AI agents for business: examples and costs explained

What is an AI agent?

An AI agent is an AI system that completes a task in several steps on its own. Germany’s Federal Office for Information Security (BSI, in German) describes AI agents as AI systems that are meant not only to give answers but to carry out actions in the digital world on their own, such as reading a request, checking the calendar and suggesting an appointment. Anthropic describes the difference from a chatbot as a self-directed loop: the agent plans, acts, checks the result and adjusts until the task is done or it needs to ask a person.

At its core, an agent has three parts, as OpenAI describes it in a guide: a language model that thinks and decides, tools it may act with (email, calendar, bookkeeping program) and instructions with clear limits. Three things are often confused:

Type	Who decides?	Example
Chatbot	You ask, it answers	Answer a question about pricing
Fixed workflow	People set the rules in advance	A confirmation email goes out after payment
AI agent	The model chooses the steps, within your limits	Read a request, check the calendar, suggest an appointment

This distinction matches the descriptions of the BSI (a chatbot answers, an agent carries out tasks essentially on its own in several steps, without anyone confirming each step) and of OpenAI (classic workflow automations follow fixed rules, guide for business leaders).

Not every task needs an agent. The Fraunhofer Institute writes in its study of February 2026 (in German) that automation can often be solved with classic, fixed methods, and then frequently more easily. Microsoft also advises in its documentation: if a task can be handled with a simple function, use that instead of an agent. Gartner puts it like this: agents when decisions are needed, automation for routine workflows, assistants for simple lookups. For us, that means we tell you honestly whether an agent helps or whether something simpler is enough.

Be careful when buying: Gartner speaks of “agent washing” when vendors rebrand existing products such as chatbots or automation software without real agent capabilities. According to Gartner’s estimate of June 2025, only about 130 of the thousands of AI agent vendors are real. So always ask what the agent actually does (see the checklist below).

How common are AI agents in business?

Not very. In a Bitkom survey (in German) of 14 September 2026 (603 companies with 20 or more employees, surveyed by phone), 11 percent of the companies that use AI or plan or discuss its use already use AI agents. 29 percent plan to, and 31 percent are discussing it. Bitkom president Ralf Wintergerst expects agents to spread about as fast as AI overall. According to the same survey, 57 percent of companies use AI at all.

We did not find an official statistic on AI agents. The Federal Statistical Office counted for 2025 that 26 percent of companies with 10 or more employees use AI (in German). The figures differ because the survey, the timing and the definition of AI differ. The official survey has not covered businesses with fewer than 10 employees since 2025.

What are examples of AI agents in business?

Typical examples are the email inbox, appointment reminders, customer questions, receipts and quotes. Tasks that work well occur often, can be described clearly and cause no big damage if something goes wrong. The overview shows what the agent takes over and what stays with you. This is our suggestion, not a rule:

Area	The agent	You
Inbox	sorts emails, writes drafts	approve replies
Scheduling	reminds via WhatsApp, fills the waitlist	check exceptions
Service	answers common questions	take the difficult cases
Bookkeeping	files receipts, prepares invoices	check and approve
Sales	finds suitable companies, writes first messages	lead the talks
Quotes	creates drafts, follows up	approve prices

Here are some examples with sources:

• Email inbox: Anyone who uses the internet at work receives an average of 53 business emails per day, shows a Bitkom survey (in German) of January 2026 (self-reported, 532 employed people). An agent sorts inquiries, prepares replies and suggests appointments. Nothing is sent until you approve it.
• Appointments and reminders: The agent reminds customers via WhatsApp, takes cancellations and offers freed-up slots to the waitlist. A review of ten studies (March 2026) found that people are on average about 11 percent more likely to show up for an appointment after a reminder. That applies to clinic patients; for text messages alone, the effect was not statistically established. How this works with WhatsApp and what it costs is covered in our article on WhatsApp appointment reminders.
• Customer service: The agent answers common questions on your website, on WhatsApp and by email and hands difficult cases to your team with all the details. In the Bitkom survey, AI users name handling inquiries in customer contact most often as a use case (72 percent). The payment provider Klarna reported in February 2024 that its AI assistant had taken over two thirds of service chats in its first month, equivalent to the work of 700 full-time staff (company claim). In May 2025, Klarna’s CEO said, according to CX Dive, that the focus on costs had led to lower quality, and that customers should always be able to reach a person. At the Berlin skilled-trades business Krone GT, more than 400 calls come in every day. According to handwerk magazin (in German, May 2026), an AI phone assistant handles the first conversation, answers common questions and passes on more complex cases. The report gives no results.
• Bookkeeping: An example of AI automation with human control: DATEV reported in June 2026 that more than 100,000 sets of books use its AI service for bank transactions (about 7,400 firms, over 10 million booking suggestions per month). Symbols show where a person should step in or check more closely. The AI service for invoices reaches over 9,000 firms and 12 million booking suggestions per month. DATEV (in German) does not state any time savings.
• Quotes and follow-ups: The roofing company Walther from Bad Dürkheim uses a “Vertriebsbutler” (sales butler), according to the German Confederation of Skilled Crafts (in German): it answers inquiries, sends quotes and sees whether they were read. It suggests follow-up emails and carries them out automatically after confirmation. The pattern: the agent suggests, the person confirms.
• Applications, with caution: The AI Act generally counts AI systems that screen or filter applications as high-risk systems (Annex III no. 4, obligations from 2 December 2027). In their guidance (in German), the German data protection authorities name as a violation of Article 22 of the GDPR an AI that evaluates applications and sends out interview invitations on its own. A person should always decide here.

Further examples are purchasing (monitoring stock), keeping track of contract deadlines, summarizing reports from several programs and answering internal questions about company knowledge. On our page about AI agents you can see examples for every department.

How much does an AI agent cost?

As a rough range, the Austrian Federal Economic Chamber (WKO, in German) names €200 to €2,000 per month on its page about AI agents, depending on complexity and depth of integration (as of 5 February 2026). That is a guideline without a stated source, because we did not find a neutral price statistic for AI agents (as of 3 October 2026). With us, it starts at €400 per month (see below). Costs are seen as a hurdle: according to Bitkom, hard-to-calculate costs keep 63 percent of companies without AI from using it. So here is what the costs are made of:

1. Setup and adjustment. The agent has to be connected to your programs, configured with your rules and tested. This is a big item: at Bitkom, AI users name as a major cost item the infrastructure such as cloud and computing power (51 percent), preparing data (50 percent) and integration into existing systems (41 percent). Subscriptions and licenses are a major cost item for only 21 percent. Anyone who only buys a license has bought AI but has not changed anything yet, says Bitkom president Wintergerst.
2. Platform. Many agents run on a platform that connects programs and controls workflows. n8n Cloud costs, for example, from €20 per month with annual billing (Starter, 2,500 executions per month), Zapier from $19.99 per month with annual billing (as of 3 October 2026). Monthly billing is more expensive in each case. If you run n8n yourself, you pay for servers and maintenance instead.
3. AI model. The model provider bills by “tokens”, small chunks of text. On 3 October 2026, one million tokens cost (input / output, in US dollars): at Anthropic Claude Haiku 4.5 $1 / $5 and Claude Sonnet 5.5 $2 / $10, at OpenAI GPT-6 Luna $0.10 / $0.50 and GPT-6.1 Sol $2 / $10, at Google Gemini 3.5 Flash-Lite $0.30 / $2.50. The price pages give no information on sales tax.
4. Messages and third-party services. If the agent sends WhatsApp messages, Meta bills per delivered message. In Germany, according to the price list effective 1 October 2026, an appointment reminder (category “Utility”) costs €0.0456 and a marketing message €0.1131. Messages from your customers to you are free. Sending 1,000 reminders per month therefore costs €45.60 (example calculation). Meta changes prices only at the start of a quarter; the next date is 1 January 2027.
5. Operation and maintenance. Programs, prices and rules change, and an agent needs supervision. Plan time or budget for checking and adjusting, especially in the first months, when you work feedback from day-to-day use into the setup.

Example calculation for the AI usage (assumed volumes, prices of 3 October 2026): An agent reads 1,000 customer inquiries per month and writes a draft reply to each. We assume 2,000 tokens of input per inquiry (email, instructions, knowledge) and 500 tokens of output. That is 2 million tokens of input and 0.5 million tokens of output. With a mid-range model at $2 / $10 per million tokens (for example Claude Sonnet 5.5 or GPT-6.1 Sol), that costs 4 + 5 = $9 per month; with Claude Haiku 4.5 ($1 / $5) it is $4.50. If the agent works in several steps, it needs more: in a technical article, Anthropic puts agents at typically about four times as many tokens as chats, and systems made of several agents at about 15 times as many. With a factor of four, it would be $36 per month.

The AI usage is therefore a small item in this example. That fits Bitkom’s figures: for 8 percent of AI users, token consumption is a major cost item, for 11 percent a minor one. Gartner, however, expects the providers’ computing costs per agent workflow to rise to more than five times as much by 2028. That is a forecast about the providers’ costs, not a price for you. So agree on a monthly limit for AI usage and messages and have the costs shown to you regularly.

Build it yourself, buy it or have it built?

There are three routes to an agent:

Route	One-time	Ongoing
Ready-made (e.g., in bookkeeping)	Turn on the feature, set the rules	Part of the subscription or an add-on
Build it yourself (e.g., n8n, Zapier)	Your time to learn and test	Platform, AI usage, maintenance
Have it built (freelancer, agency, us)	Fee for setup and integration	Support, platform, AI usage

Many programs already have AI features built in. In its spending forecast of September 2026, Gartner writes that companies use the simpler, embedded AI features of their existing software providers to automate workflows. So first check whether a program you already have can do your task. Building it yourself works if you have time to learn. Then the notes on security, data protection and liability below apply to you just the same. Having it built pays off if you want to hand off setup and maintenance.

Is an AI agent worth it?

Only your own calculation can show that: how often does the task occur, how long does it take today, and what does that time cost? The Federal Statistical Office gives average labor costs (in German) of €45.00 per hour worked for 2025 (manufacturing and services, estimate, including non-wage labor costs). Example calculation: if an agent costs €400 per month, it pays off at this average from about nine saved working hours per month (400 ÷ 45 ≈ 8.9). Use your own hourly costs. Keep in mind: saved minutes cannot always be turned into money, and checking the results takes time too.

What studies and surveys show:

• Customer service with an AI assistant: A study in the Quarterly Journal of Economics (2025) analyzed data from 5,172 customer support staff who received an AI assistant. They resolved 15 percent more cases per hour on average. That was an assistant that suggests answers, not an agent.
• Broad effects on working time and earnings (chatbots, not agents): A Danish study (NBER, March 2026) with data on about 25,000 workers found no effect of more than 2 percent on earnings and recorded working hours two years after the launch of ChatGPT. Workers who use chatbots reported saving about 3 percent of their working time on average.
• Self-reports from companies: According to Bitkom, 66 percent of companies that use AI say their competitive position has improved, and 48 percent were able to speed up internal processes significantly.
• Forecast: Gartner expects more than 40 percent of agent projects to be canceled by the end of 2027, because of rising costs, unclear business value or inadequate risk controls.

So the benefit depends heavily on what you use AI for and how. Before you start, measure what the task costs today (time, errors, waiting time), and compare after a few weeks.

What are the limits and risks?

Agents make mistakes, can be fooled by outside text and need clear limits. The WKO writes: define exactly which actions may run automatically and when a person has to approve.

• Mistakes and invented answers: According to a joint guidance from six security agencies (including CISA, NSA and NCSC-UK, May 2026), language models quite often make mistakes. OpenAI writes that made-up answers, known as hallucinations, remain a fundamental challenge for all large language models.
• Outside text can fool agents: An agent reads emails, web pages and documents. If a hidden instruction is in there, the model may take it for an order (technical term: prompt injection). OWASP, a nonprofit for software security, writes in its 2026 list that there is no reliable protection against this today. The UK’s National Cyber Security Centre considers it quite possible that the problem can never be fully fixed, and OpenAI expects that it will probably never be fully “solved”. In a test by OpenAI, an agent that received a prepared email wrote a resignation letter on the user’s behalf instead of writing an out-of-office note.
• More access, more damage: The BSI advises giving an agent only the access it needs for its task. No agent should be able to reach emails, bank account and personal files at the same time. The UK NCSC writes: the wider the access, the greater the impact if the agent malfunctions unexpectedly or is taken over. It also recommends that you can always pull the plug.
• A person approves anything with consequences: The six agencies recommend that agents do not carry out high-impact actions without prior human approval, and advise using agents only for low-risk, non-sensitive tasks. OWASP gives an example: a customer service agent may book a refund as store credit on its own, because it can be recovered, while an outgoing payout goes to a person. BSI and ANSSI expressly advise against letting AI systems work entirely without human oversight.
• Approvals have to stay real: Gartner distinguishes four levels: read only, make suggestions, act with approval, act on its own. At the “with approval” level, the check only works as long as it remains a real control. Under time pressure or fatigue it can slip and create a false sense of safety. That is why it pays to let only low-risk routine cases run on their own, step by step, and to keep putting everything important up for review.

How fast the technology is improving is shown by the Remote Labor Index of the Center for AI Safety. It measures how many of 240 real, paid freelance projects an agent completes so that the result would be accepted as commissioned work. In October 2025, the best agent completed 2.5 percent; on 3 October 2026, the leaderboard shows 20.8 percent. So even the best agent manages only about one in five of these demanding projects. For typical office tasks of small businesses, we did not find a comparable measurement. On its page about AI agents, the BSI itself writes that the technology is still at an early stage, and that security standards and legal frameworks are only just developing.

What do you need to know legally?

This is orientation, not legal advice. First of all: “AI agent” is not a legal term of its own. According to the European Commission, agents fall under the AI Act as AI systems.

• Your agent’s statements are attributed to you: On 12 May 2026, the Higher Regional Court of Hamm ruled (case no. 4 UKl 3/25, in German) that using an AI chatbot is a commercial practice of its operator, because the operator has sufficient influence on the system and sets it in motion. In that case, the chatbot of a company offering cosmetic treatments wrongly called its managing physicians specialists, and a consumer association sued to stop it. The court left open whether it is different when users provoke a wrong answer with a leading question. It is a ruling of a higher regional court, not of the Federal Court of Justice. In 2024, a Canadian tribunal ruled similarly in the case of Air Canada: it makes no difference whether information comes from a static page or a chatbot.
• A notice that people are writing with an AI: Since 2 August 2026, providers of AI systems that interact directly with people must make sure those people learn about it, unless it is obvious (Article 50 of the AI Act). In its guidelines (20 July 2026, not binding), the European Commission explicitly includes AI agents. A single clear notice before the first contact is usually enough, a sentence in the terms and conditions alone is not. Violations can cost up to 15 million euros or 3 percent of worldwide annual turnover; for small and mid-sized businesses, the lower amount applies. A provider is also anyone who has a system developed and puts it into service under their own name. Whether you count as one when you commission an agency is not expressly settled by the law. So plan for the notice in any case.
• Data protection: If the agent processes personal data, every processing step needs a legal basis, the German data protection authorities write in their guidance (in German, May 2024, not replaced since). If you use an external AI service, there is often a data processing relationship, for which you need a contract under Article 28 of the GDPR. Applications that do not use inputs for training are preferable. Decisions with legal effect should in principle be made by people (Article 22 of the GDPR).
• Staff and works council: AI systems used to select applicants are listed in Annex III of the AI Act as high-risk systems. Since the amendment by the Digital Omnibus, the obligations for them apply only from 2 December 2027. If there is a works council, you must inform it about the use of AI under section 90 of the Works Constitution Act (BetrVG). Technical systems meant to monitor the behavior or performance of employees are subject to co-determination under section 87(1) no. 6.
• AI literacy: Anyone who uses AI systems must support the AI literacy of their staff (Article 4 of the AI Act). What that means is explained in our article on the AI literacy requirement.

How do you recognize a fair quote?

By the breakdown: if you want to have an AI agent built, a quote can only be compared if it is clear what is in the price. Look out for these points:

1. Task and goal: Which task does the agent take over, and how is success measured, for example minutes per case or error rate?
2. Approvals and limits: What may the agent do on its own, and what only after your approval? Is there an emergency switch with which you can stop it immediately?
3. Access: The agent gets only the rights it needs for its task, and you know who has access.
4. Complete costs: Setup, ongoing support, platform, AI usage and message fees are listed individually in the quote. There is a monthly limit, and you can see who pays what.
5. Data and contracts: A data processing agreement is in place, you know where data is processed, and your inputs are not used for training.
6. Ownership and exit: Credentials, documentation and settings belong to you, and the notice period is clear.
7. Real technology, no labels: The vendor explains what the agent actually does and promises nothing that sounds like “agent washing”, such as “replaces staff” or “runs 100 percent on its own”.
8. Net or gross: The quote states whether VAT is added. Small businesses under § 19 of the German VAT Act (UStG) do not charge VAT.

How do you start with your first AI agent?

With a small, low-risk task and with approval. Fraunhofer recommends starting with small solutions for partial problems instead of trying a big leap right away, and OpenAI advises: start small, validate with real users and grow capabilities over time.

1. Pick one task. It occurs often, can be described clearly, and a mistake does little damage. Also check whether the task can be simplified or dropped altogether. The six agencies recommend this as well.
2. Measure the starting point. How much time does the task cost today, how often do mistakes happen, how long do customers wait?
3. Start with approval. The agent suggests, you approve. That corresponds to Gartner’s levels “make suggestions” and “act with approval”.
4. Evaluate. After a few weeks: was time saved? Were there mistakes or complaints? Only then do you let low-risk routine cases run on their own.
5. Involve your team. Write down who may do what, and train your team (AI literacy, see above).

What does an AI agent cost with us?

In the first three months, from €400 per month. That covers setup, maintenance and adjustment to your day-to-day work, so your feedback from daily use flows in right away. After that it gets much cheaper depending on scope, and after a few months the agent runs without any further charges from us. We are a small business under § 19 of the German VAT Act (UStG) and do not charge VAT. We name the exact price in the plan after the free intro call. Which third-party fees may apply to your agent, for example for WhatsApp messages, we clarify in the intro call.

Our agents work in the programs you already use, for example Outlook, Gmail, Microsoft 365, Google Workspace, HubSpot, DATEV, lexoffice, WhatsApp Business, Slack or Shopify. At the start you approve every step, and later you decide which routine cases an agent may handle on its own. How quickly an agent starts depends on how customized it needs to be: simple agents are up and running very quickly, heavily customized ones take a bit longer. More on our page about AI agents. In the free intro call, we tell you honestly whether an agent makes sense for your task.