Blog · Published · Updated
The process changed. Does the model need retraining?
Short answer
A process change does not always require retraining the model. If only the team name or the recipient changed, a mapping change in the app is enough, about €70 of work. If a category changed its meaning, you need new examples and retraining, about €1,082 in the scenario. Keep changing assignments outside the model, so that a reorganisation does not need training.
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
AI models in companies: statistics 2026
€1,082
is what retraining after a category change costs in the scenario, against €70 for changing the mapping in the app.
40%
According to Gartner (2025), over 40% of agentic AI projects will be cancelled by the end of 2027 because of costs, unclear value or weak risk controls.
48%
According to the European Commission (2026), 48% of EU SMEs and mid-caps name a lack of expertise as a barrier to AI.
Source: European Commission, Flash Eurobarometer on AI (2026) ↗
14–56%
According to Anthropic (2025), AI cut task time by 14 to 56% in studies with a control group.
Source: Anthropic, Estimating AI productivity gains (2025) ↗
20.0%
According to Eurostat (2025), 20.0% of EU companies with at least 10 employees used AI.
Figures checked: 22 October 2026
A new team name or a new meaning for a category
A company merges two teams that handle customer documentation. The model still correctly recognises the case as a request for a copy of a document, but the app sends it to the old recipient. A mapping change in the app is enough. The model works well.
Another change splits an existing category into two kinds of cases, because the company starts handling them differently. The old labels do not contain this distinction. The team has to explain the new boundaries and label examples, and the model needs retraining or a new rule.
The recipient or the team name changed
mapping in the app
€70
The priority or the handling deadline changed
a rule in the app
a few hours of work
A category changed its meaning or split
new examples and retraining
€1,082
The way content is recognised changed
retraining and full tests
priced after a review
What each path costs
- New examples labelled by an expert
- 300 × 3 min = 15 h
- Engineer's work: data, training, tests
- 16 h
- People's time in total
- 31 h (€1,082)
- For comparison: a mapping change in the app
- 2 h (€70)
Difference
€1,012
The number of examples and hours are the scenario's assumptions. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat. Count the cost of the training run itself and of the infrastructure separately.
A general line item called „retraining after a process change” can cover work the company does not need, or leave out the expert's time, without which the training will not succeed.
What to keep about the previous version
Hugging Face recommends describing in the model card the model's use, its limits, the data used to train it and the evaluation results. Such a record makes the conversation about a change easier: you know which split of categories produced the historical labels. An earlier correct decision is no error just because today's process looks different.
- The model's use and its limits
- The category split in force at training time
- The training data and its period
- Evaluation results on the test set
- A list of cases the team still checks by hand
How to tell what you need
Ask three questions. Does the model still name the case correctly, and only its recipient changed? That is a change in the app. Has the company started telling apart cases the model did not tell apart before? That means new examples and retraining. Have the documents themselves changed, for example a new form? That means a test on the new documents, and only then a decision.
The most expensive mistake is ordering retraining when a mapping change was enough. The second most expensive is fixing the mapping when the category really did change its meaning: the model will keep confusing cases, and the team will lose trust in the system.
Upkeep includes people's work
The update budget depends on the number of significant changes, the availability of checked examples and the review time. The process expert explains the new decisions, and the contractor checks what the model can do, prepares the data and the tests. The new version also has to be checked on cases the change did not touch.
Describe the next process change and the result the model would have to prepare differently after it. That is the starting point for agreeing the scope of system maintenance. Building a new system is described on the AI apps page.
Worked example · Scenario
How we count the payback: document workflow
In the review we also count the cost of upkeep after process changes.
€2,327
a month for matching 1,000 documents to cases by hand
Payback: 1.5 mo.
4 minutes per document matched to a case 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 a document in our case-file system (about €0.0002) plus matching it to the case and a note on what is missing (Claude Sonnet 5.5, about 2,000 + 300 tokens), at the ECB rate of 1.1205 USD per euro on 22 October 2026.
See the full calculation →
