Blog · Published · Updated
Does the source support your AI's answer?
Short answer
You check a citation in a company AI answer at three levels: whether it leads to a passage, whether the passage supports the whole statement and whether the answer is complete. In a sample review, 90% of answers have a link, and 60% are supported and complete. An AI application on company documents costs from €3,000 net at Syntalith. At 300 employee questions a month and 10 minutes each, 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
- net per application, fixed price before contract
- from €3,000
- from signing the contract to acceptance
- 4–10 weeks
- free prototype on your documents
- €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
Citations in company AI: statistics 2026
2.0 mo.
payback of €3,000 at 300 employee questions a month: €1,745 of work today, €267 with checking and the model. Scenario.
49%
According to Anthropic (2024), combining contextual embeddings with BM25 search cut failed retrievals by 49%, and by 67% with reranking.
11 of 13
According to the NoLiMa study (2025), at 32,000 tokens of context 11 of 13 models fall below half of the accuracy they reach with a short context.
48%
According to the European Commission (2026), 48% of small and medium-sized companies in the EU name a lack of expertise as a barrier to AI.
Source: European Commission, Flash Eurobarometer 578 (2026) ↗
15%
According to Brynjolfsson, Li and Raymond (2025), customer support agents with AI help resolved 15% more issues an hour on average, in a study of 5,172 people.
Source: Brynjolfsson, Li and Raymond, The Quarterly Journal of Economics (2025) ↗
Figures checked: 22 October 2026
A condition that vanished from the answer
The instruction says: „Materials may be sent to the client after the account owner confirms that they contain the current product catalogue”. An employee asks whether she may send the prepared deck. The assistant says yes and adds a link to the instruction. The link leads to the right document, but the answer left out the confirmation condition.
- Answers with a link to a source
- 90 of 100 = 90%
- Fully supported by a passage
- 72 of 100 = 72%
- Supported among answers with a link
- 72 ÷ 90 = 80%
- Supported and complete
- 60 of 100 = 60%
The gap between having a link and a good answer
30 percentage points
Hypothetical figures. The instruction's owner does the assessment on the questions employees keep bringing to them.
A link to the start of a long manual leaves the employee to find the passage. A useful view opens the right passage with its context. In a longer answer, some statements may be supported by the source and some not, and the application should show the gap next to the sentence concerned.
What it costs and when it pays back
- Today
- 300 employee questions × 10 min = 50 h (€1,745)
- With the system, a person checks the result in 15% of the time
- €262
- Model, 1 call per case, Claude Sonnet 5.5
- €5
- Cost of one case
- €5.82 today, €0.89 with the system
- Saving
- €1,478 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: 6,000 input tokens and 500 output tokens per call at Anthropic's rates, at the ECB rate of 1.1205 USD per euro on 22 October 2026.
The model is a small item here. The result depends on the time that searching and asking around take today, and on how much of it remains for checking the answer. A missed condition costs more than the search, because the material reaches the client without confirmation. We count such answers separately in the review.
Before we price the build, we run a test on 20 to 100 questions the team has already asked: we compare the application's answer with what the employee found by hand, and we count the time from the question to the source. The test takes 5 working days and costs €0.
Three different quality measures
Microsoft, describing the evaluation of RAG systems, separates the relevance of the retrieved context, how well the answer is grounded in that context and how complete the answer is. Automatic measures help with a large number of questions, but a person who knows the area will name the condition that must never be left out. When a user spots a missing condition, the report keeps the question, the answer and the passage, so the error can be reproduced.
Hundreds of questions to monitor
automatic grounding and completeness measures
part of maintenance
Acceptance before rollout
the owner's review on 20 to 100 questions
test €0
Errors reported by users
a report with the question, the answer and the passage
part of the application
Where to start
We run the first review on the questions employees keep bringing to the owners of instructions. The owner names the condition that must never be left out and the point where the document stops being enough.
For the first call, one question that keeps coming back to colleagues and where the team looks for the answer are enough. You do not need to send confidential documents, and you receive the written plan within 2 working days.
Worked example · Scenario
How we count the payback: answers from company documents
We run the same count for your documents in the free review.
€3,723
a month for finding answers to 800 questions in documents
Payback: 1 mo.
Our assumptions are 8 minutes to find an answer in documents or ask a colleague, and checking in 15% of the old time. €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: Claude Sonnet 5.5, about 6,000 + 500 tokens per answer from documents, at the ECB rate of 1.1205 USD per euro on 22 October 2026.
See the full calculation →
