cost is rarely the deciding factor. The larger question is whether it makes economic sense for your technology organization to build and operate AI. For many mid-market organizations, AI is not the product they sell. Every machine learning engineer hired to build and operate AI is one less engineer improving the ERP, modernizing customer-facing systems, strengthening cybersecurity, or delivering the technology initiatives that directly create business value.
Agentic platforms hide the same cost more cleverly. The license, commonly 500 to 800 thousand dollars a year [1], is the entry fee, and you still hire the automation architects, developers, and support engineers to operate the thing. You buy the software, and then you still need the people to build, operate, govern, and continuously improve it. The platform license is only one component of the total cost of ownership.
The second hidden cost: maintenance that never ends
You do not buy automation once. You rent a tuning problem for as long as you run it.
This is RPA’s open secret. Roughly 70 to 75 percent of the total cost of an RPA program is maintenance rather than development [2]. Every time a vendor changes a screen, a bot breaks. Every new exception needs a new rule. Every additional automation deepens the dependency on the specialists who keep the rules patched.
The program that looked like a one-time project quietly becomes a permanent payroll line, and the maintenance share grows with every bot you add. Any automation built on brittle, rule-based logic inherits a version of this tax, and it compounds in the direction you least want: up and to the right, forever.
The third hidden cost: the token meter nobody is watching
This is the one that looks trivial in a pilot and detonates in production, and it is the cost most proposals never model at all.
Every reasoning step, every retry, every document an agent reads burns tokens, and agentic AI is a glutton. Gartner found in 2026 that agentic AI consumes 5 to 30 times more tokens per task than a standard chatbot [3], because agents plan, call tools, observe results, and loop. On ten clean test cases that meter reads a few dollars. At production volume, across multi-step workflows and background agents that run around the clock, it becomes one of your largest and least predictable operating costs.
The trap that fools finance teams is simple. The price per token has been falling fast, with the blended cost of enterprise inference dropping around 67 percent in a single year [4], so leaders assume the bill will shrink on its own. It does the opposite, because consumption is growing faster than price is falling. More agents, longer context windows, retry loops, and always-on background tasks outrun every price cut.
The result is that 73 percent of enterprises exceeded their original AI cost projections, and the FinOps Foundation has documented organizations running three times over their token budgets [5]. Without disciplined FinOps and orchestration that routes cheap, high-volume steps to small or local models and reserves frontier models for the hard reasoning, the meter runs unbounded. This is the practical reason to never wire your operation to a single expensive model: token strategy is cost strategy.
The full picture across the options:
| Option | Sticker price | The team | Maintenance | Token / compute | What you actually pay |
|---|---|---|---|---|---|
| RPA | Low license | Specialists to tune rules | 70-75% of total cost | Minimal | License plus an endless tuning bill |
| Public AI agents (DIY) | Near zero to call | $1.5M+/yr in AI specialists | High, self-owned | Unbounded, 5-30x per task | A hiring plan disguised as a tool |
| Agentic platform | $500-800K/yr | Full team to operate it | Ongoing, on you | You manage and absorb it | License plus a department |
| Managed / outcome-based | Priced to the result | Carried by the provider | Carried by the provider | Optimized by the provider | One number, tied to value delivered |
What total cost of ownership actually means
Total cost to run, at production volume, over the life of the deployment, with the people included. Not the entry price.
The honest way to compare AI options is to price all four layers together: license plus team plus maintenance plus tokens. Do that arithmetic and the ranking usually flips. The “cheap” or “free” options reveal themselves as the most expensive, because the money moved from the invoice you can see to the payroll and the meter you did not model.
And the option that looks like a premium, a managed or outcome-based model where you pay for a delivered result and someone else carries the team, the maintenance, and the token optimization, is frequently the lowest true cost, because those three hidden layers are the provider’s problem to solve efficiently rather than yours to absorb blindly. This is the sharper version of the build-versus-partner question: it is not only about speed, it is about who eats the three-quarters of the cost that never makes it onto the first slide.
So the next time an AI proposal lands, do the math the vendor left out, at real volume, across a few years. The license is the entry fee. Ask what the full season costs, because that is the number your CFO will actually live with, and it is the one we put on the table before anyone signs anything.
Sources
- Market salary and platform-pricing estimates (Glassdoor and market data): an in-house AI build runs $1.5M+/yr in specialists; an agentic platform license commonly runs $500-800K/yr plus an operating team.
- HfS Research: 70-75% of total RPA program cost is maintenance rather than development.
- Gartner, 2026: agentic AI consumes 5-30x more tokens per task than a standard chatbot.
- Industry inference-cost analysis, 2026: blended enterprise inference cost fell ~67% year over year, even as consumption outran the savings.
- FinOps Foundation, State of FinOps 2026: 73% of enterprises exceeded original AI cost projections; some ran 3x over their token budgets.