AI agent development company for custom agents that do real work in your systems
JustAutomate is an AI agent development company that designs, builds and runs custom AI agents for B2B companies in the UK, Poland and the DACH region, working remotely and in English. An agent we build gets its own tools for each system it needs, such as the CRM, the finance package or the shared drive, a written list of what it may read and change, and a test set made from your real cases. You see it pass that test before your team depends on it.
- 2 weeksof pilot with a small group on real work, before the agent goes to the whole team
- EUR 3,500 netper week of the process workshop, where the agent's job, tools and limits are agreed
- 29,854operations in 30 days that our own production agent ran with no person in the session
The problem
Why do so many AI agent demos never reach production?
Building an agent that looks good in a demo takes an afternoon. Give a model a few tools, ask it a question it was prepared for, and it answers well. The trouble starts on the second day, when real staff ask real questions: a client name spelt two ways, a deal that sits in the CRM and in a spreadsheet with different values, a request the agent should refuse but tries to complete.
Most of what separates a demo from a working agent is unglamorous. Which records may it read for this person and not for that one? What does it do when a system times out? Which actions need a human to approve them first? How do you know that last week's prompt change did not quietly break the answers about invoices? None of that shows in a demo, and all of it decides whether people keep using the agent after the first month.
So the choice of an AI agent development company is mostly a choice about that second part. Ask any supplier how they test an agent before go-live and how you will see what it did afterwards. If the answer is a demo and a promise, the project is likely to stall at the pilot.
- A pilot that impressed the board and then quietly stopped being used
- An agent nobody trusts with a write action, so it only ever summarises
- Prompt changes made by feel, with no test showing what they broke
- Knowledge of how the agent works sitting with one developer or one supplier
Why now
What makes a custom AI agent practical for a mid-sized company?
Until recently each connection between a language model and a business system was written from scratch. Since Anthropic published Model Context Protocol in November 2024, an open standard for exposing data and actions to a model, an agent's tools can be built as separate modules, one per system, each with its own permissions. Adding the finance system later does not mean rebuilding the part that talks to the CRM.
That changes what custom development costs you. The expensive work is no longer plumbing, it is deciding what the agent should do, what it must never do, and how you will check it. That is where we spend the workshop, and it is the part a generic tool cannot decide for you.
If what you need is a model inside one or two systems rather than an agent that acts across the company, our AI integration services are the smaller step. If you want an agent for company data that is already built, BOND, our business AI agent, is the ready-made route.
What you get
What does a custom AI agent built by us include?
Tools shaped around your processes
Each system the agent touches gets its own tool: look up a client in the CRM, check an invoice status, find a clause in a contract, create a task. The agent works through those tools, not through a general login.
One tool per action, each with its own permission
A written scope of what it may and may not do
Read only, read and write, or write after a human approves. We agree this per system and per action, and the agent inherits the permissions of the person asking.
Scope document signed off before the build starts
A test set built from your own cases
Before the pilot we collect real questions and tasks from your team, including the awkward ones, and record what a correct result looks like. Every change to prompts or tools is run against that set.
Pass rate on your cases shown before go-live
A log of every step the agent took
Who asked, which tools the agent called, what it read and what it changed. When an answer is wrong, the log shows where it went wrong, so the fix is a decision rather than a guess.
Full call log available for audit
Refusal and hand-over rules
The agent says when data is missing, refuses requests outside its scope and passes the case to a named person when a decision is needed. These rules are tested like any other behaviour.
Refusals and hand-overs included in the test set
Documentation you keep
Tools, permissions, prompts and the test set are written down and handed over. Another supplier or your own developer can pick the agent up from that document.
Handover document at the end of the build
Proof
Have you built AI agents that run in production?
Yes. The one we can describe in numbers without a client's permission runs inside JustIdea Agency, the agency our founders run. It was built with the method on this page, and the figures below come from its own call log, together with what they do not prove.
Built for ourselves. These are not client deployments but implementations inside the founders' own companies. We say so plainly, because proof you cannot verify is not proof.
JustBottheir agentJustIdea Agency, a performance marketing agency of around 30 people
- 23
- people from the company work with the agent
- 38
- systems wired in permanently
- 112
- processes run without a human
- 475
- thousand lines of the agent’s code
A custom agent with tools for the CRM, the task system, Google Ads, Meta, Analytics, Search Console, an SEO tool and the data warehouse. Development began on 1 March 2026. Today 23 people at the agency use it.
8 of 38 systems it works in
Google Ads
Meta Ads- HubSpot
- ActiveCollab
- Ahrefs
GA4
Search Console
BigQuery
- Before
- Pulling together what was happening on one client meant several panels, a spreadsheet and a colleague who knew where to look, so cross-system checks happened only when someone found the time.
- After
- Staff ask the agent directly, and a large part of its work now runs on schedules and events with nobody in the conversation. People step in where a judgement is needed.
29,854 of 64,903 operations in 30 days ran with no person in the session
Tool calls made by the agency agent each month of 2026, rising from 83 in March to 62,985 in August as users grew from 4 to 23. An agent that people abandon after the pilot would show a flat or falling line.
Measurement: Agent call log in the data warehouse, March to August 2026, one tool call counted as one operation.
And in your company
Is a custom AI agent the right call for us?
Four doubts UK buyers raise most often before they hire an AI agent development company, and our honest answer to each.
Our CRM already has agent features built in.
Use them if your questions live in that one system. A custom agent earns its cost when the answer needs the CRM and the finance package and a document at once, or when the agent has to act across systems. We will say in the consultation which of the two you are.
You are not based in the UK.
We work with UK companies remotely and in English, the same way we work across Poland and the DACH region. Personal data handling follows UK GDPR and the <a href="https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/" rel="noopener">ICO guidance on AI and data protection</a>, and the processing terms are agreed before any data reaches the agent.
What if the agent does something it should not?
That is why write actions are scoped per system, the riskier ones wait for human approval, and the test set includes requests the agent must refuse. Everything it does is logged, so a mistake can be traced to the step that caused it.
We might want to bring it in-house later.
Good. The handover document covers every tool, permission, prompt and test case, so your developers start from a written description of a working agent rather than from someone's memory.
- Language models
- Claude and GPT models, chosen per task and kept in configuration
- CRM and sales
- HubSpot, Salesforce, Pipedrive, Zoho
- Finance
- Xero, sevDesk, ERP systems with an API
- Documents and team tools
- Google Workspace, Microsoft 365, Slack, Notion
- Data
- BigQuery, Google Sheets, PostgreSQL and other SQL databases
- Workflow layer
- n8n, Make, Zapier, custom API integrations
Four ways to get an AI agent, and where each one stops
Hiring a development company is not always the right answer. This is how we would compare the routes if we were in your seat.
| Route | What you get | Fits when | Where it stops |
|---|---|---|---|
| Agent features inside a tool you already pay for | an agent that knows that one product's data | the questions and actions live in a single system | it cannot read or act in your other systems |
| No-code agent builder (n8n, Make, Microsoft Copilot Studio) | an agent assembled from ready blocks by your team | the logic is simple and someone in-house owns it | testing, permissions and logging are left to you, and it gets hard to maintain as tools are added |
| Custom agent from a development company | an agent designed around your processes, with tools per system, a test set and a call log | the work spans several systems or the agent has to change records, not just read them | it needs one clear use case and data good enough to start; for a single-system question it is more than you need |
| Your own engineering team | full control of the agent and its roadmap | you already employ engineers with language model experience and plan to treat the agent as a product | salaries run before the first result, and the know-how stays with whoever built it |
The routes combine. A common pattern is a custom agent for the cross-system work while built-in features in the CRM stay in use for simple tasks inside it.
How we work
How we develop an AI agent, from first workshop to production
Each stage ends with something you can judge before you agree to the next.
Workshop: the job and its limits
We list the tasks and questions that cost your team the most time and pick one the agent should own first. The output is a written job description for the agent: systems, actions, permissions and what it must refuse.
Test set from your real cases
Before writing a line of the agent, we collect real examples from your team with the correct outcome for each, including edge cases and requests that should be refused.
Build: tools, prompts, permissions
We build one tool per action and system, connect them through ready connectors or custom integrations, and run the test set after every change until the results hold.
Pilot with a small group
Five to ten people use the agent on real work for 2 weeks. We read the call log for wrong answers, missing data and requests nobody expected, and add them to the test set.
Production and review
The agent goes to the whole team with monitoring of failed calls. At 30 and 90 days we look at what people actually ask and decide the next task or system, or that the agent has enough on its plate.
Before and after
What changes once the agent is live
Before
- Cross-system questions wait for the one person who knows where everything is
- Routine updates in the CRM and task system done by hand, or not at all
- AI use limited to drafting text in a chat window
- No record of what the AI was asked or what it touched
After
- Anyone on the team asks the agent, within their own permissions
- Agreed routine actions done by the agent, riskier ones queued for approval
- The agent works on live company data through tested tools
- Every question, tool call and change sits in one log
The pilot group works with the agent 2 weeks after the build. Whether it has become part of daily work shows at 30 days, whether to extend it at 90.
For the person signing this off
What a managing or operations director should ask for
Ask to see the test set and the pass rate before go-live. It is the one artefact that tells you whether the agent works on your cases, not on the supplier's demo, and it lets you judge every later change on the same terms.
Ask, too, where the knowledge of the agent lives. With a written scope, a handover document and a full call log, the agent is an asset of the company. Without them it belongs to whoever built it.
- Job description and permission scope in writing before the build
- Test set from your own cases, run after every change
- Full log of questions, tool calls and changes for audit
- Review at 30 and 90 days, so extending is a decision on usage figures
What it costs
How much does AI agent development cost?
The process workshop has a published price: EUR 3,500 net per week, from one to seven weeks, and one or two are usually enough for a first agent. We quote the build after the workshop, because the cost depends on things nobody knows before it: how many systems the agent needs, how many of its actions write rather than read, and how many exceptions the process has.
Model usage and any workflow platform you pay directly to the provider, with no margin from us.
Process workshop
EUR 3,500 net
per week, one or two weeks for a first agent
A piece of work that stands on its own. It ends with the agent's job description and permission scope, which any developer could build from.
- Tasks and questions that cost the most time, ranked
- The first job for the agent and what it must refuse
- Systems, actions and permissions, written down
- Metrics to check at 30 days
- Most popular
Workshop and custom agent
Quoted after the workshop
the cost follows the number of systems, actions and exceptions
We build the agent described in the workshop, test it on your cases, pilot it with a small group and take it to production.
- Everything in the workshop
- Tools for each system, ready or custom
- Test set from your real cases
- Two-week pilot, then production with logging
- Handover document of tools, prompts and permissions
Agent care after go-live
Quoted after the workshop
for companies that want us to run and extend the agent
We monitor the agent, keep the test set current and add tasks or systems after each review.
- Monitoring of failed calls
- Test set re-run after every change
- Reviews at 30 and 90 days, then on an agreed rhythm
The workshop ends with a document that stands without us: the agent's job, its tools and its permissions. If it shows that a feature in a tool you already pay for covers the need, we say so and the engagement ends there.
Frequently asked questions
What does an AI agent development company do?
It designs, builds and runs AI agents for a specific company: software built on a language model that reads data and carries out actions in the company's own systems through a set of tools. The work covers choosing the agent's job, building the tools for each system, setting permissions, testing on real cases, a pilot with a small group and monitoring in production. JustAutomate does this for B2B companies in the UK, Poland and the DACH region.
How much does it cost to develop an AI agent?
At JustAutomate the only fixed price is the process workshop, EUR 3,500 net per week, and a first agent usually needs one or two weeks. The build is quoted after the workshop, because the cost follows the number of systems, how many actions write data and how many exceptions the process has. Model fees go directly to the provider.
How long does it take to build a custom AI agent?
One or two weeks of workshop, then the build, then a 2-week pilot with five to ten people before the agent goes to the whole team. The build itself depends mainly on whether the systems have documented APIs and how quickly access is granted, and we check both in the workshop before giving a date.
Should we build an AI agent in-house or hire a development company?
Build in-house if you already employ engineers with language model experience and plan to run the agent as a product. Hire a development company if you need the first agent working without hiring, or want a tested design your team can take over later. A handover document of tools, permissions, prompts and test cases makes the second route lead into the first.
What is the difference between an AI agent and a chatbot?
A chatbot answers from content prepared for it in advance. An AI agent decides which tools to use for the request in front of it, reads current data from company systems and, where allowed, makes changes such as updating a CRM record or creating a task. Our guide <a href="/en/guides/what-is-an-ai-agent/">what is an AI agent</a> explains the difference in more detail.
How do you test an AI agent before it goes live?
With a test set built from your team's real questions and tasks, each with a recorded correct outcome, including cases the agent should refuse or hand to a person. The set is run after every change to prompts or tools, and the pilot adds new cases from the call log. You see the results before deciding on go-live.
Do you work with UK companies?
Yes. We work with UK companies remotely and in English. Personal data handling follows UK GDPR and ICO guidance on AI, and processing terms are agreed before any data reaches the agent.
Find out whether a custom AI agent is worth building for you
The free consultation takes 30 minutes. Tell us which tasks eat your team's time and which systems hold the data. We will tell you whether a custom agent makes sense, whether a tool you already pay for would do, and which job the agent should take first. You do not need to share any data at this stage.
We take at most two rollouts at a time, and you get the start date before the contract.
Book a free consultationPS. Before you hire anyone to build an agent, ask how they will test it on your cases. The answer tells you more about the supplier than any demo.