Service · Web development

Custom AI agent development.

An AI agent does more than answer questions. It carries out tasks on behalf of your employees: retrieving data, proposing decisions, calling tools, choosing next steps. We build domain-specific agents that work on your data, your systems and your rules — model-agnostic.

An AI agent is not a chatbot.

A chatbot answers a question and stops. An AI agent is given a goal — "research this lead", "classify this ticket", "check this invoice against our purchasing terms" — and decides for itself which steps are needed. It consults your documents, calls your APIs, compares options, validates intermediate results and returns a well-founded final result, often with a trail of sources and decisions.

That difference makes building one harder. The agent needs to know which tools are available, when it may act, when it must escalate, and how it accounts for its steps. A chatbot talks; an agent does something that matters. That places hard demands on reliability, observability and cost control. We build AI agents for businesses that do all of this in a production-ready environment — with monitoring per agent run, cost budgets per task and human-in-the-loop for high-impact actions.

Under the hood, agents rely on concepts and platforms such as Anthropic's tool use, OpenAI Assistants, Microsoft Copilot Studio and open-source frameworks like LangChain and LlamaIndex. Which combination fits best depends on your use case, data residency requirements and the models you have access to. We work model-agnostically: Anthropic Claude, OpenAI GPT, Azure OpenAI, Google Gemini, or open-source models such as Llama and Mistral within your own VPC. No vendor lock-in, and no dogma about a single framework.

Not every task suits an agent. For tightly defined, deterministic processes, a standard integration or a rule-based workflow is often cheaper and more reliable. An agent adds value where the work requires judgement, needs to combine data from multiple sources, or where the input is so irregular that a traditional if-then flow can't keep up. In every introductory conversation, we help you make that trade-off honestly.

What makes an agent different.

Tool use
The agent calls your APIs and databases
RAG
Your documents as context, not model training
Memory
Long-term customer and task context
Guardrails
A rules engine checks every output

Three types of AI agent.

Depending on how much autonomy the agent gets and how deeply it integrates with your existing systems.

Flavour 01

Assistant agent (read-only)

Compact project

The agent follows along, looks things up, summarises and makes suggestions. No write actions on your production systems. Ideal as a starting point: low risk, quick value, easy to monitor. Think of a sales research agent that combines LinkedIn, your CRM and public sources into a briefing for a meeting, or a research agent that gathers market signals every week and summarises them for your management team.

RAG pipelineTool useStreaming UIAudit log
Flavour 02

Action agent (with human-in-the-loop)

Mid-sized project

The agent proposes actions and carries them out once a human approves. For example, a support agent that classifies a ticket, drafts a reply with source references and waits for one-click approval before it goes to the customer. Or a purchasing agent that drafts a request for quote and only sends it after review. The review step is built in deliberately: trust is something you build up gradually, not something you switch on.

Multi-step reasoningApproval flowCRM writeNotifications
Flavour 03

Autonomous agent (production-critical)

Larger project

The agent acts independently within tight boundaries. For example, an operations agent that classifies monitoring alerts, runs standard runbooks and only escalates in case of deviations or uncertainty. This requires strong guardrails, observability at every step, a fallback path for when the agent is unsure, and clear limits on which tools the agent may call in which context. Only suitable for tasks where the real-world risk of a wrong action is low or easy to reverse.

Tool orchestrationRules engineObservabilityCost controls

What you get at the end.

A production-ready agent, running in your environment or hosted by us — with everything around it to manage, adjust and extend it.

The agent itself

Production and staging, in your cloud (EU region) or ours. Your own prompts, your own tools.

Codebase and prompts

Full source code, a versioned prompt library, and the tool definitions the agent uses.

Observability dashboard

Insight into which prompts, which tools, which costs and which errors occur per run.

Eval suite

A test set to catch regressions when you change prompts, models or tools.

Management (optional)

Model updates, prompt tuning, cost monitoring, adding new tools.

What we build agents for.

Eight use cases we guide clients through. The common thread: repetitive work that requires judgement, on data that lives in multiple systems.

Sales

Lead research agent

Combines your CRM, LinkedIn, public sources and your own sales notes into a meeting briefing. Can draft a first email and update CRM fields.

Customer support

Support agent

Classifies incoming tickets, finds similar cases and proposes a reply with source references. Escalates automatically when it detects signs of urgency or a complaint.

Operations

Operations agent

Detects anomalies in monitoring data, links them to runbooks and proposes a follow-up action. Generates weekly reports based on your own KPIs.

Compliance

Compliance agent

Reviews documents against internal rules or external standards. Runs sanctions list checks and due diligence steps, and flags deviations for human review.

HR

HR agent

Answers questions about terms of employment using your own documents. Screens CVs against job criteria without using bias-sensitive fields.

Purchasing

Procurement agent

Searches supplier catalogues, compares prices and drafts quote requests. Takes your procurement policy and approved suppliers into account.

Finance

Finance agent

Codes incoming invoices, checks expense claims against your expense policy and flags possible fraud patterns. Writes to your accounting system after approval.

IT

IT agent

Triages monitoring alerts, automatically runs standard runbooks and routes escalations to the right team with the context handed over.

How an engagement works.

01Use-case scoping→ 02Eval & prototype→ 03Build→ 04Rollout & management
1 hour

Use-case scoping

We map out the task: the goal, the data, the tools and the risk boundaries.

A few sprints

Eval & prototype

We build a working prototype on a single model, plus an eval set to measure quality objectively.

A few sprints

Production build

Guardrails, observability, cost controls, RAG pipeline and integrations with your systems.

Ongoing

Rollout and maintenance

A pilot with a small group, followed by a broader rollout. We keep tuning models and prompts.

Not sure about a large project yet?

Test your idea first: a working prototype in 1 day

With OneDayBuild, we make your idea tangible in a single day for €1,150, so you know whether further development is worth the investment. Decide to go ahead with the full build? Then we deduct the cost in full.

View OneDayBuild →

Frameworks we work within as standard.

An agent that carries out tasks on behalf of your organisation touches on privacy, security and compliance. We identify these frameworks at the start and build them into the design — not a tick-box exercise afterwards.

GDPR

Data minimisation and DPIA

The agent only gets access to the data needed for the task. We carry out a DPIA for every project that processes personal data, and we set retention limits in code, not in policy.

EU AI Act

Risk classification upfront

At the start of the project we determine which category your agent falls into — minimal, limited or high risk — and tailor documentation, transparency towards users and monitoring accordingly. For high-risk use cases (HR, government, critical infrastructure) we build in additional audit trails and human escalation points.

Data residency

EU-only where needed

For regulated data, everything runs within an EU region: Azure OpenAI in West Europe or Sweden Central, Anthropic via AWS Bedrock EU, or your own on-prem Llama deployment. No model data, logs or embeddings need to physically leave the EU for the US.

Guardrails

Output validation at multiple layers

The agent can't do everything. A rules engine checks output for forbidden content, tool usage and cost budget. High-impact actions require a second model validation or human approval. We log failures in full, including the prompt, response and tool trace, for later review.

Models and stack we work with.

For each project we choose what fits best in terms of quality, latency, cost and data residency. We are model-agnostic — no religion about a single vendor or framework. For specific smart API integrations between the agent and your systems, we work with the same principles. Where an agent is too heavy for the task, we fall back to a simpler solution, and where your use case goes well beyond a single agent, we guide the broader enterprise AI implementation.

Models
Anthropic ClaudeOpenAI GPT-4o / o1Azure OpenAIGoogle GeminiLlama (on-prem)Mistral (on-prem)
Agent frameworks
LangChainLlamaIndexOpenAI AssistantsAnthropic SDKVercel AI SDKCustom orchestration
Infra & observability
pgvectorWeaviateQdrantLangSmithLangfuseAzure / GCP / AWS

Frequently asked questions.

What is the difference between an AI agent and a chatbot?
A chatbot answers questions based on a model and, optionally, your documents. An AI agent is given a goal and decides for itself which steps to take to reach it: calling tools, making decisions, choosing follow-up steps and checking its own output. It carries out tasks, not just conversations. For a focused question-and-answer flow, an AI chatbot is often enough; for tasks that span multiple steps and multiple systems, choose an agent.
Which AI model will you choose for my agent?
That depends on your task. For complex reasoning and large contexts, we often choose Anthropic Claude Sonnet or Opus. For fast, low-cost work, Haiku, GPT-4o-mini or an open-source model. For strict data residency, everything runs within Azure OpenAI or your own on-prem Llama deployment. We benchmark on your own tasks during the eval phase before we settle on a choice.
Will our data stay secure, and will it be used for training?
With the commercial APIs from Anthropic, OpenAI and Google, data is not used for training by default; we document this per project in the contracts and API settings. For regulated data, we use Azure OpenAI or an on-prem model, so the data never leaves your VPC. We build retention limits, audit logging and data residency into the design.
Can the agent run in an EU region?
Yes. Anthropic Claude is available in an EU region via AWS Bedrock, OpenAI via Azure West Europe / Sweden Central, and Google Gemini via Vertex AI in the EU. For open-source models, we run Llama or Mistral within your own VPC or with an EU cloud provider. No data needs to physically leave the EU.
What does the EU AI Act mean for our agent?
The EU AI Act classifies AI systems by risk. Many business agents (sales, support, operations) fall into the "limited risk" category, with transparency obligations. Agents that support HR decisions or run within government bodies may be "high risk", with stricter documentation and monitoring requirements. We map out the classification at the start of the project so the design is compliant from day one. We already have specific experience with this in AI for municipalities and government.
Can the agent run on-prem without external APIs?
Yes. We use Llama or Mistral as models, with vLLM or Ollama as the runtime, on your own GPUs or in a private cloud tenant. Quality is slightly lower than with the top-tier commercial models, but for well-defined tasks that is often perfectly fine. For heavier agents, we sometimes combine a local model for routine work with a hosted API for the difficult steps.
What determines the cost of an AI agent?
Three factors. One: the build scope. A read-only assistant takes a few sprints, while an autonomous agent with deep integrations takes many more. Two: the runtime costs of the chosen model multiplied by the expected volume. Three: maintenance after go-live. We lay out these three items transparently and manage to a cost budget per agent run during the build. For broader context on AI implementation, see enterprise AI implementation.
How quickly can we go live?
A well-scoped assistant agent can be up and running in a few sprints. An action agent with a human in the loop and integrations with your systems takes several sprints. Autonomous production agents with strict guardrails and compliance requirements need more time in the eval and hardening phase. We work in two-week sprints, so you see a working version every two weeks.
What happens after the MVP?
The agent first goes to a small group of users. We measure quality (eval set + human review), latency and cost. Based on that, we tune prompts, tools and, if needed, the model. Then we widen the rollout. Models change quickly, so we schedule periodic evaluations to check whether a newer model or a cheaper variant can replace the current setup without losing quality.
How do you measure whether the agent works well?
We build an eval suite of representative tasks with expected output. On that set we run automatic scores (correctness, completeness, source citation) plus sampled human review. We also log every production run with prompt, tool calls, model, cost and latency, so regressions after a prompt change or model update stand out immediately. What isn't measured can't be reliably improved, and without evals an agent is always a gamble.

Talk to us about your AI agent.

A no-obligation introductory call of half an hour. We listen to the task you want to automate, ask sharp questions about data, risk and volume, and give you an honest view of whether an agent is the right answer, or perhaps a chatbot, or no AI at all.

Want to start a project?

We'd love to hear from you.

Long story, or would you otherwise have emailed it? Choose Detailed briefing: headings and bullet lists, images in the text and attached files.

Mies

Contact Mies

Business Developer

Contact Martijn

Founder of Appfront

Martijn

Your message has been sent

Thanks for your interest! We'll get back to you as soon as possible, usually within 1 working day.

Edit content