Why your chatbot costs you money and what takes its place
Many companies introduced a chat window and noticed after a few weeks that nobody uses it any more. This article explains why a chatbot creates work instead of removing it, what an agent does differently and which limits it needs.
Fabian Weiss Key Takeaways
- A chatbot delivers text that a human has to read, check and transfer into another program. The work stays, it only moves.
- An agent does not return text but a function call that your system executes. The human as translator between AI and software disappears.
- An agent needs fixed limits in code: amount caps, separate read and write permissions and a handover to a human as soon as it is unsure.
Why your chatbot costs more than it delivers
The wave of text generators since 2023 created a market built on a false assumption: that producing text is the same as getting work done. Many companies built a chat window into customer service, HR or sales. After a few weeks nobody uses it any more. We see this regularly in first conversations and usually it is the reason the conversation is happening at all.
The reason is simple. A chatbot is a consumption good. It is only useful while a human is operating it. An agent is a capital good. It works without anyone sitting next to it. The difference is the same as between a service that ties up staff with every use and a machine that produces on its own once bought.
A chatbot binds attention instead of freeing it. Someone types the question. Someone reads the answer. Someone checks it, copies it into the actual program and clicks save there. The AI is an additional step in the workflow, not a replacement for one. Why the pure computing costs hardly matter any more is covered in the article on cognitive deflation.
There is a second effect many decision-makers miss. A chatbot creates the feeling that the AI topic is handled. That feeling is exactly what delays real automation by months. Why service providers often have an interest in keeping that feeling alive is covered in the article on the agency paradox.
What has changed since 2022
2022. Language models generate text. Companies build chat windows. Every request costs computing but completes no transaction. The human reads the answer and carries out the actual action by hand.
2026. Agents carry out actions. The model does not answer in prose but with a structured command, such as “send invoice 123” or “set customer to status closed”. Your system executes the command. The human no longer sits between AI and software.
The technical lever behind this is called function calling. The model is given a list of permitted functions and answers with a call from that list, in a format a program can read directly. What the model is not allowed to call, it cannot call. The program code that executes the calls is ordinary software and behaves the same way every time.
The point in one sentence
A chatbot gives you text you still have to turn into work yourself, an agent gives you the completed transaction.
What this looks like in practice
A typical case from our projects is procurement in a mid-sized manufacturing company. The manual workflow: check stock levels, compare suppliers, maintain an Excel list, write the order email. Four steps, three programs, one human as the glue.
The automated workflow consists of four small agents with one task each. An observer queries stock levels at fixed intervals. A researcher fetches current prices from the stored suppliers. A checker compares the best offer against the approved budget. An executor places the order and writes the order number into the inventory system.
The human steps in only if the budget would be exceeded or a supplier does not respond. In projects we see processing time per order drop from close to an hour to under a minute. How many orders go through without intervention depends heavily on how clean the supplier data is. With well-maintained master data it is most of them. With chaotic data the workflow stops constantly. That is exactly what it should do.
The saving is not only time. The head of purchasing, who previously spent a large part of his week on routine orders, now negotiates framework agreements. The agent handles the volume, the human handles the decisions that need experience. How such a workflow fits with data protection is covered in a separate article.
How we stop the agent from ordering nonsense
The most common objection: what if the agent orders ten thousand screws? This is where clean architecture separates from tinkering. We limit what the agent can do at all, in code, not in the instructions to the model.
An amount cap is fixed in the program. The agent cannot place an order above it, no matter how confident it is. Read and write permissions are separated: the researcher may only read and cannot break anything. Only the executor may write and it is doubly secured. As soon as the model rates its own answer as uncertain, the workflow stops and reports to a human, for example via Slack or email.
That is not a decision by the AI but a fixed rule in code. The agent does not know it is constrained. It simply is. More on why such workflows are easier to audit than human routine is in the article on the security of automated workflows.
| Step | Chatbot | Agent |
|---|---|---|
| Trigger | A human types a question | An event in the system, such as low stock |
| AI response | Text | Function call |
| Execution | A human copies and clicks | The system executes the call |
| Limits | Wording in the instructions | Fixed in code |
| Growth | More staff | More computing |
What this means for your growth
Software made workflows faster. Chatbots simulated automation. Only agents that work without a human in between decouple growth from headcount planning. When your company grows, you do not hire more people for every transaction. You give the workflow more computing. Why that changes the most important metric in the company is covered in the article on revenue per employee.
Most mid-sized companies we advise pay for a long list of subscription tools. Each solves a partial problem. None of them talks to the others on its own. Humans make the connection in between, with copy and paste. That translation work is the real cost block, not the licences. A few agents with clear tasks replace it. Not because agents are cheaper than subscriptions, but because they make the humans between the tools unnecessary.
What you can do next
- Find the workflow in your company with the highest volume and the lowest complexity. Procurement, incoming invoices, scheduling or pre-sorting applications are typical candidates.
- Measure how long one transaction takes today and how many programs are involved. Every program a human copies between is a place for an agent.
- Start with one workflow, not five. The first one teaches you which limits and handovers you need. Every subsequent one becomes cheaper because of it.
Anyone still buying a chat window as an AI strategy today is building tomorrow’s legacy system. If you want to know which workflow qualifies first in your case, write to us.
Research for technical decisions
New reports, benchmarks and technical analyses on SaaS economics, AI engineering and owned infrastructure.
Original research Public sources No sales mail
By subscribing you receive new analyses and updates from FW Delta by email. You can withdraw your consent at any time. Further information is available in the privacy policy.