Human hours per completed unit of work fall by 65 to 86 percent when an agent takes the steps, but the hours that remain cost three to eight times more each. That is the finding of the agentic AI cost model I built this week across five enterprise processes. The volume tier disappears, the judgment tier stays, and the largest line on the invoice becomes the person who signs. Unit cost falls 40 to 68 percent in four of the five cases and barely moves in the fifth.
New here? I publish one agentic AI governance post every weekday. Subscribe to the blog and it lands in your inbox the moment it goes live.
This follows yesterday’s post on the agentic AI judgment tier, the work that stays with a person once agents run the steps. That post said accountability does not run on a GPU. This one prices it. I modeled the numbers below rather than reporting them: typical handle times, loaded 2026 labor rates, and agent costs at current token prices. Treat them as shape. Plug in your own rates before you treat them as fact.
The four rates in the model
Every case in the agentic AI cost model uses the same four loaded rates. The delivery rate, 20 dollars an hour, prices the volume tier that does intake, entry, triage and routine build. At 60 dollars, the review rate prices the person who checks an agent’s result against a definition of done. Then the specialist rate, 110 dollars, prices the underwriter, the adjuster, the architect and the enhanced-due-diligence analyst. And the judgment rate, 160 dollars, prices model risk, large-loss authority and the harness owner.
The agent line covers tokens, tool calls, retries and the eval run, priced the way I set out in cost per completed task. It is never the price of one model call. It is the whole loop, including the runs that fail.
Loan origination: one funded application
Before agents, a funded file takes about 11.75 human hours across four roles. After, it takes 1.5, and the underwriter still signs.
| Line | Before | After |
|---|---|---|
| Processing at the delivery rate | 8.0 hrs, $160 | none |
| Review and closing at the review rate | 2.0 hrs, $120 | 0.25 hrs, $15 |
| Underwriter at the specialist rate | 1.5 hrs, $165 | 1.0 hrs, $110 |
| Model risk and harness at the judgment rate | 0.25 hrs, $40 | 0.25 hrs, $40 |
| Agent loop | none | $55 |
| Total | 11.75 hrs, $485 | 1.5 hrs, $220 |
Cost per funded file falls 55 percent. The underwriter’s hour barely moves, because the exception is still the exception. What vanished is the eight hours of preparation in front of the signature.
Insurance claims: one simple claim
A simple claim inside authority takes 4.25 human hours before and half an hour after. Agents do the first notice, the photo estimate and the settlement letter. A reviewer checks the band, and a supervisor signs.
| Line | Before | After |
|---|---|---|
| Desk adjusting at the delivery rate | 3.0 hrs, $75 | none |
| Adjuster review at the review rate | 1.0 hrs, $60 | 0.4 hrs, $24 |
| Supervisor at the specialist rate | 0.25 hrs, $28 | 0.1 hrs, $11 |
| Agent loop | none | $35 |
| Total | 4.25 hrs, $163 | 0.5 hrs, $70 |
That is a 57 percent drop on the simple claim. But a large-loss claim takes eight specialist hours before and eight after, so it costs 880 dollars either way, plus 40 for the agent. Blend a book that is 70 percent simple and the saving is about 35 percent. The agentic AI cost model only pays where there was volume to remove.
Customer care: one resolved contact
Customer care is the case that surprised me. Human minutes per contact fall from about 12 to 4, and the blended unit cost falls only 12 percent, from 5.47 dollars to 4.80.
| Line | Before | After |
|---|---|---|
| Routine contact, 70 percent of volume | 8 min at the delivery rate, $2.67 | agent only, $0.80, now 75 percent of volume |
| Complex contact, 30 percent of volume | 20 min, blended $12.00 | 16 min at the review rate, $16.80 with agent, 25 percent of volume |
| Blended per contact | 11.6 human min, $5.47 | 4 human min, $4.80 |
The residual contact is harder and goes to a care expert at three times the old rate. So the per-contact line stays almost flat while the wage bill falls about 45 percent. The return here is quality, fewer transfers and a human on every dispute, and it belongs in the payoff post argument: govern the loop, not the spend.
More on the agentic AI cost model
- The Judgment Tier: The work that stays with a person once agents run the steps, and why it needs a named owner.
- Six Cost Lines: Every agentic AI deployment has six cost lines, and most budgets fund one.
- Sell Judgment, Not Hours: What six agent hours per human hour does to pricing and staffing in IT services.
- FinOps Falls Short: Why the discipline built for cloud spend cannot govern agent spend.
Software delivery: one shipped feature
A mid-size feature took a ten-person pod about 410 human hours. The same feature with agents takes 59, and two of those roles did not exist before.
| Line | Before | After |
|---|---|---|
| Builders and testers at the delivery rate | 320 hrs, $6,400 | none |
| Lead or reviewer | 60 hrs, $2,700 | 35 hrs at the review rate, $2,100 |
| Architect at the specialist rate | 30 hrs, $3,300 | 20 hrs, $2,200 |
| Harness and model risk at the judgment rate | none | 4 hrs, $640 |
| Agent loop | none | $2,500 |
| Total | 410 hrs, $12,400 | 59 hrs, $7,440 |
The saving is 40 percent against a low-rate pod and 73 percent against a pod billed at the review rate. Either way the after column costs the same, because the delivery tier is gone from both. The architect signs the intent, the reviewer signs the pass rate, and the pyramid becomes a diamond.
Know-your-customer: one alert
Alert triage is the purest volume tier in banking, because nine alerts in ten are false positives. Before, an alert took 20 minutes at the delivery rate plus five minutes of second-line review, 11.67 dollars. After, the agent closes 90 percent inside tolerance for 1.20 dollars and no human minutes. The one in ten that escalates takes 25 minutes at the review rate, 26.20 dollars with the agent. Blended, an alert costs 3.70 dollars and 2.5 human minutes, a 68 percent drop. Onboarding a standard client falls 43 percent, from 90 dollars to 51, because the enhanced-due-diligence hour stays with the specialist who signs the file.
What the five cases say together
Three things hold across all five cases in the agentic AI cost model. First, human hours per unit fall 65 to 86 percent, but judgment hours stay within 20 percent of where they were. The exception gets harder once the easy cases are gone, so the underwriter and the architect do not get faster. Second, labor goes from all of the unit cost to 60 or 75 percent of it, with the agent loop at 15 to 35 percent. Third, the biggest single line after is the judgment role’s hours, not the agent.
That last point changes what the agentic AI cost model is sensitive to. Every 10 dollars on the delivery rate moves the before column by 10 to 15 percent. Every 10 dollars on the judgment rate moves the after column by 5 to 8 percent. Agents make the unit cost insensitive to the rate gap and sensitive to judgment pay. The model that used to turn on the cheapest hour now turns on the most accountable one.
What I would do
In my work in regulated loan origination, I price every process this way before I build the agent: the hours by role, the rate by role, the agent loop as a line of its own, and the eval run inside it. Then I name the person on each judgment line, because that line is now the one the CFO will ask about. A model with no name on the signature row is not done.
So do the arithmetic before the pilot. Count the hours the agent removes and the hours it leaves. Price the ones it leaves at what a signature is worth, not at what a seat used to cost. Instructions in, results out was IT, and it was priced by the hour. Intent in, outcomes out is agentic AI, and it is priced by the signature.
Which line in your agent business case is the judgment rate, and does it have a name on it?
Was this post useful? One click tells me what to write next.
Subscribe to the Blog: I publish a new post every weekday on one agentic AI governance question. Enter your email and each post arrives in your inbox the moment it goes live.
For more on this topic, please check out my three books on agentic AI for regulated enterprises:
- Intent In, Outcomes Out: the thesis, why intent in, outcomes out replaces instructions in, results out.
- Earned Autonomy: the method, the EAI-1000 framework that scores how much autonomy an agent has earned.
- The Governed Loop: the architecture, the six layers that decide outcomes, govern the harness, not the model.
- All three books: reading order, formats and buy links on one page.
© Dr. Harish Kotadia, Ph.D., All Rights Reserved, 2026
Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock, and Google Vertex AI.
Disclaimer: This blog post is based on publicly available academic publications, vendor documentation, open standards, and news items from reputed media sources linked above. This post is intended for educational purposes, to help the enterprise agentic AI community build a shared vocabulary from public, authoritative sources.
Views and opinions expressed here are my own and do not represent those of any employer or client, past or present. The analysis presented is my independent interpretation of the published sources linked above and does not constitute legal, financial, or consulting advice of any kind.