What AI agent and RAG development costs
There are four packages, each with a from-price in euros. After I read your brief I give a range; after we have written the scope down, a fixed price. This page shows what each package includes and what moves its price.
- Currency
- Euros
- How prices work
- A from-price, then a fixed price for a written scope
- Model usage
- Billed to your own provider account
at a glance
The schedule
packages
Package by package
from
€1,500
LLM hardening audit
A review of an existing LLM feature for prompt injection, runaway spend, missing logs and unmeasured quality, with fixes ranked by risk.
included
- A review of one LLM feature in one Python codebase.
- A written report: findings ranked by risk, each with its evidence and a suggested fix.
- A one-page map of where text enters and where output goes.
- An evaluation baseline of 20 to 40 cases and a script that scores them.
- A fix list in order, and a readout call.
what raises the price
- More places that call a model, and more prompts.
- A model that can use tools or write to other systems.
- A larger codebase around the feature.
- No staging environment, so one has to be set up first.
- A larger evaluation baseline.
- A backend that is not written in Python: ask me first.
from
€4,000
RAG assistant
A question-answering assistant over your own documents that cites its sources, refuses when they do not support an answer, and is measured against your real questions.
included
- An assistant over one set of documents, as a simple web chat or an API.
- Documents split along their own structure, each piece with its citation label.
- Cited answers, and a refusal when the documents do not support one.
- A question set of 30 to 40 of your real questions, and a scored report.
- Logging of every question and its outcome, and spend caps.
- Tests, documentation and a handover session.
what raises the price
- More documents, or messier ones: scans, tables, mixed formats.
- Several sources, or syncing with a live one such as a shared drive or SharePoint.
- Different people seeing different documents.
- More than one language.
- Embedding it in your site or chat tool.
- Higher stakes: a review step or a second-model check.
from
€7,500
Agent workflow build
An AI agent wired into the tools your team already uses, which drafts the work and waits for a person to approve anything that matters.
included
- One agent workflow, connected to the tools it touches.
- An approval step that records who approved which version.
- Dates and rules in plain code; model output checked against a fixed format.
- Each write to your tools happens once, even when a call fails or repeats.
- Spend caps, and untrusted text kept out of the instructions.
- Tests, scenario evals, documentation and a handover session.
what raises the price
- More tools to read from and write to.
- More cases and exceptions the workflow must handle.
- An approval screen built for you, rather than approving inside a tool you already use.
- Reacting to changes in your tools, not only writing to them.
- Several roles with different approval rights.
- A tool with a poor or no API.
from
€1,800 a month
Retainer
Ongoing improvements, monitoring and evals for a system already in production, with a fixed number of days each month.
included
- A fixed number of days each month, set in the proposal.
- Watching spend, errors and refusals.
- Evals re-run after any model, prompt or document change.
- Small improvements and fixes.
- A short written note each month: what changed and what I recommend.
what raises the price
- More days each month.
- More systems in production.
- A system I did not build. I start with the audit.
included
What every engagement includes
- Code in your repository
- Everything I write lives in a repository you own, not in mine.
- Tests
- Automated tests for the code I write.
- Evals, where they apply
- Agents and document assistants are scored against your real cases before and after every change. The audit leaves you a baseline to re-run.
- Spend caps
- Every model call is counted against a cap enforced in the code, set to the budget we agree.
- Documentation
- How to run it, configure it, change it, and what to watch.
- A handover session
- A session with the people who will own it.
- Fixes for 30 days
- I fix defects in what I built for 30 days after handover at no charge.
model costs
How model and API costs work
Model usage is billed by the provider, whether OpenAI, Anthropic or another, to your own account. It does not pass through me and carries no mark-up. You hold the account and the keys, so you can see every charge and set a limit at the provider as well.
In the code, every model call is counted against a spend cap before it is made, and the cap is set to the budget we agree. Before we start I estimate the monthly cost from your expected volume.
Hosting is a separate bill from whoever hosts the system. Where it should run is part of the scope we write down.
payment
Payment terms
These are my default terms. If your finance process needs something different, say so in the brief.
- Fixed price
- Builds and audits are one fixed price for a scope written down before work starts.
- Part up front
- Half is invoiced when we agree the scope, to book the time. The other half is invoiced at handover.
- Retainers
- Invoiced monthly in advance. You can stop with one month’s notice.
- Changes to scope
- Quoted before I do the work, never after.
- Currency
- Invoices are in euros.
Questions people ask
Why are these “from” prices?
Scope varies more than a price list can show: ten documents or ten thousand, one tool or four. A from-price is the smallest realistic project of its kind. After your brief I give a range, and once the scope is written down, a fixed price.
How do I get an exact price?
Send a brief. I reply within two working days with a rough scope, a price range and whether AI is the right tool for the job at all. If it is, we talk, write the scope down together, and I send a fixed price.
Do you charge by the hour?
Not for builds and audits: those are one fixed price for a written scope. A retainer buys a fixed number of days each month. I do not bill open-ended hours.
Can I start smaller?
Yes. The audit is the cheapest way in, and a document assistant over one set of documents is the smallest build. What is built for the first piece, such as the question set, the spend caps and the logging, carries over to the next.
What if AI is the wrong tool for my problem?
I say so before you pay anything. The reply to your brief includes that judgement, and if a spreadsheet formula or a no-code automation does the job, I will tell you.
Who pays for the model usage?
You do, directly to the provider, from your own account, with no mark-up from me. The spend cap in the code is set to the budget we agree, and you can set a limit at the provider as well.
What if we stop partway through?
You pay for the work done up to that point and keep the code written so far.
What is in the retainer?
A fixed number of days each month, agreed in the proposal. They go on watching spend, errors and refusals, re-running the evals after any model, prompt or document change, and small improvements. Each month I send a short written note on what changed and what I recommend.
Send a brief and I will price it.
Send a short brief: what the work is, who does it now, and what it costs you. I reply within two working days with a rough scope, a price range and whether AI is the right tool for it at all.
Send a briefOr write to keel.agents@gmail.com.