Skip to content
SSyntalith
Book a free process review (30 min)

Blog · Published · Updated

Transport documents ready for review: how the workflow works

Short answer

A legibility gate measures the brightness, contrast and sharpness of a photo of a transport document before it is read. When the image is unfit, the system records that no reading was attempted and asks for a new shot, in place of returning a full set of fields of unknown reliability. In our demonstrator, 5 of 7 image variants were stopped without calling the model. At Syntalith document reading automation costs from €3,000 net and pays back in 2.0 months.

How we workDetails ↓

  1. 01Call30 minutes with an engineer, free
  2. 02PrototypeA working system on your process and data, free
  3. 03ImplementationTarget and fixed price in the contract; 50% at signing
  4. 04Acceptance50% once the agreed target is met
analysis with an engineer and a free prototype
€0
team in ChatGPT or Claude, up to 10 people
from €1,150
automation, agent or application for one process
from €3,000
Book a free process review (30 min)

We belong to the OpenAI and Anthropic partner programmes and learn through their official courses. What this means ↓

Author: Syntalith · 3 min read

Example: the morning request inbox
Agent · ready

Statistics

Documents and AI: statistics 2026

5 of 7

image variants in the Syntalith demonstrator that the gate stopped before reading, without calling the model. This proves the mechanism works. We measure quality on the client's photos.

Figures checked: 22 October 2026

Is this photo readable in the first place?

A driver photographs a consignment note in the cab, in poor light and at an angle. The reading model returns a full set of fields anyway, because that is how it was built: some values are correct, some are guessed, and the employee cannot tell which are which. A legibility gate works earlier and stops an image that is not worth reading. The driver gets a request for a new photo while still at the unloading point. A week later, when the document reaches the office, a new shot is often impossible.

The gate before reading
  1. 01MeasureBrightness, contrast and edge features of the image.
  2. 02DecisionThe image goes on to reading or is stopped.
  3. 03RecordWhen stopped, the result says „no reading attempted”, with the reason.
  4. 04RequestThe driver or the sender gets a request for a new shot.
  5. 05ReadingA legible image goes to OCR, and the fields sit next to the parts of the image they came from.

What our demonstrator shows

The operator chooses one of seven images, rotates the preview and points to a field on the document. The two legible examples let you download the saved fields, and the other five keep the state of a refused reading. The archive shows no model calls for the five blocked variants. The photos and synthetic values come from an earlier demonstrator, and the limits are set out in the source note. This is no measure of production quality and no result from a client rollout.

Three states of an image

  • Legible

    to reading

    The fields go for checking next to the parts of the image.

  • Illegible

    stopped

    No reading attempted, with the reason: too dark, blurred, cut off.

  • Borderline

    to a person

    An image at the threshold goes to a manual reading.

What a legibility gate costs and when it pays back

Scenario: 1,000 photos of transport documents a month
Today
1,000 photos × 3 min = 50 h (€1,745)
With the system, a person checks the output in 15% of the time
€262
Model, 1 call per case, Claude Sonnet 5.5
€7.34
Cost of one case
€1.75 today, €0.27 with the system
Saving
€1,476 a month

Payback of €3,000

2.0 months

The volume, the time and the checking share are the scenario's assumptions. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat. Model: 2,500 input tokens and 320 output tokens per call at Anthropic's rates, converted at the ECB rate of 1.1205 USD per euro on 22 October 2026.

The gate also lowers the model cost, because stopped images are never read, but the main gain is the absence of falsely complete data. According to Anthropic (2025), AI cuts the time of document tasks by an estimated 87%, but only when the result does not need every field checked again.

A first trial in the business

Setting the threshold on your own photos
  • Accepted and rejected photos from one process
  • The number of false rejections the team can handle
  • The point at which a manual reading is needed
  • The time from the request for a new shot to receiving it
  • The fields that most often need correcting after reading

A gate that is too strict shifts the work onto drivers, and one that is too lenient lets guessing through. That is why in the first month we count both kinds of error and move the threshold together with the person who receives the documents. We set the threshold on the client's photos, and the next steps are described in the posts about reading CMR documents and fine-tuning for transport documents. Google describes evaluating Document AI processors on a test set, which helps compare the variant with the gate against the one without it.

Worked example · Scenario

How we count the payback: reading transport documents

We run the same count for your documents in the free analysis.

€1,745

a month for copying 1,000 documents by hand

Payback: 2 mo.

3 minutes to copy the fields from a document and checking in 15% of the old time are our assumptions. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat, for companies with 10 or more employees. Model cost about €0.007 per document: OCR in our system cost €1.17 for 181 files, and a separate read about €0.0008, converted at the NBP rate of 21 October 2026 and the ECB rate of 1.1205 USD per euro on 22 October 2026.

See the full calculation →

This is how we implement it in companies

All case studies →

Client results.

All projects →
foap.comcustomer service

foap.com customer service: 95% of replies are sent automatically.

Users from several countries write to foap.com, in several languages. The system reads the entire conversation and the terms of that person, and then sends a response in that person's language. Questions about money and exceptions go to the team with a ready draft, so they have to check and send them.

the system sends the answers itself, without human typing
95%
time from the arrival of the e-mail to the ready response (median)
30 s
e-mails that require a response receive a ready-made draft
100%
less work for every 1000 answers (estimate)
47 h

Data from foap.com production, October 20-21, 2026. Estimate: 3 minutes for handwritten response.

How foap.com customer service works →
Sample messages. Money matters always come to the person.

Other implementations

jaktako.design

studio projektowe

−87%

Time to look for new customers, every day

4 hours → 30 min

approx. 73 hours less per month (21 days of work)

according to the studio

How the agent loop works →
Signature Estates

real estate

approx. −90%

Operator time spent browsing portals, per day

2–3 hours → 20 min

approx. 45 hours less per month (21 days of work)

respect · contact the seller 5 minutes after the announcement

How the agent finds property listings →
Kuba Koziej, CEO and co-founder of MoreGrowth and board member at Natu.Care

Kuba Koziej

CEO and co-founder, MoreGrowth

„We started with my own AI agent. During the training that followed, I learned to build new applications myself. Today I am bringing this way of working with AI into my business. Working with Syntalith gave me exactly the skills and tools I needed.”

View the original training review on LinkedIn ↗

Start with the process that takes the most time each week.Book a free process review (30 min) →

How we build automation and agents

From a conversation to a working system in four steps.

  1. 01 · 30 min

    Free process review

    €0

    You're talking to an engineer. You receive the written plan within 2 business days.

  2. 02 · 5 days

    Free prototype

    €0

    The system works on 20-100 cases that your team has already handled. You see every result.

  3. 03 · 2–16 weeks

    Build

    50%

    Fixed price and goal written in the contract. First installment upon signing.

  4. 04 · after reaching the goal

    Reception

    50%

    You pay the second installment when the system reaches its target on your cases.

How a free prototype is created

Mon.

We set a goal

What result do you consider good?

Tue–Wed

The system is working

It goes through your affairs, one by one.

Thu

We compare

We check every difference with the team's work.

Fri

Result and price

List of errors and fixed implementation price.

You choose the matters. We anonymize them upon request, they stay in the EU and we delete them after the test.

On the market, such a test usually costs from PLN 3,000 to PLN 8,000.

Book a free process review (30 min)

We establish a prototype during a free analysis.

Data checked on October 25, 2026.

„Is my data safe?”

Yes. You choose where they go: from your company account to the model on your server.

  1. 00Employee's private account

    Customer and company data ends up in an account that the company cannot see or control. There is no data processing agreement, so it is a violation of the GDPR. When an employee leaves, the chat history goes with them.

  2. 01ChatGPT or Claude corporate account

    The provider does not use your data to train models. A copy of the queries remains in his logs for up to 30 days.

  3. 02API with zero retention (ZDR)

    The provider processes the request and does not save anything. It requires his consent, we prepare the application together.

  4. 03Processing in the EU

    Queries are counted and stored in Europe: OpenAI in the EU region, Claude via clouds with the EU region.

  5. 04Model at your place

    The open model runs on your server or in your cloud in the EU. The data does not go outside the company.

What we do with every implementation

  • Confidentiality agreement and data processing agreement before we receive any data.
  • You choose the cases for the prototype and we anonymize them upon request.
  • The data stays in the EU and we delete it after the test.
  • Corporate accounts with an administrator on your side.

When data cannot leave the company

The open model runs on your server or in your cloud in the EU. To choose from:

White-tailed eagle
Polish model, SpeakLeash and ACK Cyfronet AGH
PLLUM
Polish model, HIVE consortium commissioned by the Ministry of Digitization
Qwen
Alibaba
GLM
Zhipu AI
DeepSeek
DeepSeek

We select a model for the task and check it on your cases. We run it, tune it and maintain it.

Source: Bielik and PLLuM, ACK Cyfronet AGH ↗ · PLLuM, Ministry of Digitization ↗

Do you have a lot of your own data? We will fine-tune the small model.

  1. 1 Your examples
  2. 2 tuning
  3. 3 test on new cases
  4. 4 model at your place

When you have thousands of similar emails, documents or reports, we teach a small open model based on your examples. It works for you, responds quickly and costs less to maintain than a large model. We measure quality on issues that the model has not seen while learning.

Przemysław and Piotr did it in large programming and consulting companies. Artem did it before and does it today at Syntalith. It's about models on your own infrastructure, tuning and quality measurement.

Private AI models →

Source: OpenAI, training on data from private accounts ↗ · OpenAI, who sees conversations ↗ · Anthropic, changes for private accounts (August 2025) ↗ · OpenAI, data in API ↗ · AWS Bedrock, EU profile ↗

You choose the protection level during the interview. We sign the data processing agreement before we see any data.Schedule a free analysis →

AI implementation

Start with the process that takes the most time each week.

The engineer chooses one process with you and calculates how much it costs today. Then we build a free prototype on 20-100 cases that your team has already handled. We write down the goal and the fixed price in the contract.

See AI implementation →
analysis with an engineer and a free prototype
€0
team in ChatGPT or Claude, up to 10 people
from €1,150
automation, agent or application for one process
from €3,000
Book a free process review (30 min)

or call+48 888 78 48 78WhatsAppMon–Fri 9:00–17:00 CET

Questions

  • What is a legibility gate before OCR?

    An image check before reading: it measures brightness, contrast and edge features. When the image is unfit, the system records that no reading was attempted and asks for a new shot, in place of returning fields of unknown reliability.

  • What does the Syntalith demonstrator show?

    Seven variants of an image of a transport document. Two legible ones go on to reading, and five are stopped without calling the model. It shows the mechanism and does not measure production quality.

  • How much does automated reading of transport documents cost?

    At Syntalith from €3,000 net. At 1,000 photos a month with 3 minutes of retyping each, the payback is 2.0 months.

  • How do you set the legibility threshold?

    On accepted and rejected photos from one process. Decide how many false rejections the team can handle and when a manual reading is needed, because a gate that is too strict shifts the work onto drivers.

Let's talk

Just write, call or choose a date.

We work fast because we hate corporate cotton wool. We don't have salespeople or managers to keep an eye on managers. You talk immediately to engineers who then build your system. We respect each other's time and yours too.

We will accept 3 more companies by the end of the year.

Write
contact@syntalith.ai
Call
+48 888 78 48 78
WhatsApp
write on WhatsApp

Mon–Fri 9:00–17:00 CET

What our OpenAI and Anthropic partnerships mean for you.

OpenAI makes ChatGPT and Anthropic makes Claude. We belong to the partnership programs of both companies, so we work with their tools every day and learn from the creators themselves.

OpenAI Select PartnerClaude Partner Network
01We will select a tool for the task
We implement both ChatGPT and Claude. We check both on your documents, and you choose the one that makes fewer corrections and saves more time.
02Knowledge straight from the creators
We learn from official OpenAI and Anthropic courses and materials. Your team learns about the features and settings that work today.
03Licenses without our margin
When we roll out ChatGPT, Claude or Copilot, you buy the licences directly from OpenAI, Anthropic or Microsoft and they invoice you. We add nothing to the licence.
04Certified engineers
On the team, we have an AWS certification in machine learning and an Anthropic certification in systems architecture on Claude.

Participation in programs does not mean that OpenAI or Anthropic endorses our services.