AI document processing for SMEs, with a person still in control
ZT Flow AI is workflow automation with AI steps added where rules alone are not enough, such as reading invoices and contracts, sorting incoming requests, and summarising long messages. It is for companies that receive a lot of documents or free text and have people reading all of it.
Who it's for
- Finance and admin teams that read and key in supplier invoices or orders
- Companies that receive requests by email or forms and sort them by hand
- Teams that need a short summary of long threads, calls or documents
What we need from you
- A sample of real documents or messages, with names and personal data redacted if you prefer
- A description of what a correct result looks like
- A person who can review uncertain cases during the first weeks
What it does
An AI step reads something a rule cannot, then hands a clean result to the rest of the workflow.
- Extract data from documents. Supplier, date, amounts, line items and references are pulled from PDFs and scans into your accounting or ERP system.
- Classify requests. Incoming emails or form submissions are labelled, for example as a quote request, a complaint or an invoice question, and sent to the right person.
- Summarise. Long threads, call transcripts or documents become a few lines, sent to whoever needs to act.
- Check and flag. The AI compares a document with an order or contract and flags differences for review.
AI where it helps, rules everywhere else
Language models are strong at reading and weak at being exact every time. We keep calculations, matching and approvals as plain rules and use AI for reading and writing only. Each AI step has a defined output, a check on that output, and a fallback to a person.
What a day looks like after launch
Documents arrive, the workflow extracts and checks them, and most pass through. The ones the AI is unsure about appear in a short review list with the source document next to the extracted values. Your team confirms or corrects them in seconds, and corrections are logged so we can improve the setup.
How the work is organised
We start by testing on your real examples, because that tells us whether the project makes sense before you commit. Then we build, run in review mode, and move to automatic processing once the accuracy is proven. [PLACEHOLDER: typical project duration]
What you own
The workflow, prompts and documentation are yours. We build on accounts you control where possible.
Four steps from first call to launch.
- 01
Collect real examples
We ask for a sample of your documents or messages, including the messy ones, to see what the AI has to handle.
- 02
Decide where AI helps
We use AI only for the steps that need it. Everything that can be a clear rule stays a rule.
- 03
Test accuracy on your data
We measure how often the result is right on your examples, and set when a person must check it.
- 04
Launch with review in place
Uncertain results go to a person first. As accuracy is proven, the review can be reduced.
How your data is handled
Documents are processed only for the automation you ordered. We tell you which AI provider processes the data and under what terms, and we use data-processing agreements and EU processing options where the provider offers them. Personal data can be redacted before it leaves your systems. Every AI result is logged next to its source document.
No supplier invoice lost: automatic routing for a B2B distributor
Supplier invoices are pulled automatically from the national e-invoicing system and routed to the right department folder, instead of five people searching one inbox.
Common questions
How accurate is AI document extraction?
It depends on the documents. We measure accuracy on a sample of your own files before launch and agree a threshold. Results below it are sent to a person instead of being processed automatically.
Will my documents be used to train an AI model?
We choose providers and settings that do not train on your data, and we put this in writing in the proposal. [CONFIRM: providers and settings to name publicly]
Do we need AI at all?
Not always. If your documents are standard, such as structured e-invoices, rules are enough and we will say so. AI is added only when it removes real manual work.
What types of document can it read?
Invoices, orders, delivery notes, contracts, forms and emails are common ones. We confirm on your samples before proposing a price.
Which task would you eliminate first?
Tell us in 30 minutes how your team works. We will show you what can be automated and how much time you would get back.
Book a free call