Guide
What AI agent development costs, to build and to run
The eight things that move a build quote, the seven lines on a monthly run bill, and a worked example priced from live provider pages.
Published Updated 8 min read
What an AI agent costs to build and run
An AI agent has two prices that behave differently.
The build is a project cost measured in engineering weeks, four to six for a typical first scoped piece, and it is driven far more by the state of your data than by which model you pick. Running it is a recurring monthly bill with about seven line items, and for an internal question-answering agent the model tokens are usually not the biggest one.
Anyone who quotes a build price before looking at your data sources is guessing. That is arithmetic, not a sales tactic: the same brief costs four weeks against a clean database and sixteen against thirteen years of scanned PDFs. What follows is the structure of both numbers, plus a worked monthly model priced from provider pages we read on 4 August 2026, so you can substitute your own volumes and get a figure you can defend.
One line shows up late in most budgets. The part that makes an agent safe to put in front of staff is stopping it from inventing answers. That work is a real line in the estimate, and it is the one cheap quotes leave out.
What actually drives the build price
Eight things move the number. None of them is the model.
| Driver | Cheap end | Expensive end |
|---|---|---|
| Number of data sources | One system, one export | Six systems, three legacy, one with no API |
| State of the data | Clean text or a real database | Scanned PDFs, merged cells, tables split across pages |
| Question types | One shape of question, one field | Twenty shapes, several needing arithmetic across sources |
| Actions | Answers only | Writes, sends, books, pays |
| Where it plugs in | One chat window | CRM, intranet, and a Slack bot |
| Security and tenancy | One team, one permission level | Per-user permissions enforced at retrieval time |
| Evaluation depth | A dozen spot checks | A graded set with "abstain" and wrong-value checks |
| Who operates it | Your team, after handover | The builder, on the hook, with an on-call path |
Two rows do most of the damage.
The state of the data is the one buyers consistently underestimate. Getting numbers out of a scanned twenty-year-old report is not a model problem. It is weeks of extraction and verification work that no amount of prompt writing shortens.
Per-user permissions are the other. An agent that answers from one shared pile of documents is a fraction of the work of one that must return different answers to different people and never leak across the boundary, because the permission check has to happen at retrieval, before the model sees the text.
Use the table as a scoring sheet on your next vendor call. Read the eight rows out loud and ask which end of each one the quote assumes. A bidder who cannot answer has not scoped the job.
The two costs people confuse
Build and run are separate budgets with separate owners. Mixing them is how projects get approved and then quietly starved. The build is one-off and finite. The run cost starts the day the first user asks a question, never stops, and grows with adoption. Approving a build without a run budget produces a working system that gets switched off in month five.
What running it costs each month
Seven lines. Model tokens for generating answers. Embeddings, both the one-off pass over your corpus and the ongoing refresh as documents change. Vector storage. Reranking, which reorders candidate passages before the model sees them. Hosting. Monitoring. And human review time for the answers the system flags.
Here is a model you can copy. The assumptions are stated so you can swap in your own.
An internal question-answering agent for 120 staff, each asking 5 questions on each of 22 working days, so 13,200 questions a month. A corpus of 250,000 documents averaging 1,200 tokens each, which is 300 million tokens, with 5 percent of it changing every month. Each answer sends 10,000 input tokens (system prompt plus twelve retrieved passages) and returns 700 output tokens. Each question reranks 50 candidate passages of 600 tokens.
| Line | Arithmetic | Monthly |
|---|---|---|
| Generation, input | 13,200 x 10,000 = 132 MTok at $3.00 | $396.00 |
| Generation, output | 13,200 x 700 = 9.24 MTok at $15.00 | $138.60 |
| Reranking | 13,200 x 30,000 = 396 MTok at $0.02 | $7.92 |
| Embedding refresh | 15 MTok at $0.02 | $0.30 |
| Query embeddings | 0.26 MTok at $0.02 | $0.01 |
| Vector database | Pinecone Standard minimum | $50.00 |
| Metered total | $592.83 |
Embedding the entire 300 million token corpus once, to stand the thing up, costs 6 dollars. That number surprises people more than any other on this page.
The prices behind those numbers
Every unit price above was read on a live provider page on 4 August 2026.
- Claude Sonnet 5: 3 dollars per million input tokens and 15 dollars per million output tokens from 1 September 2026, with introductory pricing of 2 and 10 dollars in effect until 31 August 2026. Claude Haiku 4.5 is 1 and 5 dollars. Source: Anthropic's pricing documentation. The model above uses the standard rate, not the introductory one, because the introductory rate expires in under a month.
- OpenAI text-embedding-3-small: 0.02 dollars per million input tokens. text-embedding-3-large is 0.13. Source: OpenAI's API pricing page.
- Voyage rerank-2.5-lite: 0.02 dollars per million tokens, with rerank-2.5 at 0.05. Source: Voyage AI pricing.
- Pinecone Standard: 50 dollars per month minimum usage, 0.33 dollars per GB per month for storage, and 16 to 18 dollars per million read units depending on cloud and region. Source: Pinecone pricing.
Where the money actually goes
Now add the two lines with no published unit price. Hosting and monitoring come out of the cloud and observability contracts you already hold, and we are not going to invent a figure for a stack we have not seen.
Then human review. If 5 percent of those 13,200 answers get flagged for a person to check, that is 660 reviews, and at four minutes each it is 44 hours a month. Multiply by your own loaded hourly cost. At any professional rate, that single line is larger than every metered line above it combined.
That is what most cost articles miss. Inference is rarely the biggest number. The biggest numbers are the people who check the output and the engineers who keep retrieval honest as the corpus drifts. A quote that shows you a token estimate and nothing else has priced the cheapest part of the system.
Where the hidden money is
Four things move a run bill in ways nobody budgets for.
Tokenizers change. Anthropic's pricing documentation notes that Claude 4.7 and later models use a newer tokenizer producing roughly 30 percent more tokens for the same text. A per-token price that looks flat can still raise your bill when you upgrade models.
Pinning inference to a region costs more. On Amazon Bedrock and Google Cloud, regional and multi-region endpoints carry a 10 percent premium over global endpoints, per the same page. If your compliance team requires regional routing, that is a real line, not a checkbox.
Server-side tools are billed separately. Anthropic's pricing documentation lists web search on the Claude API at 10 dollars per 1,000 searches on top of tokens, and code execution as free for the first 1,550 hours a month per organization, then 0.05 dollars per hour per container. An agent that searches on every question has a second meter running.
Caching cuts the bill only while the prefix holds. On the same Anthropic page, a cache read costs a tenth of the base input price and a five-minute cache write costs 1.25 times it. Change one byte near the front of your prompt and every request after it pays full price again. Teams see the saving in week one and lose it in week six when someone adds a timestamp to the system prompt.
What money cannot buy
Grounding and citations do not make mistakes impossible. They make them visible and traceable, instead of silent. No budget converts an AI agent into a system that is right by construction, which is why the evaluation and human review lines are permanent rather than a launch expense you retire in month three. Any quote that shows those costs falling to zero after go-live is describing a system nobody is checking.
There is also a real question of whether to build at all. For a common workflow with an off-the-shelf tool that already does it, the honest answer is often to buy, and we cover that trade in build versus buy for AI agents.
How Ilayer prices this
The build is fixed scope, with the number agreed before we start rather than metered by the hour, so the estimation risk sits with us instead of you. After launch there is a flat monthly rate to operate and improve it, and no open-ended retainer. A first fixed-scope piece runs 4 to 6 weeks from kickoff to something working on your real data, and the code lives in your repositories under your license.
We do not publish a price list, for the reason at the top of this page: the same brief is four weeks or sixteen depending on what your PDFs look like. What we do publish is how a fixed-scope engagement runs, step by step, so you can see exactly what you are agreeing to before the number exists. If a workflow needs a guarantee AI cannot give, we say so before you spend a dollar.
Three questions that move a quote more than any other
If you are collecting quotes and they are coming back miles apart, the spread is almost never the engineering rate. It is that each vendor made a different assumption about one of these three, and none of them wrote the assumption down.
What state is the data in? A folder of clean, text-layer PDFs and a folder of scanned faxes are the same sentence in a brief and different quarters of work. Ask each vendor which one they assumed. The honest ones will ask you for a sample before answering.
Who decides what a correct answer is? Somebody on your side has to sit with the test set and say yes or no, and that person is usually the busiest person in the building. A quote that does not name this role has quietly moved it onto you, and it is the single most common reason a build slips.
What happens on the day it is wrong in front of a customer? The answer determines how much review, logging and abstain behaviour the system needs, and those are the lines that separate a demo price from a production price. If a wrong answer is embarrassing, the number is one thing. If it is contractual, it is another.
Send the same three answers to every vendor and the quotes become comparable. Until then you are comparing four different projects that happen to share a title.
Questions people ask
How much does it cost to build an AI agent?
AI agent development cost is a project cost measured in engineering weeks, not a licence fee, and a first scoped piece typically runs 4 to 6 weeks of embedded engineering. Multiply those weeks by whatever day rate you are quoted and you have your build number. What moves the figure is almost never the model: it is how many systems the agent has to read, whether that data is already structured, whether the agent takes actions or only answers, and whether per-user permissions have to be enforced at retrieval time.
What does it cost to keep an AI agent running each month?
A monthly run bill has about seven lines: model tokens for generating answers, embeddings, vector storage, reranking, hosting, monitoring, and human review time. For an internal question-answering agent handling roughly 13,000 questions a month, the metered API and database lines come to under 600 dollars at August 2026 provider prices. Human review of flagged answers is usually the largest single line, and hosting and monitoring depend entirely on the cloud contracts you already have.
Why do AI agent quotes vary by ten times for the same brief?
Because the same sentence describes three different projects. A quote for an agent over your documents can assume clean text in one system or scanned PDFs across six, can include an evaluation set or none at all, can stop at handover or cover operating the thing in production, and can price a demo or a system that survives contact with real users. Ask every bidder the same four questions: which data sources are in scope, how will we prove it is right, who is on the hook after launch, and what is explicitly excluded. Most of the ten-times spread collapses once those answers are on paper.
Keep reading
- 4 August 2026How to stop AI hallucinationsFour mechanisms that cut invented answers, the number to demand for each, and the eight questions that expose a vendor who has never measured one.
- 4 August 2026Build vs buy AI agents, and the fourth option nobody sells youFour routes to an AI agent, what each one costs you in speed, control and lock-in, and the five-step rule that picks between them.