This piece helps with the choice and gives ten criteria, each with a question for the provider and the sign of a good answer.
At a glance
- Check who approves, what the agent may do and what gets logged.
- Work through the ten criteria below before you sign an offer. Each can be settled with one question and one demonstration.
- Start with a narrowly defined routine process, such as checking incoming invoices, and let the agent only propose at first.
Why control decides
Gartner expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027. The forecast names rising costs, unclear value and inadequate risk control as the reasons. The analyst also warns of “agent washing”, meaning products that relabel a chatbot or a rule-based automation and sell it as an agent (source 1).
Trade articles, for example at t3n, often see the cause in the architecture around the model and less often in the model itself (source 2). An agent stands out in daily work when it does something wrong and nobody can trace why. So look at approvals, permissions and the log before you settle on a model or an interface.
Chatbot, automation or agent
The three terms are often mixed up. The table separates them by what counts in daily work.
| Feature | Chatbot | Automation | Agent |
|---|---|---|---|
| Task | Answers a question | Runs fixed steps by rules | Works towards a goal and carries out several steps |
| Flow | A conversation, answer by answer | The same path every time | The path branches depending on intermediate results, within fixed limits |
| Tools | Usually none | Hard-wired | From a list of permitted tools |
| Start | An input | Schedule or incoming item | Schedule, incoming item or event |
| A person decides | Always, they read the answer | At fixed points or not at all | At approval points |
| Example | A question about the leave policy | Move an invoice attachment from a mail into a folder | Read an invoice, match it with the order, prepare a posting |
Not every process needs an agent. If it runs the same way every time, an automation is enough and is easier to check. An agent pays off when the next steps depend on the content, for example because every document looks different.
Ten criteria for choosing
Put each question in writing and ask to see the answer on the product. A slide is not enough.
| No. | Criterion | Question for the provider | How to recognise a good answer |
|---|---|---|---|
| 1 | Fit with ERP and documents | Which systems does the agent work with directly, and what needs extra programming? | The provider names connections that run today and keeps them apart from plans. |
| 2 | Templates | Which ready-made flows exist for my department, and how do I change them? | You see templates with trigger, steps and approval point, and you can adapt them. |
| 3 | Triggers | What starts a run, and what prevents a double start? | Schedule, incoming item and event are possible, and a double start is caught. |
| 4 | Approvals | Where does the agent put something in front of a person, and who covers during absence? | Approval can be set per step, and the default is cautious. |
| 5 | Permissions | With which permissions does the agent work? | It never may do more than the person or role it runs for. |
| 6 | Log | Can I see for every run which steps ran and what they relied on? | Every step can be read back with source and approval, and exported. |
| 7 | Testing | How do I test an agent before it changes anything? | There is a dry run without effect and an operating mode in which the agent only proposes. |
| 8 | Model choice and operation | Which models run where, and can I switch? | Model and location can be chosen per task, and the contract says whether inputs are used for training. |
| 9 | Cost | What drives the cost, and how do I cap it? | There are limits per agent or tenant, and the pricing logic is available in writing. |
| 10 | Exit | How do I get my data and flows out if I switch? | Interfaces and export routes are documented. |
How we meet the ten criteria
The table sums up what Beyonetix OS offers on each point. Which function runs in which module is described on the linked pages.
Our templates mostly run fixed processes with approval points. In the comparison of chatbot, automation and agent earlier in this article, such a process counts as an automation. In ENGIVERITY you set the goal, the permitted tools, the step limit and the budget for each agent.
| Criterion | What Beyonetix OS offers | More |
|---|---|---|
| Fit | Templates run in Documents and in ENGIVERITY. Ad Pilot has agent roles on request, and the ERP and CRM run the assistant and fixed processes such as invoice approval. | Agent platform |
| Templates | Ready-made flows per department, in Documents switched on per tenant. Your own versions can be saved in ENGIVERITY. | Templates for agents |
| Triggers | Schedule, event, upload and manual start in ENGIVERITY, arrival of a document and deadline in Documents. A double start is caught. | Building agents |
| Approvals | An approval point per step, chains with deputies and the four-eyes principle in Documents, and in Ad Pilot levels of autonomy. Ad Pilot does not write back into advertising accounts today. | Approval and control |
| Permissions | In ENGIVERITY, every agent gets limits, such as projects, object types and read-only or write access. In Ad Pilot, upper limits apply per change, plus a quota per tenant. The chat in Documents never has more rights than the interface. | Approval and control |
| Log | In ENGIVERITY there is a step log, a trigger log and a log of every tool call. | Approval and control |
| Testing | In ENGIVERITY there are a dry run, shadow mode and a test week with sign-off. | Dry run and shadow mode |
| Model choice and operation | In the ERP, a central, replaceable model layer that allows transfer to external providers only with permission per company. | Sovereign infrastructure |
| Cost | Step limit and budget per agent in ENGIVERITY, page quota per tenant in Documents, quota per tenant in Ad Pilot. | Building agents |
| Exit | REST interfaces in ENGIVERITY and Documents, signed webhooks in Documents. | Ask about export routes |
What the law says
First settle which role you hold under the EU AI Act, provider or deployer. Since 2 August 2026, first transparency and labelling duties apply, including to deployers, for example for AI-generated texts on matters of public interest (source 3). Whether a use counts as high risk depends on the application. Staff selection and performance evaluation are among them. The deadlines are sorted out in the article The EU AI Act for mid-sized companies.
For data protection, legal basis, data processing agreements and deletion apply as for any other software. The five points to check are in the article AI and the GDPR in business. This does not replace legal advice. The legal position is that of November 2026, so check it again before you decide.
How to start
Pick a process that recurs often and can be narrowly defined. Typical starting points are the handling of incoming invoices and the sorting of support requests. Fix a figure in advance by which you will measure success, such as minutes per case. Let the agent only propose at first and approve every step by hand. When the first process works with evidence, the next one follows.
What this looks like with Beyonetix OS is shown on the page about the agent platform. Questions and commands to your company data are covered on the page about the AI assistant. For an overview of use cases and controls, read the article AI agents in the company.
Sources
- Gartner, press release of 25 June 2025, forecast on cancelled agentic AI projects by the end of 2027: gartner.com
- t3n, article on AI agents failing because of architecture mistakes: t3n.de
- IHK Nürnberg, notice on transparency and labelling duties from 2 August: ihk-nuernberg.de