You can bring AI into a business process in six steps: pick one process, describe how it runs today, split the work between rules, AI and people, launch on a small volume, measure and fine-tune, and keep the solution supported. AI automation for your task starts at €250 on average, and development takes up to a month.
The roadmap: six steps
| Step | What we do | Result |
|---|---|---|
| 1. Pick a process | Find the spot where a person currently decides by hand | A clear first target for automation |
| 2. Describe the process | Where data comes from, who does what, what counts as an error | A list of checks |
| 3. Split the work | What rules handle, what AI handles, what a person confirms | A trust threshold |
| 4. Launch small | The most important scenarios, on your own examples | A first working version |
| 5. Measure and fine-tune | What went through automatically, what went to review and why | A solution that fits how the team really works |
| 6. Keep it supported | Hosting, backups, support, warranty | The solution keeps working after launch |
Step 1. Pick one process
Start with one process that eats up time and where someone makes decisions by hand every day. These are the processes we work with:
- Documents and email: invoices, statements and receipts that are now copied into the books by hand;
- Requests: messages and enquiries that need to be sorted and classified;
- Reports: figures that are now pulled together in spreadsheets by hand;
- Publications: material from websites that has to be collected, cleaned up and posted to a channel, with translation or rewriting if needed.
Starting with one process gives you a clear before-and-after comparison.
Step 2. Describe how the process runs today
Write down where the data comes from (email, website, Telegram, an existing system), who does what with it and where the result ends up. Above all, agree on what counts as an error. For documents, for example:
- the line items don't add up to the total;
- the wrong recipient;
- a duplicate of a document already received;
- someone else's company or vehicle on the invoice.
In the Fleet Control case study, invoices and fuel control used to be handled manually, on paper and by email, and mistakes and unnecessary costs were spotted too late. Now cases like refuelling with another vehicle's card or a wrong odometer reading are flagged by the system as problems.
Step 3. Decide what AI does and what rules and people do
AI isn't needed everywhere. Documents with a text layer are handled by ordinary rules; the model comes in where rules are not enough.
| Who | What they do |
|---|---|
| Rules | Read documents with a text layer, check amounts, recipient and duplicates |
| AI model | Identifies the document type and recipient where rules fall short; reads non-standard invoices; classifies requests; answers questions about your data in plain language |
| A person | Confirms doubtful cases and amounts above the threshold |
The model's answer goes through the same checks as all other data, and a decision log shows what it did and why. At this step we agree on the trust threshold: what gets posted automatically and what waits for confirmation. More on our AI in processes page.
Step 4. Launch on a small volume
The first version covers the most important scenarios, built on your own examples, rather than everything at once. We recommend starting with one source and part of your documents or requests, then expanding once the checks work as they should. Your team starts working in the system, and we stay close by.
Step 5. Measure and fine-tune
After launch, judge by what you can see in the system itself:
- how many documents or requests went through without manual input;
- how many went to a person for review and why: the problem list explains the reason;
- which mistakes and unnecessary costs were found.
In Fleet Control's first week, about 30 PDF invoices were turned into tables and charts, 30 fuel anomalies were detected automatically, and an unnecessary €500 invoice turned up that another contractor should have paid. From there, the system is adjusted to how the team actually works.
Step 6. Support after launch
The solution can run on our server. A new source or a new metric is set up separately.
| What | Price | Note |
|---|---|---|
| Hosting on our server | €8.50 per month | Regular backups included |
| Priority support | €50 per month | Paid separately |
| Warranty | 12 months | On solutions we build |
Who is responsible for what
We recommend defining three roles from the start:
- Process owner decides what counts as an error and where the confirmation threshold lies, and signs off on the result.
- Operator or administrator works with the system every day: goes through the problem list, confirms doubtful cases and reports what isn't working right.
- Developer (us) sets up the rules, the model and the checks, launches the solution and supports it.
What it costs and how long it takes
AI automation starts at €250 on average, the final price depends on the task, and custom development takes up to a month. To quote a price, we need to know what you want to automate, who will use the solution, what it should integrate with, your industry, your timeline and your budget. See How much does AI automation cost for details, and all prices on our pricing page.
About your data. Data is sent to the model provider via API only to process a specific request. What exactly is sent is agreed before launch.
FAQ
Where should I start with AI in my business?
With one process that takes up time and where a person decides by hand: documents, requests, reports or publications. Then describe how it runs today and what counts as an error.
How much does it cost to bring AI into a business process?
AI automation starts at €250 on average, and the final price depends on the task. After launch, hosting on our server costs €8.50 per month.
How long does AI implementation take?
Custom development takes up to a month. We give the exact timeline once we have clarified where in the process a model is needed.
What if the AI makes a mistake?
AI output goes through the same checks as all other data: mismatches go to a person for review. Amounts above the threshold are not posted without confirmation.
Do I still need a person once AI is in place?
Yes. A person confirms doubtful cases and amounts above the threshold, and a log of the model's decisions shows what it did and why.
Does my data end up in public models?
Data is sent to the model provider via API only to process a specific request. What exactly is sent is agreed before launch.