Pickware AI Chat works on request: you ask a question and get an answer. An agent (AI agent), on the other hand, works independently in the background. It starts on a schedule, when an event occurs in your shop, or when you start it, checks your data, and puts its result into the inbox as a message. This way, an agent takes over recurring checks and reports, such as a daily warehouse briefing, a revenue report in the morning, or a check of master data.
You create an agent via chat: you describe in your own words what it should take over, or you start with a template such as the AI warehouse manager.
Where do I find the agents?
You find the agents in the Admin menu under Agents. There are two areas: your agents' messages land in the Inbox, and under Agents you see all created agents with task, trigger and last run.
You create a new agent there via Create agent. Alternatively, open Pickware AI Chat and select New agent.
Which plans include agents?
Until 31 December 2026, you can use agents in all Pickware plans and also during the trial period.
After that, agents are included from the Advanced plan: they analyze your data and report their results in the inbox. The full range of agent features is available from the Professional plan.
How do I create an agent?
After you click Create agent, the chat opens with the question What should the agent take care of for you? Describe the task the way you would explain it to a colleague, for example:
Every weekday at 6:30 a.m., check which items are oversold, meaning more is ordered than we can deliver. For each item, show me the available and the physical stock, the open inbound quantity with delivery date, and how many open orders are affected. As a short list, the most urgent at the top. Only notify me if there are oversold items.
If a key detail is missing, the chat asks, for example about the time. It then shows you the draft as an Agent definition card with task, trigger and settings.
With Create agent you save exactly this state. If you want something different, click Customize task or simply write it in the chat. The card then shows the new state.
Instead of your own description, you can choose a ready-made task under Or start with a template, such as the AI warehouse manager. The chat then only asks for what it needs to know about your shop. You can find which templates are available and examples of custom agents in the article Which agents can I set up with Pickware AI?
Note: An agent works on its own and cannot ask questions during a run. So write everything it needs into the task: which data it should check, at what value something is noteworthy, and what should be in its message.
When does an agent start?
An agent starts in one or more of these ways. You can see which triggers an agent has on its detail page under Configuration.
Schedule: for example daily at 7 a.m. or every Monday. The time applies in the agent's timezone, which is also shown under Configuration. If it is not correct, name the desired timezone in the chat when creating or customizing the agent, for example Europe/Berlin.
Event in the shop: for example when a goods receipt is completed, a supplier order is created or its delivery date changes, an order has been shipped, or a product is no longer available. Each event starts its own run. Which events your shop offers depends on what is installed. Ask about it in the chat when creating the agent.
Manually: With Run on the detail page, you start a run immediately.
There is no trigger for "a value has changed" or "a threshold has been fallen below". You solve such cases with a schedule and a check in the task, for example: "Check every morning whether a product has a margin below 20%."
How long does a run take?
Depending on the task and the amount of data, a run can take 20 minutes or longer. So set the schedule so that the result is in the inbox before work starts. This is why the AI warehouse manager suggests 6:30 a.m. on weekdays.
During a run, the detail page shows Working. The agent does not report intermediate states; the result and transcript appear only at the end. If a run could not be completed, this is shown in the run history.
How do I manage my agents?
Clicking an agent under Agents opens its detail page. There you see its open messages, the last and next run, the run history, and under Configuration its task, its triggers and its timezone.
At the top of the detail page you find these actions:
Customize opens the chat asking what should change. You describe the change in your own words and save it on the card with Save change. It applies from the next run.
Run starts a run immediately. The result appears in the inbox.
Pause stops the agent until you start it again with Resume. A paused agent does not run.
It is best to start small: create the agent with a simple task, look at the first results, and then refine the task via Customize.
How do I work with the inbox?
All messages from your agents come together in the Inbox. The Open, Assigned to me and Archive filters sort them by processing status. The number next to Inbox shows how many messages are open. There are no notifications by email or push.
Above each message, it says what it is about:
I need your help: The agent needs a decision, an input or an action from you.
You should take a look at this: Something did not go as expected.
For your information: a result or report.
Not every run produces a message. An agent reports when it needs you, when it notices something important, or when its task calls for a report. If the task calls for a report, it sends it with every run, even if there is nothing to report. If you only want a message for anomalies, write that explicitly in the task.
Orders, products and other records are linked in the message. A click opens the record while the message stays pinned. This way you work through longer lists item by item.
Under Deep dive at the end of a message, you see which data the agent relies on (Context) and which steps it took to get there (Agent activity).
With Assign, you take over a message. After that, only the person it is assigned to can chat with the agent about it. Once a message is dealt with, you mark it as Done. It moves to the Archive, and chatting about it is no longer possible afterwards.
How do I implement what the agent suggests?
An agent does not change any data during its runs. It analyzes, assesses the situation and suggests what to do. You can implement this directly in the Admin or in the chat for the message.
Below each message, you chat with the agent about its content (Ask your agent about this …). There you ask follow-up questions, have individual points explained in more detail, or give an instruction, for example "Create the suggested supplier order." The chat works with your user permissions and within the scope of your plan. Write actions in the chat are included from the Advanced plan, see What is Pickware AI?
You do not change the agent itself in the chat for the message. If it should work differently in future, open its detail page and click Customize. However, you can also have facts about your business that apply to all chats and agents remembered from the chat for the message. More on this in the article How can Pickware AI remember knowledge about my business?
What should I keep in mind with agents?
Check results: An agent is based on an AI language model. Two runs of the same task can turn out differently, and numbers can be inaccurate. Check critical values in the Admin before you act on them.
No emails: An agent does not send emails to customers or suppliers. It can, however, prepare a text, for example for a follow-up with a customer.
Name thresholds concretely: "below 25% margin" works as a specification, "too low a margin" does not.
Task before knowledge: If an agent's task specifies something different from a note in Knowledge, it follows its task in case of doubt. If it should work differently permanently, adjust the task.
Permission: Only users whose user role has the permission for it can create and change agents.
Agent or Flow Builder?
The Flow Builder maps fixed, rule-based processes: when a defined condition occurs, a defined action is executed. The result is completely predictable. An agent is the better choice when a task requires judgment:
Room for discretion: Which delivery delay is critical today, and which can wait?
Pattern recognition: Anomalies across many records for which there is no rigid rule.
Unstructured information: Content such as order comments or addresses that first have to be read and categorized.
The two are not mutually exclusive. Clear-cut automations remain well placed in the Flow Builder. With the flow action Trigger Pickware AI, a flow can also start an agent that listens for this event, if an agent should handle the assessment.