What can be automated, task by task?
For each of nine common tasks the table says what rules can handle, what AI may only suggest and what stays with a person. Five have a more detailed page.
| Task | Rules can | AI may suggest | A person decides | More detail |
|---|---|---|---|---|
| Website enquiries | Save the original message, create a task, assign a person, confirm receipt | A short summary for review, next to the original | The reply, the assessment of the goal, the quote | Enquiry handling |
| Documents for accounting | Check attachments, duplicates and required fields, build a list of missing documents | The document type, a number read from it, a draft message to the client | What is accepted, corrected and booked | Guide for accounting offices |
| Customer questions | Give answers from an FAQ page, hand the conversation to a person | An answer from your own content, only when it exists there | Prices, deadlines, complaints, exceptions | Chatbot guide |
| Quotes and offers | Fill a template from items you approved yourself, add up totals | An opening text and a summary of the customer’s request | The price, the terms and sending it | No separate page |
| Reports and summaries | Collect the same figures from the same sources in the same way every time | A short text on what changed | What the figures mean and what to do about them | No separate page |
| Campaign planning | Gather offer, website and keyword data in one place | Keyword groups and ad drafts | Budget, goals, approval and launch | Campaign planning |
| Checking tracking and measurement | Compare what should be set up with what is actually set up | An explanation of the findings and a likely cause | What changes and when | Tracking checks |
| Reminders and follow-ups | Send a reminder on an agreed date, open a task if there is no reply | A draft message | Who is contacted and what exactly is said | No separate page |
| Data entry between tools | Copy fields from a form into a sheet or record, reject duplicates | Reading data from free text or a scan | Corrections to whatever the system flagged as unclear | No separate page |
A rule is enough when what you check is either true or false. Add AI only where a rule cannot cope, such as free text or a scan, and even then it does not decide.
What does Midaz know first-hand, and what is general method?
The enquiry flow on our own site has been running since 2 October 2026. It keeps the original message, creates a task for Danijel and sends a confirmation, on rules and without AI judgment. We checked that it delivers and behaves as agreed, and we measured no time saving. See enquiry handling.
For our own work we have also built internal systems: one for research and campaign planning, one that compares bank transactions with invoices (its AI advisor only suggests, and a person decides) and one for checking tracking. They show building experience, not client results.
Read everything else in this guide as general method.
How do you pick the first task?
We use the same test in our other guides. A good first task:
- repeats every week or more often
- has known inputs and a clear result
- has a person who can judge whether the result is good
- lets you spot a mistake before the work moves on
If there are several candidates, take the one where a wrong result is cheapest. A second task makes sense only once the first one works.
What should you not automate first?
Anything where a mistake is expensive or hard to notice:
- decisions with consequences: approving, refusing, accepting a complaint
- prices and promises you have not approved yourself
- legal, tax or medical advice
- booking, paying or sending to clients without a person reviewing it
- work done rarely, because there is nothing to learn from or check
- a process nobody can describe step by step
What does a first pilot look like?
A pilot is one bounded process, not the whole business, with a responsible person from your team involved from the start. It is a separate job with a written quote based on scope.
- Process mapping: today’s work, tools, how often the task occurs and where it gets stuck.
- Pilot agreement: inputs, tools, human review and the success criterion, before anything is built.
- A pilot of limited scope: a first example checked on representative tasks.
- Review with your team: the result, time, quality and the manual work that remains, then a decision on what next.
Measure the whole task, including review and corrections, not only the speed of the first result. A sample list of test cases is in the guide for accounting offices. What affects a quote is explained in the guide on the cost of AI automation.
What can go wrong?
We do not promise there will be no mistakes. These are the usual traps:
- a wrong answer delivered in a confident tone
- the process stops and nobody sees it
- the format of the inputs changes, for example a new form or a different kind of scan
- data goes to an outside service without you knowing
- the review becomes a formality, so mistakes slip through
- the review takes longer than the old manual work
That is why, before launch, we agree where a failed run is visible, who receives the task and how the work continues by hand.
Frequently asked questions
Does every automation need AI?
No. Where the rules are clear, ordinary automation without AI is enough. Add AI only where text has to be understood, information summarized or a draft prepared, and always with a review.
Which task should come first?
One that repeats, has known inputs and gives a result whose accuracy someone can judge. If several qualify, take the one where a wrong result is cheapest.
Can AI answer customers on its own?
It can, but we would not start there. Answers about prices, deadlines and terms have to pass through a person. The chatbot guide shows how to decide and how to test.
What if the AI gets it wrong?
We do not promise it will not happen. That is why steps that need a decision have a person reviewing them, and before launch we agree where a failed run is visible and who takes the task.
Do I have to change the tools I already use?
The aim is that you keep them. Before building we check what their integrations support and where the limits are, and only then agree whether a new tool is needed.
What happens to my data?
Your business data stays yours. Before building we agree which data is needed, who has access and which outside services the solution uses, and we explain that before anything is connected.
Have you done this for clients?
We have no AI client project to show. We have the enquiry flow on our own site and internal systems from our own work, and they show building experience, not client results.
How much does it cost?
There is no price list for AI automation. A pilot is quoted by scope, and the guide on the cost of AI automation explains what makes it smaller or larger.
Write down five tasks your team repeats every week. Next to each, note who does it, where the input comes from and who can judge the result. The first one that passes all three questions is your pilot candidate.
See also how much AI automation costs, where to start with AI automation, AI assistant or automation, AI and automation or contact.
How Midaz can help ↗