AI Agents for Business: Build In-House or Hire a Developer

The basics

What you get when you build AI agents for your business

An AI agent is a program that pursues a goal instead of only answering a question. It reads inputs such as emails, PDFs or forms, uses a language model to understand them, calls tools and systems and files the result where your team carries on working. Where a decision has consequences, it stops and puts it to a person. That takes its capabilities further than chatbots and AI assistants, which wait for every input.

If you want to tackle several workflows at once, our page on AI automation for business gives you the overview of use cases, fit and safeguards. This page covers the question before that: how a single agent is built, who builds it and what you need for it.

Three terms, one difference

·Chatbot: answers questions from a fixed knowledge base and stays in the conversation.
·AI assistant: helps one person write or research and waits for the next request.
✓AI agent: completes an entire workflow across several steps and systems, with approval at the points you define.

Guide: what AI agents deliver in the enterprise today →

What holds companies back · Destatis 2025

·26 percent of enterprises in Germany used artificial intelligence in 2025, 36 percent of those with 50 to 249 persons employed.
·Another 20 percent had considered AI but not yet introduced it.
·Of these, 72 percent named a lack of knowledge as a reason, 62 percent unclear legal consequences and 60 percent data protection concerns. Costs that were too high were named by 32 percent.

Source: Federal Statistical Office of Germany (Destatis), survey on ICT usage in enterprises 2025, tables on AI usage and on reasons against using AI (in German), as of 24 November 2025.

So the technology is rarely the hurdle. Knowledge, law and data protection are, and these are exactly the three points this page answers: which building blocks an agent consists of, when you can build it in-house and which framework the EU AI Act and the GDPR set.

Building blocks

The four building blocks of every AI agent for business

Whether you build in-house or hire a developer: every agent that holds up in daily work consists of the same four parts. If one is missing, the agent stays a demo. The questions under each building block are the ones we ask in every intro call.

1 · Task and goal

An agent gets exactly one workflow with a clear start and end. Which input triggers it, which result does it deliver, how can your team tell in seconds that it is right? The narrower the task, the sooner the agent holds up.

2 · Access and data

The agent needs read access to its sources and, where necessary, write access to your systems: inbox, file storage, CRM, ERP, knowledge bases. It only gets the rights the responsible staff have, and only the company data the purpose requires.

3 · Rules and approvals

What may the agent do alone, what only with approval, when does it hand over to a person? Payments, commitments to customers and contract changes only go out with approval. We write these rules down with your business team, not in IT alone.

4 · Operation and measurement

Every step is logged with its source. Before the start we set the metric, such as hours per week, turnaround time or error rate, and check it in operation. A named owner maintains rules and examples whenever the workflow changes.

A multi-agent system, in which several agents work together, follows the same building blocks, just per agent and with orchestration on top. To get started, a single agent with one clear task is almost always enough.

Getting started

Which AI agents for business make a good start

Among the many types of AI agents, the best first ones have frequent digital input and a result your team can check quickly. The first use case should be small enough to hold up within weeks and frequent enough to pay off. These are the four we build most often.

Invoice intake

The agent reads incoming invoices from inboxes and portals, matches them against purchase order and delivery note and proposes the booking. Your accounting team approves.

AI invoice processing →

Sorting enquiries and emails

Shared inboxes are sorted by request, urgency and owner, and standard questions get a ready-made draft reply.

AI agents as a service →

Monitoring tenders

A research agent screens procurement portals and funding databases daily, scores matches against your criteria and puts deadlines into the calendar.

Tender monitoring with AI →

Customer service

The agent answers recurring questions from your own knowledge base and hands special cases to your team with the full history. That takes load off the team and raises customer satisfaction.

Customer service with AI →

Decision guide

Build in-house or hire an AI agent development company?

There are four ways to your own AI agent. They differ less in the AI than in who runs the agent afterwards, maintains it and answers for its mistakes. The table puts them side by side.

Four ways to build an AI agent, compared by effort, time, risk and fit.
Approach	Effort for your team	Time to live operation	Risk	When it makes sense
In-house, with an agent builder from a large software vendor	Low: write instructions, connect documents as a knowledge source	Short, as long as the agent only reads and drafts	Rights and data flows follow the vendor's defaults, the agent stays tied to that environment	Personal helpers for research, summaries and drafts
In-house, with a no-code workflow platform	Medium: assemble workflows, set up interfaces, catch error cases	Medium, depending on the number of connected systems	Operation often depends on one person, you build logging and approvals yourself	Tech-savvy team, few systems, no decisions with consequences for customers
In-house code, with programming libraries for agents	High: development, testing, hosting, security	Long, and maintenance stays in-house for good	Knowledge depends on a few developers, models and interfaces change constantly	Own development department, agents as part of your own product
Hire Weidner & Friends	Low: one subject-matter owner, real examples, approvals in the first weeks	First measurable effects after four to six weeks, live operation after around twelve weeks	Low: our documentation of every agent is set up so that your IT or another provider could take over	The agent writes to core systems, handles customer data or is meant to run daily and measurably

Build in-house if

✓the agent only reads, summarises and drafts,
✓no personal data of third parties is involved,
✓your team has the resources to maintain it,
✓an outage would not hold anyone up.

Hire a development partner if

✓the agent writes to your ERP, CRM, accounting system or inbox,
✓customer or employee data is processed,
✓several systems and many exceptions come together,
✓the workflow is meant to run every day and be measurable.

How to spot a trustworthy partner →

The two do not exclude each other. Many organisations start with their own assistant for individual staff and have the first agent with system integration built. The know-how from their own attempt makes the collaboration faster afterwards.

Tools

What you can build your own AI agents with

Besides general AI tools for writing and research, there are three classes of tools for building agents. Which one fits depends on how deeply the agent reaches into your business processes. We work model-agnostic with different AI models and choose the tool by the workflow, not the other way round.

Agent builders from large software vendors

Built into office software and collaboration platforms. On these AI agent platforms you describe the task in plain language, often start from a template, connect documents as a knowledge source and share the agent with your team. Strong for beginners and for helpers at individual workplaces.

No-code workflow platforms

They connect inbox, file storage, CRM and spreadsheets through ready-made building blocks and call an AI model at the right points. Many are available as open source to run on your own servers in the EU. Strong for workflows across several systems.

Programming libraries and frameworks

For developers who build agents with their own logic, memory and orchestration of several agents. Full control over data flows and hosting, at the price of your own effort for testing, security and maintenance.

Anyone starting with their own AI agents soon notices: whatever the tool, the quality of instructions and examples decides. As a rule, an agent does not learn your workflow by training its own model, but through clear rules, good examples from your daily work and your team's feedback in the first weeks.

Process

How we build your AI agent: five steps with realistic timings

The weeks apply to an agent with one bounded workflow. After each step you know where you stand and decide whether to continue.

01
Process review · weeks 1–2
We walk through the workflow with the people who do it today, collect real examples and agree on the goal, the metric and the approvals.
02
Access and data protection · weeks 2–3
Systems, rights and data flows are clarified, data processing, EU hosting and deployment settled before the agent sees real data.
03
Build and test · weeks 3–4
The agent is built and tested on your examples, especially the difficult ones. Exceptions get a fixed hand-over to your team.
04
Suggestion mode · weeks 4–6
The agent works with real input and your team approves every proposal. This is where the first measurable effects show.
05
Live operation · from around week 12
Once the agent runs reliably, it handles more steps on its own. We measure against the targets from step 1 and plan the next workflow.

Everyday example · agent for a shared inbox

08:12Enquiry received via the contact form, three requests in one message
08:12Requests separated: quote, delivery time, question about an invoice
08:13Customer found in the CRM, open order and latest invoice matched
08:13Draft reply on delivery time created, quote request passed to sales
08:14Invoice question handed to accounting with the document attached
09:02Draft checked by customer support and sent

What you contribute

✓one subject-matter owner who knows the workflow,
✓real examples from daily work, including the difficult ones,
✓access to the systems the agent needs,
✓approvals in the first weeks, until you hand more over to the agent.

Moving AI pilots into production →

Cost

What it costs to have an AI agent built

Cost depends less on the AI model than on your system landscape: how complex the integration is, how many systems are connected, how clean the data is and how many exceptions the workflow has. A pilot for one workflow is usually in the five-figure euro range with us, plus running costs for operation and model usage.

Our cost guide for AI agents lists the typical setup ranges: simple rule-based or FAQ agents between 3,000 and 10,000 euros, process agents with system integration between 10,000 and 40,000 euros, multi-agent systems from 50,000 euros. Building in-house saves the provider and costs working time for building, testing and maintenance.

Before we start, we set the benefit against it: hours per week, turnaround time, error rate. So you know before delivery whether the agent pays off. Public funding for digitalisation in Germany and the EU can cover part of the cost.

What the EU AI Act and the GDPR require from AI agents for business

Two sets of rules apply to agents used by companies in the EU: the European AI regulation, known as the AI Act, and the GDPR. For typical office agents the obligations are manageable if they are settled before the start. This overview reflects the law as of 1 November 2026.

Provider and deployer

The AI Act, Regulation (EU) 2024/1689, distinguishes providers and deployers. A company that develops an agent, or has one developed, and puts it into service under its own name for its own use can be both under Article 3(3) and (11). Which obligations follow depends on how the agent is classified.

AI literacy under Article 4

Since 2 February 2025, providers and deployers take measures on the AI literacy of their staff. Since 27 July 2026, the version amended by Regulation (EU) 2026/1744 applies: the measures remain mandatory, but nobody has to guarantee a specific level of AI literacy.

Transparency under Article 50

If an agent interacts directly with people, for example in customer service, it must be designed so that they are informed they are interacting with an AI system, unless this is obvious. This obligation applies since 2 August 2026.

High-risk only in certain areas

High-risk obligations hit agents in the areas listed in Annex III, for example when they screen job applications or evaluate employee performance. Following the amendment by Regulation (EU) 2026/1744, they apply there from 2 December 2027. An agent for invoice intake or enquiries usually falls into none of these areas.

GDPR: people decide

Under Article 22 GDPR, decisions with legal effects for a person may in principle not be based solely on automated processing. Data processing agreements, EU hosting and limited access rights come on top.

GDPR-compliant AI agents →

This overview is not legal advice. We record the classification, data flows and approvals of your agent in writing during the process review, so that your data protection officer and, where applicable, your works council can check them.

FAQ

Questions about AI agents for business

How do I build an AI agent for my business?

Start with one narrowly defined task, then settle the access and data the agent needs, the rules and approvals for anything with consequences and the metric for live operation. Agents are built with an agent builder, a no-code workflow platform or programming libraries. For agents that connect to your systems and handle customer data, an experienced partner pays off.

How long does it take to have an AI agent built?

For one bounded workflow, first measurable effects usually show after four to six weeks, and live operation is in place after around twelve weeks. In the first weeks the agent works in suggestion mode and your team approves every step.

What does a custom AI agent cost?

A pilot for one workflow is usually in the five-figure euro range with us, plus running costs for operation and model usage. According to our cost guide, process agents with system integration usually cost between 10,000 and 40,000 euros to set up.

Can I build an AI agent for free?

Many open source platforms and some agent builders can be tried without licence fees. In operation you pay for model usage, hosting and above all the working time for building, testing and maintenance. That is often enough for a personal helper, rarely for an agent that handles customer data.

Do I need programming skills to build AI agents?

Not for simple helpers: agent builders and workflow platforms work without code. As soon as an agent is meant to write safely to your ERP, CRM or accounting system, catch errors and log every step, you need technical know-how, in your own team or with a partner.

Does an AI agent have to be labelled under the EU AI Act?

If the agent interacts directly with people, for example in customer service, Article 50(1) of the AI Act requires that they are informed they are interacting with an AI system, unless this is obvious. The obligation applies since 2 August 2026. For agents in contact with customers we build the notice in from the start.

What is the difference between an AI agent and a chatbot?

A chatbot answers questions from a fixed knowledge base. An AI agent pursues a goal across several steps, uses your systems to do so and puts decisions with consequences to a person for approval.

Guides

AI agents and automation: our guides

Costs, data protection, how agents differ from RPA and how pilots reach production, written from our project work.

Which workflow should your first AI agent take over?

Tell us in a 20-minute intro call. Afterwards you know whether you can build the agent in-house or whether we should build it for you, and what a pilot would look like.

Book an intro call (20 min) Start a project

AI agents for business: how an agent is built that works inside your operations.