Blog · Published · Updated
Fine-tuning or RAG: what will improve product answers?
Short answer
When AI gives out-of-date product availability, a current source is missing, and fine-tuning will not supply one. You give the model changing data from the system at every question. In a scenario of 1,200 product answers a month, a current data source cuts errors from 10% to 3%, and fine-tuning to 2%, which together save 6.4 hours of corrections, or €223 a month.
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
Product data and AI: statistics 2026
6.4 h
fewer corrections a month at 1,200 product answers: a current data source cuts errors from 10% to 3%, fine-tuning to 2%, at 4 min per correction. Scenario.
49%
According to Anthropic (2024), combining contextual embeddings with BM25 search cut failed retrievals by 49%, and by 67% with reranking.
20.0%
According to Eurostat (2025), 20.0% of EU companies with at least 10 employees used AI.
Source: Eurostat, 20% of EU enterprises use AI technologies (2025) ↗
−87%
According to Anthropic (2025), AI cuts the time of tasks on invoices, memos and other documents by an estimated 87%.
Source: Anthropic, Estimating AI productivity gains (2025) ↗
14–56%
According to Anthropic (2025), studies with a control group found smaller time savings, from 14% to 56%.
Source: Anthropic, Estimating AI productivity gains (2025) ↗
Figures checked: 22 October 2026
Is a current fact missing, or is the model using it incorrectly?
A sales assistant describes a product as available, although the order system marks it as unavailable. The team thinks about training the model further. Further training on old answers will not supply today's order status, though: a model trained in October does not know what changed in November. What you need is the source where the company updates availability, read at every question.
It is different when the model received the status „unavailable” and still told the customer the product could be ordered right away. The information arrived, but the answer ignored it. That is a problem with the model's behaviour, which better instructions fix, and fine-tuning when the error keeps coming back. OpenAI distinguishes improving the context when information is missing or out of date from improving behaviour when a task is done inconsistently.
- Did the model receive the current product status at all
- Is the source it reads the one where the company updates the data
- Does the model still answer badly when the status is correct
- How many answers employees correct, and why
What a simpler change gives, and what training adds
- Today: 10% errors × 1,200 × 4 min
- 8 h (€279)
- After a change without training (a current data source): 3%
- 2.4 h (€84)
- After fine-tuning: 2%
- 1.6 h (€56)
- The simpler change saves 5.6 h, fine-tuning adds 0.8 h (€28) a month
Compare fine-tuning with the cost of examples, training and a new evaluation
The error shares and the correction time are the scenario's assumptions, which we measure on the client's cases not used for training. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat.
In this count, a current data source gives most of the effect for a fraction of the cost. Training makes sense at a larger scale or when the remaining errors are expensive.
- Evaluation cases kept apart from training examples
- The same set for the current model, the simpler change and the adapted model
- Correct answers prepared by the people who correct the results today
- Errors counted separately for each kind of mistake
- A plan for a new evaluation after a change of base model or process
The set of training examples has to be maintained: when the process, the naming or the base model changes, the evaluation is repeated on the same test set. We put that cost in the quote together with the training itself.
When an employee's existing tool is enough
For an occasional comparison of approved product sheets, a ready tool with access to files is enough, for example ChatGPT Work or Claude Cowork. For product availability, a direct read from the system that keeps the status is better than creating a document and adding it to a search engine. A separate app makes sense when the same way of fetching data is needed in many answers a day.
Data changes every day
a read from the system at every question
integration
Product sheets and specifications
RAG on approved sheets
from €3,000
Correct data, wrong form of answer
instructions, then fine-tuning
after a trial
Define the scope around the team's answers
We start with the answers employees correct today and with where the right information was at that moment. We adapt models as part of our private LLMs, and we build the app where an employee sees the result next to its source as part of our AI apps. For the free analysis, a few examples of an error that keeps coming back and a note of who corrects it today are enough.
Worked example · Scenario
How we count the payback: AI in e-commerce
We run the same count for your products in the free analysis.
€2,792
a month for handling 800 orders that bypass the checkout
Payback: 1.3 mo.
6 minutes to retype an order or answer a question 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: reading the order in our system (about €0.0006) plus matching the lines (Claude Sonnet 5.5, about 3,000 + 500 tokens), at the ECB rate of 1.1205 USD per euro on 22 October 2026.
See the full calculation →
