Most published answers to "what does an AI agent cost" give you one range, and one range is the least useful shape the answer can take. An agent that has not been built yet, an agent running in a pilot with ten friendly users, and an agent answering real customers at three in the morning are three different financial commitments. Blending them into a single number hides the only transition that actually decides whether the project succeeds.
So here is AI agent development cost split by stage, with a source attached to every figure, and the running cost worked out from published API prices rather than asserted as a band.
Sourcing note, stated plainly. Two different kinds of number appear below. The build bands are market observations taken from published vendor cost guides. They are secondary sources, and they describe what firms advertise rather than what any project settles at. The run-rate arithmetic is different: it is calculated from a published price list you can open and check. Where a figure is one rather than the other, it says so.
Cost by stage, not by a single range
| Stage | Published range | What the money buys |
|---|---|---|
| Proof of concept | $10,000 to $50,000 | One narrow task, prompt tuning, a couple of integrations, enough to judge whether the idea holds |
| Task agent in a pilot | $50,000 to $120,000 | Tool orchestration, fallback logic, real QA, a small group of real users |
| Retrieval-grounded knowledge agent | $80,000 to $180,000 | Knowledge ingestion, vector storage, semantic search, memory |
| Multi-agent system in production | $150,000 to $400,000 and up | Coordination between agents, task decomposition, resilience, governance |
Ranges as published by Azilen's 2026 agent cost guide, which is one of the more granular of the vendor guides on this question and breaks the same totals down by component.
Regulated work sits above all of it. The same guide quotes healthcare at $150,000 to $400,000 and financial services at $120,000 to $350,000, with the premium going to audit trails, guardrails against hallucination, and the validation that a regulator will ask to see. If you are building in one of those sectors, the general bands above are a floor rather than a range.
The expensive step is pilot to production, and it is not close
This is the number worth planning around. Cost guides citing Gartner report that taking a model from 90% accuracy to the 99% that production normally demands can multiply implementation effort by three to five times, and that a $60,000 proof of concept regularly becomes a roughly $250,000 production system once the surrounding layers are built. We could not reach the original Gartner publication to verify the figure at source, so treat it as widely-repeated industry guidance rather than a primary citation.
The market data says the same thing from the other direction. Camunda's 2026 State of Agentic Orchestration and Automation, a survey of 1,150 IT leaders published in January 2026, found that 71% of organisations say they use AI agents in some capacity, while only 11% of agentic use cases reached production in the last year.
That pairing is worth reading carefully, because it is one survey with one sample rather than two statistics from different studies stitched together, and a lot of the versions circulating online are the stitched kind. Most agent spending is currently buying pilots. The gap between the two figures is not a technology problem. It is the cost of everything that is boring: monitoring, access control, error handling, and knowing what the agent did when somebody asks next quarter.
Budget the production step as its own line with its own approval, not as a rounding error on the pilot. A pilot that cannot be funded through to production is a sunk cost with a demo attached.
The run rate, worked from the published price list
Every vendor guide quotes something like "$1,000 to $5,000 a month" for model usage. The band is real and it is nearly useless, because the figure is set by engineering choices you control rather than by a market rate you accept. Here is the arithmetic, using the published Claude API prices.
Take a support triage agent. It reads a ticket, searches your knowledge base, checks order status, then drafts a reply. That is three model turns, not one, and each turn re-sends the conversation so far. Assume a stable system prompt and tool definitions of 20,000 tokens, about 13,700 tokens of ticket and retrieved content across the three turns, and 770 tokens of output. Run it 20,000 times a month.
On Sonnet 5, at $2 per million input tokens and $10 per million output:
- The system prompt is sent on all three turns, so 60,000 tokens per ticket at $2 per million is $0.120.
- Ticket and retrieved content, 13,720 tokens at $2 per million, is $0.027.
- Output, 770 tokens at $10 per million, is $0.008.
That is about $0.155 a ticket, or roughly $3,100 a month.
Now turn on prompt caching. The system prompt is identical on every single run, so it reads at $0.20 per million instead of $2. The 60,000 tokens per ticket drop from $0.120 to $0.012, and nothing else changes. Per ticket falls to about $0.047, or roughly $940 a month.
Same agent, same traffic, same answers. One configuration decision, and the bill falls by a factor of three.
Model choice moves it again. The identical cached workload costs roughly $470 a month on Haiku 4.5 and roughly $2,360 a month on Opus 5, a five-fold spread across the tiers for work that many agents could route between depending on how hard the ticket is.
Every one of those figures sits inside the "$1,000 to $5,000" band the guides quote, which tells you the band is honest and tells you nothing else. The number that matters is which end of it your architecture puts you at.
Model spend is not the whole bill. Azilen's breakdown adds retrieval infrastructure at $500 to $2,500 a month, monitoring at $200 to $1,000, ongoing prompt tuning at $1,000 to $2,500 and security tooling at $500 to $2,000, landing an all-in production agent at $3,200 to $13,000 a month.
The maintenance line, again
A build quote answers what it costs to deliver. It does not answer what it costs to own. Vendor guides commonly plan around 15% to 20% of build cost per year for retraining, prompt updates and enhancements, and that is on top of the monthly running cost, not inside it.
This is the same separation our 2027 planning worksheet argues for on every software line, and the same one our 2026 guide to pricing custom builds makes for non-agent work. An agent adds a third line that traditional software does not have, because the model layer keeps changing underneath you whether or not you ship anything.
How to actually use these numbers
Deliberately absent above: our own rates. Those are being settled and publishing a half-decided number would be worse than publishing nothing, so the guidance here is about reading other people's quotes rather than selling you ours.
Two things make quotes comparable.
Ask for the stages separately. A single total that blends proof of concept, production hardening and year-one running cost cannot be compared against another vendor's single total, because you cannot see which stages each one included. Insist on the split.
Ask for engineering weeks alongside the currency. Effort is the thing you are actually buying, and it is the one number that survives a change of vendor. Once you have weeks from each bidder, the differences in their totals resolve into either a rate difference or a scope difference, and those are two very different conversations. The vendor questions we would ask covers the rest of that conversation, and handing every bidder one written scope using our ungated RFP template is what stops them each pricing a different project.
If you want to see what the finished thing looks like before committing to a number, our WhatsApp assistant teardown walks through an agent we built and run ourselves, including the parts that turned out harder than planned.
Get the staged version for your own agent
The bands above are a planning tool and they will be wrong for your project the moment you need a real number. Send a one-paragraph description of what the agent should do and we will come back with a stage-by-stage range within two business days, NDA-first if you would rather not describe it in the open. See how engagements are structured, then get in touch.
Common questions
Published 2026 vendor guides put a proof of concept at roughly $10,000 to $50,000, a working task agent at $50,000 to $120,000, a retrieval-grounded knowledge agent at $80,000 to $180,000, and a multi-agent system at $150,000 to $400,000 or more. Regulated sectors sit higher again, with healthcare quoted at $150,000 to $400,000 and financial services at $120,000 to $350,000. These are market observations from vendor price guides, not a quote for a specific project.
Because the last few points of accuracy are the expensive ones. Cost guides citing Gartner report that moving a model from 90% accuracy to the 99% production usually demands can multiply implementation effort by three to five times, and that a $60,000 proof of concept regularly lands at around $250,000 once monitoring, guardrails, access control and edge-case handling are added. The pilot proves the idea works. Production proves it keeps working when nobody is watching.
It is an engineering decision more than a market rate. Worked from published Claude API prices, a support agent handling 20,000 tickets a month with three model turns each costs about $3,100 a month with no prompt caching and about $940 a month with caching on the stable system prompt. The same workload is roughly $470 on Haiku 4.5 and roughly $2,360 on Opus 5. Vendor guides quote a $3,200 to $13,000 a month all-in band once retrieval, monitoring and security tooling are included.
Ask each one to price the stages separately rather than as a single figure, and to state which stage their number covers. A quote that blends proof of concept, production hardening and the first year of running cost into one total cannot be compared against anything. Ask for the build in engineering weeks as well as currency, so you can apply each vendor's own day rate and see where the difference actually sits.
The maintenance line and the running cost. Build quotes answer what it costs to deliver, not what it costs to own. Vendor guides commonly plan around 15% to 20% of build cost per year for retraining, prompt updates and enhancements, and that sits on top of the monthly model and infrastructure spend, not inside it.