AI agent development services: how a custom agent gets built

AI agent development services, stage by stage: what each stage produces, the go/no-go question it answers, what drives cost, EU rules and a vendor checklist.

Published
Reading
7 min
By
Authoricy

AI agent development services turn one job your team does by hand into software that decides the next step and acts in your systems. It works inside rules it cannot break. A good build runs in stages, and each stage ends with a go/no-go question you can answer before you spend more.

Our AI agent development team runs every project this way, starting with a fixed-price discovery.

What you are buying, and when not to buy it

When the project ends, you should hold six things:

  • The code, in your own repository and cloud account.
  • A tool layer: the exact actions the agent may take in each system, with the permission each one needs.
  • The rules, written in code: limits, required fields, who must approve what.
  • An eval set: real past cases with the right answer, used to test every change.
  • A run log that shows what the agent saw, decided and did on every case.
  • A runbook: who owns the agent, what to do when it stops, how to widen what it may do.

If the input looks the same every time, a rule-based flow is cheaper and easier to trust; our guide to AI workflow automation covers that route. If the job lives inside one tool you already pay for, that tool’s own agent may be enough. Build custom when the work crosses several systems, the input is messy free text, or the rules are yours alone. To score candidate tasks first, use the scorecard in AI agents for business.

The build, stage by stage

Discovery takes two weeks at a fixed price. The whole path to a first live agent usually takes 6 to 10 weeks.

Stage What it produces Go/no-go question
1. Discovery Task map, a sample of real past cases, list of systems and access Can we write down what a correct result looks like?
2. Blueprint Tool list with permissions, rules, hand-off points, fixed build price Is every action the agent needs reachable through an API or a safe workaround?
3. Build and eval Working agent, eval set, scores per case type Does it match your team on the eval set, case type by case type?
4. Shadow run Agent drafts, a person sends; differences logged Do its drafts match your team’s on live cases, across the normal mix?
5. Controlled launch Agent acts alone on the safest case types Are errors caught before they reach a customer or the books?
6. Run Monitoring of every run, tuned rules, wider permissions as earned Is someone named as owner, and do they read the log?

A “no” at stage 1 costs two weeks. The same “no” found after launch costs a customer.

One agent, start to finish: order intake

Picture a wholesaler with a small sales office that gets orders by email. Some are typed in the email body, some are PDFs from the customer’s own system, a few are photos of a handwritten list. A coordinator keys each one into the ERP.

Discovery. Read a few months of real orders and sort them by shape. The go/no-go answer is yes, because the coordinator can say exactly what a correct order line looks like.

Blueprint. The agent gets four tools: read the inbox, look up customers, articles and stock in the ERP, create a draft order, and email the customer a question. Confirming an order, changing a price and touching a credit limit stay with a person.

One run, as the agent sees it:

  1. Trigger. A new email arrives with a PDF.
  2. Read. It extracts customer, lines, quantities and requested date.
  3. Look up. It matches each line to an article and checks stock.
  4. Check. Code, not the model, compares prices with the customer agreement and flags any gap.
  5. Decide. Clean orders become draft orders. Anything unclear becomes a question to the customer or a task for the coordinator, with a short summary.
  6. Log. The input, each tool call, the result and the reason are stored.

Build and eval. The eval set holds past orders with the order the coordinator actually entered. We score the agent per order shape. The launch can start with the shapes that pass.

Shadow run and launch. At first the coordinator approves every draft. Once the drafts match what the coordinator would have entered, clean, typed orders from known customers go straight to draft without review. The coordinator still confirms each order. More examples of jobs like this one are in AI agent examples.

What drives the cost

The model is seldom the largest line. Five things move the price of a custom AI agent more:

  • Systems and their access. An ERP with a clean API is quick to connect. One without an API is a project of its own.
  • Input variety. Three order formats are easier to test than thirty.
  • How much the agent does alone. Every action it may take without approval needs rules, tests and a way to undo it.
  • Language. An agent that reads and writes Swedish must be tested in Swedish, on Swedish cases.
  • Running it. Model usage grows with runs and document length.

So every project starts with a fixed-price discovery, and discovery ends with a fixed price for the build.

Why agents stall between demo and production

A demo uses hand-picked cases. Production sends the strange ones. Most stalls trace back to a skipped step:

  • No eval set. Nobody can say whether a prompt change made things better or worse. Stage 3 exists to fix this.
  • Permissions too broad. The agent got write access to everything “for now”. The blueprint names each action instead.
  • No shadow period. The first real mistakes happen in front of customers. Stage 4 lets them happen in a draft.
  • No owner. The agent drifts as products, prices and staff change. Stage 6 names a person.
  • A model is retired. Providers replace models on their own schedule. With an eval set, switching to a new model is a test run, not a rewrite.

EU rules that change the design

The AI Act, Article 50. A system meant to talk directly with people must be designed so they know they are dealing with AI, unless it is obvious. They must be told at the latest at the first interaction (AI Act Service Desk, Article 50). The Act has applied in general since 2 August 2026, and Article 50 falls under that date (AI Act Service Desk, Article 113). In our example, the agent’s questions to customers carry a line saying they come from an AI assistant. A voice agent, such as an AI receptionist for a small business, says it in the greeting.

GDPR, Article 22. A person has the right not to be subject to a decision based solely on automated processing that has legal or similarly significant effects on them (GDPR on EUR-Lex). A draft order is not such a decision. Refusing credit to a sole trader might be. So the blueprint marks where a person must decide, and that point is built as an approval step, not a promise.

Where the data is processed. Choose the model provider’s region before the build starts. With OpenAI’s API, European data residency is chosen when a project is created and an existing project cannot be moved later (OpenAI, data residency in Europe). Sign a data processing agreement with each provider and list them in your records.

Questions to ask any AI agent development company

  1. Who owns the code, the prompts and the eval set when the contract ends?
  2. Can we see the eval set and the scores before launch?
  3. Which actions will the agent take alone, and where is each limit enforced?
  4. What happens when the model provider retires the model you built on?
  5. Where is our data processed and stored, and by which sub-processors?
  6. What does the run log show, and who can read it?
  7. Who owns the agent after launch, and what does that cost to keep?
  8. How do you test quality in Swedish, or in the language our customers write in?

Vague answers to any of the eight are a warning sign. When you want those answers for your own task, book a scoping call about custom AI agent development.

Frequently asked questions

How long does it take to build an AI agent?

A first custom agent usually takes 6 to 10 weeks to reach production, including a two-week discovery, the build, evaluation and a shadow run beside your team. A narrow agent inside one tool can be faster.

How much does AI agent development cost?

It depends on how many systems the agent works in, how varied the input is and how much it may do without approval. Model usage tends to be a smaller share than build and upkeep. Authoricy starts every project with a fixed-price discovery that ends in a fixed price for the build.

Do we need a lot of data to build an AI agent?

No large training dataset is needed, because the agent uses an existing model. What you need is a set of real past cases with the correct outcome, plus access to the systems the agent will act in.

Should we build a custom agent or buy an off-the-shelf one?

Buy when the job lives inside one tool and fits how that vendor sees it. Build when the job crosses systems, the rules are your own or the data must stay in a given region. A chat tool on its own is a different thing, as Is ChatGPT an AI agent? explains.

Talk to the people who build it

Name the task your team dreads.

Book a call and we tell you honestly whether an agent should do it, and what it would take.