Services
AI Agent Development Services: What to Buy and How to Judge
Checked 2026-10-31
Loading poll...
AI agent development services are projects where a provider builds an agent for you: software that uses a model to plan steps, call your tools and finish a task. A good engagement covers scoping, data audit, build, evaluation, guardrails, launch and handover. The directory is open for listings and names no agency yet, so you can request quotes through one brief.
What you are actually buying
The phrase covers very different work, so start by naming which kind you need.
Kind of project | Example job | What the provider builds |
|---|---|---|
Support agent | Triage tickets, draft replies, look up orders | A loop with helpdesk and order tools, human review before sending |
Coding workflow agent | Review pull requests, write tests, update docs | Rules, skills and MCP tools around a coding agent such as Claude Code |
Operations agent | Reconcile data between systems, chase approvals | Connectors, scheduled runs, an audit trail |
Research agent | Gather and summarise sources on a topic | Search tools, citation handling, review steps |
Customer-facing assistant | Answer product questions on your site | Retrieval over your content, safety filters, escalation |
The first two lines of any proposal should say which row you are in and what finished looks like. If the proposal says "an AI agent for your business", ask again. Our step-by-step guides show what a first version involves: how to create an AI agent, build a support triage agent and build a code review agent. Reading one before you talk to providers makes you a better buyer.
What a good engagement includes
A sound project moves through phases, and each one leaves something behind that you can inspect.
Phase | What happens | What you should receive |
|---|---|---|
Discovery | The job, the users, the data, the risks | A one-page scope with success tests |
Data and systems audit | What the agent will read and change, and how | An inventory of data and tool access, with permissions |
Prototype | A thin version of the loop on real examples | A working demo on your data, not sample data |
Evaluation | A set of real cases with right answers | The test set and a score for each version |
Integration | Connectors, authentication, hosting | Running agent in your environment |
Guardrails | Approvals, limits, logging, kill switch | Documented controls, tested |
Launch | A limited rollout with monitoring | A rollout plan and a review date |
Handover | Docs, code, prompts, access | Everything in your accounts and repositories |
Three of these are often skipped, and they decide whether the agent still works in six months.
Evaluation. An evaluation set is a list of real inputs with the answers you accept. It tells you whether version two is better than version one, and whether a new model breaks something. Models change on the vendor's schedule: Anthropic, OpenAI and others retire models and publish new prices, and a provider without an evaluation set cannot tell you what the change did.
Guardrails. Agents act, so someone must decide what they may do alone. The controls vendors document, such as approvals, sandboxes and managed settings, are covered on enterprise AI coding agent security, and the policy side on AI agent governance. Ask the provider which of these they will set up and test.
Handover. The project is only done when you hold the keys. Code in your repository, prompts in version control, accounts in your name, and someone on your side who can explain how it works.
What drives the cost
We do not quote market rates, because we have no source for them and they differ widely by country and team. What we can give you is the list of cost drivers, which a good provider will price one by one.
Driver | Why it moves the price |
|---|---|
Number of systems the agent touches | Each connector needs building, authentication and testing |
Quality of those systems' APIs | A messy or undocumented API costs more than a clean one |
Autonomy | An agent that acts alone needs more testing, guardrails and monitoring than one that drafts |
Evaluation depth | More cases and more edge cases take more effort, and they are worth it |
Data rules | Zero retention, residency or private hosting narrow the choices and add setup |
Model choice | Token cost grows with volume, see below |
Interface | A chat window, an email flow and a dashboard are different jobs |
Support period | Monitoring and fixes after launch are a continuing cost |
Token cost is the part you can estimate yourself from public prices. At Claude Sonnet 5.5 rates of $2 and $10 per million tokens, a month of 10 million input and 2 million output tokens costs $40, and the same volume on Claude Haiku 5.5 at $0.10 and $0.50 costs $2. That arithmetic is ours and ignores caching. It shows why token cost is rarely the big number in a quote. See LLM API prices and the vendor pages for Anthropic and OpenAI. The engineering and the support period are usually where the money goes.
Ask for the quote in three parts: build, running cost (models and hosting), and support. Then ask what happens to each if volume doubles.
Build, buy or hybrid first
Not every job needs a custom build. If a mature product does the task, buying is often cheaper, and a hybrid of a bought agent extended with your own tools is common. Our guide on build vs buy AI agents gives a scoring sheet. A trustworthy provider will tell you when you do not need them, or when a smaller project will do. Be wary of anyone who proposes a custom build for a job a seat-priced product already does.
How to judge a provider
Because the directory has no listed agencies yet, we cannot point you to anyone, and we will not rank providers. Here is how to judge any provider you find, on our site or elsewhere.
What to look at | A good sign | A weak sign |
|---|---|---|
Evaluation | They describe a test set built from your cases, before building | "We will test it as we go" |
Demo | A working demo on your data in the first weeks | A polished demo on sample data only |
Model changes | A plan for re-running tests on a model switch | No answer |
Ownership | Code, prompts and accounts in your name, in writing | Their repository, their accounts |
Data handling | They ask about retention, residency and training terms first | They have not read the vendor terms |
Scope | A short scope with a success test | "AI transformation" |
Handover | Documents and a walkthrough for your staff | A retainer as the only way to run it |
References | Named clients who will talk to you | Logos with no one to call |
References deserve a note. We will not invent or list any, and neither should you trust a provider page that quotes a client who cannot be reached. Ask for one conversation with a past client and ask what went wrong.
Red flags
- No success test. If the proposal cannot say how you will know the agent works, it will be judged on feelings.
- Autonomy by default. An agent that sends, pays or deletes without review on day one is a risk you have not priced.
- Black-box prompts. Prompts and tools that live only in the provider's system lock you in.
- Secrets in the wrong place. Keys pasted into prompts or committed to code are a sign of weak practice.
- No logging. You cannot debug or audit what you cannot see.
- Fixed price with no change process. Models and APIs change, and a fixed price with no mechanism for it produces disputes.
- Vague data answers. If they cannot say where your data goes, they have not looked.
- Promises of accuracy figures. No honest provider can promise a number before they have seen your cases.
Questions to ask
Topic | Question |
|---|---|
Scope | What does finished look like, and how will we measure it? |
Evaluation | How many real cases will we collect, and who labels them? |
Models | Which models do you propose, and what happens when one is retired? |
Data | Where does our data go, for how long, and is training off? |
Control | What can the agent do alone, and what needs a person? |
Security | How are keys, permissions and logs handled? |
Ownership | Who owns code, prompts, evaluations and accounts at the end? |
Cost | What are build, run and support costs, and how do they scale? |
Support | Who fixes it when an upstream API changes? |
Exit | If we leave, what do we get, and in what format? |
Write the answers into the statement of work. A promise on a call does not bind anyone.
A statement of work outline to ask for
A good provider will send something close to this. If they send a single price and a paragraph, ask for the sections below.
Section | What it should say |
|---|---|
Objective | The job in one sentence, and the success test |
Out of scope | What the agent will not do, so nobody assumes it will |
Data and access | Every system and data class the agent can read or change, and the permission for each |
Models and hosting | Which models, which provider terms, where it runs |
Evaluation | How many cases, who supplies them, the pass mark, how often it is re-run |
Controls | Approvals, spend caps, logging, kill switch |
Milestones | Dated checkpoints, each with something you can inspect |
Ownership | Code, prompts, evaluation set, accounts, documentation |
Change process | How a model retirement or API change is handled and paid for |
Support | Who responds, how fast, for how long |
Exit | What you receive if the contract ends early |
Milestones matter more than the total. A project that has no inspectable output for two months is a project you cannot steer. Ask for a working version on your data early, even if it is rough, and judge the provider by how they react to what it gets wrong.
What usually goes wrong, by project kind
Project | The common failure | The question that catches it |
|---|---|---|
Support agent | It answers confidently from outdated knowledge | Where does the knowledge come from, and who keeps it current? |
Coding workflow | It works on the demo repository and fails on yours | Can we run it on our own repository in week one? |
Operations agent | It acts on a record it misread, and nobody notices | What does it log, and who reviews the log? |
Research agent | It cites sources that do not say what it claims | How are citations checked? |
Customer-facing assistant | It says something your company would not | What filters and escalation rules exist, and how are they tested? |
Every one of these is about the loop around the model, not the model. That is a useful way to read a proposal: count how many of its pages are about the loop and how many are about the model.
Working with the agent after launch
Launch is the start of the running cost. Plan for three routines.
- A weekly look at the logs. A person reads a sample of what the agent did and flags oddities. It is the cheapest quality control there is.
- A re-run of the evaluation on every change. A new model, a new prompt or a changed API should trigger it. Treat a failed run as a blocked release.
- A quarterly review of vendor terms. Retention and training terms change, and so do prices. Our vendor pages carry a checked date for this reason.
If the provider will run these for you under a support plan, ask what the report looks like and who reads it on your side. If you will run them, ask for the scripts on day one.
How to request quotes here
The service listings under /services/ are open, and no agency is listed yet. When providers join, each listing will carry the plan tier it chose, and any paid placement will be labelled and will not change ratings or order. Until then the useful route is the quote form.
- Write a one-page brief. The job, the users, the systems involved, the data rules, the timeline, and a budget range if you have one.
- Say what you will not accept. For example, no training on your data, or no unattended actions.
- Send it once through get quotes. The plan is to match a brief to up to three providers whose listings fit.
- Compare quotes using the table above. Price, evaluation approach, ownership and handover matter more than the headline number.
If you are a provider, see claim your listing and plans and sponsored slots.
Our take
The best predictor of a good outcome is a provider who talks about evaluation and handover before they talk about models. Models are interchangeable inputs that change every few months. The test set, the guardrails and your ability to run the thing yourself are what you keep. If a quote leaves those out, it is a quote for a demo.
We have not tested or ranked any provider, and this page makes no claim about any. The price examples use published token prices as read on 2026-10-31, with arithmetic shown.
Related pages
Get quotes, MCP server development services, AI automation agencies, AI agent frameworks, enterprise AI agents.
Questions
What do AI agent development services include?
Scoping the job, auditing your data and systems, building the agent loop and tools, creating an evaluation set, adding guardrails and logging, deploying, and handing over documentation and ownership. A good engagement ends with you able to run and change the agent.
How much do AI agent development services cost?
It depends on the cost drivers, not a market rate we can quote: how many systems the agent touches, how much evaluation it needs, the autonomy you allow, the data rules, and how long you need support. Ask for a scoped quote with each driver priced.
How do I know a provider is good?
Ask to see an evaluation approach on your data, how they handle model changes, who owns the code and prompts, and a handover plan. A provider who demos on clean sample data only, or cannot say how they will know the agent works, is a risk.
Are there agencies listed on this site?
Not yet. The services directory is open for listings, and we do not name or rank any agency. You can send one brief through the quote form.
Loading poll...
