An assistant for preparation and understanding
An assistant can find information, summarize a document or prepare a draft. Define its sources and how it shows the basis for an answer. For important decisions, users should be able to check the source and correct the result.
Automation for an agreed workflow
When the steps are known, focus on inputs, rules, access and reliable completion. For example: receive an enquiry, check required fields and prepare a task in an internal tool. Define how duplicates and errors are handled too.
Combine them when there is a reason
AI might suggest a document category while the system prepares an item for review. A person confirms the next step. This works best when decisions, data handling and work history are clear.
The same topic, two different deliverables
If a team asks what the returns policy says, an assistant can locate the source and prepare an answer for review. Recording an approved return requires a workflow that checks required data, prevents duplicates and records the result. Conversation and execution have different quality criteria. Before requesting a proposal, specify whether you need help understanding information, preparing a suggestion or completing a predefined task.
Test the solution on difficult examples
Alongside a clean example, include a message with missing information, a contradictory document and a repeated request. Observe whether the system asks for clarification, shows a source or indicates it cannot continue. Agree which actions always wait for human confirmation. This reveals more about usability than a single perfect demonstration. Record what works, what fails and who handles each exception.
A Midaz example: from enquiry to assigned task
On our own website, we connected the enquiry form to contact records and internal tasks. The original message stays with the contact, and Danijel receives an assigned task containing the enquiry details. The visitor receives an acknowledgement in the form's language. We put this workflow into production on 2 October 2026.
The data transfer and task preparation follow defined rules without AI judgement. The automatic acknowledgement confirms receipt. Danijel then reviews the goal and prepares the next step; the acknowledgement is neither an answer to the business problem nor an accepted proposal.
Before launch, we checked both language forms, the original content in the assigned task and repeated processing of the same enquiry. We also checked incoming email and WhatsApp replies. This demonstrates an implemented, tested workflow; it is not a measured time saving or a promise of more sales.
For a similar project, first define what must be retained, who owns the task and where a failed handoff becomes visible. An AI summary can be an additional draft for review, while the original message and responsibility for replying remain clear.
Frequently asked questions
What is the difference between an AI assistant and an automation?
An AI assistant helps people find information, summarize a document or prepare a draft, and a person checks the result. An automation carries out an agreed sequence of steps, for example receiving an enquiry, checking required fields and preparing a task. Conversation and execution have different quality criteria.
What do I need if I want my team to find answers in documents faster?
Usually an assistant. It can locate the source and prepare an answer for review, for example when the team asks what the returns policy says. For an important decision, a person should be able to check the source and correct the result.
When do I need an automation?
When the steps are known and repeat, for example receiving an enquiry, checking required fields and preparing a task in an internal tool. What matters then is the inputs, rules, access and reliable completion. You also need to define what happens with duplicates and errors. Such a workflow often does not need AI.
Can an assistant and an automation work together?
Yes. AI might suggest, for example, a document category while the system prepares an item for review, and a person confirms the next step. This makes sense when it is clear who decides, which data is processed and how the work history can be seen.
How do I test a solution before I start relying on it?
Alongside a clean example, try a message with missing information, a contradictory document and a repeated request. Observe whether the system asks for clarification, shows a source or says it cannot continue. Agree in advance which actions always wait for human confirmation and who handles each exception.
Describe the result in a simple sentence. If it explains what should happen and who confirms it, choosing the solution becomes easier.