Blog · Published · Updated
Buying local AI with a test on your documents
Short answer
Before you buy a local document assistant, compare its answers to your own questions. Does it point to the right version of the source, does it admit when information is missing, and how long does checking take? At Syntalith the app costs from €3,000 net. With 500 questions a month and a GPU server at €214, it pays back in 2.0 months.
How we workDetails ↓
- 01Call30 minutes with an engineer, free
- 02PrototypeA working system on your process and data, free
- 03ImplementationTarget and fixed price in the contract; 50% at signing
- 04Acceptance50% once the agreed target is met
- model on your server or in the cloud in the EU
- quoted after the review
- same work in application with EU cloud model
- from €3,000
- task, data and equipment analysis
- €0
We belong to the OpenAI and Anthropic partner programmes and learn through their official courses. What this means ↓
Author: Syntalith · 3 min read
Statistics
Local AI and company documents: statistics 2026
2.0 mo.
payback of a €3,000 local app at 500 questions a month of 8 minutes each: €2,327 of work today, €796 with checking and a GPU server. Scenario.
230,085
According to a Syntalith measurement (2026), Qwen3.8-27B on an RTX 3090 card recalled 3 of 3 checked facts from an input of 230,085 tokens, and the run took 563.4 seconds.
Source: Syntalith, long-context measurement of Qwen3.8-27B (2026) ↗
146
According to Bartanowicz and Jassem (LREC 2026), in a RAG system on Polish technical documentation assessed on 146 reference questions, Bielik-11B-v2.3-Instruct consistently beat PLLuM-12B-nc-chat.
83%
According to Smits and Walkowiak (EACL 2026), in RAG on the Polish Wikipedia assessed on 1,000 PolQA questions, PLLuM-12B matched the quality of 70B models with 83% less energy.
48.8%
According to Eurostat (2025), 48.8% of EU companies that considered AI and did not use it named concerns about data protection and privacy.
Source: Eurostat, Use of artificial intelligence in enterprises (2025) ↗
Figures checked: 22 October 2026
One question, two versions of a document
A product demo and a benchmark score do not tell you whether a tool will answer questions from your company's documents. Take a procurement team that compares supplier quotes. The folder holds an older quote with a delivery date and a newer one that changes that date. An employee asks: which delivery date applies to the October order?
The assistant may find the older quote and give its date in a fluent, confident tone. The sentence rests on a document, yet the conclusion about the date that applies is wrong. This example becomes a test task: the question, the set of sources, the document version, the expected answer and a description of the errors that disqualify the result.
- 01Questionthe employee's own words from a real case
- 02Sourcesthe documents the assistant can reach, with version and date
- 03Expected answerthe date from the newer quote, with a quotation
- 04Disqualifying errora date from the older version, or one guessed when an amendment is missing
- 05Checking timethe employee's minutes from the answer to a checked finding
Criteria written down before the result
We take the tasks from real mistakes and real employee questions. For each one we record the allowed sources, the expected result and how the basis must be shown. The trials used to improve the system stay apart from the tasks reserved for acceptance. Anthropic's guide to evaluating agents recommends fixed tasks with a clear success criterion and an assessment over many runs, because a single trial easily misleads.
- Test conditions: models, versions, the range of documents and the date
- Accepted and rejected examples, with the reason for each rejection
- Source retrieval scored apart from answer quality
- The time an employee needs to check the result
- The range of tasks the decision applies to
Testing local AI on equal terms
Three options can take part in the test: today's document search, an approved cloud service and a local system. Each gets the same questions, the same range of documents and the same criteria. If the assistant cannot find the newer document, the problem lies in retrieval or in the order of the documents, and a different model will not help. Our long-context measurement showed that Qwen3.8-27B on an RTX 3090 recalled 3 of 3 facts from 230,085 tokens. The run took 563.4 seconds. For questions asked several times a day, the response time counts too.
- Today
- 500 questions × 8 min = 66.7 h × €34.90 = €2,327
- With the app, an employee checks the output in 25% of the time
- €582
- Hetzner GEX45 GPU server
- €214
- Saving
- €1,531 a month
Payback of €3,000
2.0 months
The volume, the time and the checking share are the scenario's assumptions, which we measure in the test. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat. Price of the GEX45 with 24 GB of GPU memory from Hetzner's website on 22 October 2026, without the one-off setup fee of €209.
What to prepare before the call
A failed test is also a result. A wrong source version points to the order of the documents, and a guessed date points to the instructions or the model. If the test confirms that you need a local system, the next decision is the hardware, described in local AI on a workstation or a server.
Documents may go to an approved cloud
a cloud service with search that shows the source
from €3,000
Documents may not leave the company and there are many questions
a local app on a GPU server
from €3,000 plus the server
The assistant cannot find newer versions
put the documents in order before choosing a model
the team's own work
We build local assistants as part of private LLMs. For the free process review, prepare two or three questions your employees regularly look for answers to, and say which documents they use today.
Worked example · Scenario
How we count the payback: the procurement team
We run the same count for your documents in the free process review.
€2,792
a month for checking 800 supplier invoices by hand
Payback: 1.3 mo.
6 minutes to compare an invoice with the order and the goods receipt 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: €0.0040 per invoice, measured in our invoice coding system (0.004514 USD), at the ECB rate of 1.1205 USD per euro on 22 October 2026.
See the full calculation →
