AI and data
AI process automation where rules are not enough
Standard automation handles “when X arrives, do Y”. Once the input is free text, a scanned invoice or a customer email, rules break down. That is where I use AI: it reads, sorts, extracts the data and prepares the output for the next system. A person confirms sensitive steps.
Who it is for: For companies drowning daily in emails, invoices, orders or forms in inconsistent formats that someone has to read, sort and retype by hand.
- From
- CZK 25,000
- Typical timeline
- 4–6 weeks
- You get
- A running process, monitoring, documentation
- Oversight
- A person approves sensitive steps
01
This is probably what hurts.
- Someone retypes data from PDFs or emails into a system every day.
- A shared inbox fills with messages that must be read just to know who owns them.
- The same document types arrive in a different format every time, so templates fail.
- Processing stalls because “someone has to take a look”.
- Retyping errors surface only in invoicing or in the warehouse.
If you nodded at least twice, it is worth a conversation.
02
What we actually solve.
Typical cases clients bring. Yours may look different; we always start from what you need.
Email triage and routing
AI reads an incoming message, identifies topic and urgency and assigns it to the right person or queue. Complaints, enquiries and order-status questions no longer land in one pile.
Invoice and order extraction
From a scan or PDF it pulls supplier, amounts, items and dates, checks them against rules and sends them to your accounting or warehouse system. Suspicious cases are held for review.
Draft replies
For typical questions it prepares a draft from your material. A person reads it, edits it and sends it with one click.
Summaries and minutes
From a call or a long thread it produces a summary with tasks and deadlines and writes them where your team tracks work.
Data-quality checks
It finds duplicates, missing fields and inconsistencies in your CRM or catalogue and proposes fixes for approval.
03
How it works, step by step.
For each step you see what I do, what I need from you and what you hold afterwards. No black box.
- 01
Process mapping
Week 1I do
I walk the process step by step, count how much time it takes and decide which steps suit AI and which do not.
You do
Show the process live and send samples of real inputs, including the messy ones.
You get
A process map, a list of AI steps and an estimate of the benefit.
- 02
Design and rules
Week 1–2I do
I design the data flow, the output format and the rules for when AI continues alone and when it waits for approval.
You do
Confirm the boundaries and name who approves.
You get
An approved flow design including exceptions.
- 03
Build
Week 2–4I do
I build the processing, connect inputs and target systems and add validation and error handling.
You do
Give access to the systems and test data.
You get
A working process in test operation.
- 04
Verification on samples
Week 4–5I do
I run the process on your real samples, measure accuracy and tune. I log everything that behaved unexpectedly.
You do
Review the results and say which errors are unacceptable.
You get
An accuracy overview and a list of refinements.
- 05
Go live
Week 5–6I do
I launch the process in production, set up monitoring and failure alerts and hand over documentation.
You do
Decide who watches the approval queues.
You get
A live process and an operating guide.
04
What it can look like in practice.
Model situation: orders by email at a supplier
A supplier receives orders by email, sometimes in the message body, sometimes as an attachment. Two colleagues retype them into the warehouse system every day.
Before
- Retyping takes up part of the working day.
- Typos only show up at dispatch.
- Seasonal peaks cause delays.
Process
- AI reads the email and attachment and extracts items and quantities.
- It checks them against the catalogue and the customer.
- Clear orders are created in the system.
- Unclear ones go to a colleague with a note of what is missing.
After
- Routine orders create themselves.
- Colleagues deal only with exceptions.
- Every step is logged and traceable.
An illustrative scenario, not a specific client and not a promised result. We establish accuracy and savings on samples of your documents in the Verification phase.
05
What I build with, and why.
- Language models and document extraction
- Chosen by input type: text, scan or table. What matters is accuracy on your documents.
- Make / n8n and custom scripts
- Orchestration, queues and approvals. I write code only where a connector is not enough.
- APIs of your systems
- Accounting, ERP, CRM, helpdesk, e-shop or Google Workspace.
- Validation and approval queue
- AI proposes, rules check, a person confirms sensitive steps.
- Monitoring
- Failure alerts and an overview of how much was handled automatically and how much went to people.
What you get at the end.
- Process map and automation design
- AI-based input processing
- Integration with target systems
- Validation, approval queue and error handling
- Monitoring and alerts
- Documentation and handover of operation
06
What it costs and how long it takes.
from CZK 25,000
Typical timeline: 4–6 weeks
AI usage and tool licences are paid directly to the vendors. You get an estimate of running costs before we start.
Full price listWhat moves the price
- Variety and quality of the input documents
- Number of connected systems
- Required autonomy versus approval
- Processing volume and data-protection requirements
When it does not make sense.
I would rather say so now than have you pay needlessly.
- The process is simple and inputs are uniform. Classic automation without AI is cheaper and more predictable.
- Nobody can say what a correct result looks like. Without criteria, quality cannot be verified.
- You want AI to decide on money or contracts on its own with no human review.
07
What people ask most.
Does AI have to decide everything?
No. Routine steps run automatically, and as soon as AI is unsure or the decision is sensitive, the item goes to an approval queue. We set the boundary together and it can be moved over time.
How accurate is the extraction?
It depends on the document type, so I measure accuracy on samples of your real documents before go-live rather than quoting generic figures. A document the system is unsure about goes for review.
How does it differ from classic automation?
Classic automation follows fixed rules over structured data. AI automation can read free text, scans and inconsistent formats. I often combine both: AI processes the input and rules verify the result.
Where is my data sent?
Before we start we define which services process the data, where it is stored and the vendor’s contractual terms. For sensitive data, options with EU processing or minimised data transfer are available.
What happens when the process fails?
The error is logged, the item stops in a queue and you get an alert. Nothing is deleted or forwarded half-done. Failure behaviour is part of the design.
Can we start with one process?
I recommend it. One bounded process gives a fast, measurable result and shows whether it is worth continuing.
Can you connect to our older system?
I check whether it has an API, an export or another interface. If not, we look for a workaround (such as file import) and I will tell you honestly where the limits are.
08
Let’s talk it through.
Write a few lines about what you are dealing with. I will reply within 24 hours with a view on whether and how it makes sense, before you commit to anything.