How Salus Controls Romania cut unanswered customer questions by 60%
Salus Controls Romania supports more than a thousand heating controls, thermostats and connected devices, many with near-identical model codes. Its virtual assistant was answering about the wrong product, and in more than a third of conversations it was telling customers it could not find an answer at all. Svennis diagnosed the cause, repaired the retrieval layer and moved the engine to Claude. A month of production traffic later, that failure rate had fallen by 60%.
The story in short
Sally is the virtual assistant that answers product and technical questions on the Salus Controls Romania website, in Romanian, at any hour. It was answering, but too often about the wrong thing. A customer who had selected an SQ610RF thermostat could be walked through a procedure belonging to a different model, and customers were told a manual could not be found when that manual was published and indexed.
Working with Svennis Cloud Solutions, Salus rebuilt Sally on dedicated infrastructure with Claude as its conversational engine. The fix was not only the model. Svennis diagnosed the retrieval layer, found that only 57% of the knowledge base carried a usable product code, repaired the tagging across the entire corpus, and scoped every search to the customer's own product.
fewer conversations ending in “I could not find an answer”, measured over the first 30 days in production
retrieval accuracy on a fifty-question Romanian benchmark
of conversations resolved without escalation to a human agent
What actually happened in production
The honest test of any of this is what customers experienced afterwards, so the outcome was measured against the live conversation record rather than estimated. The first 30 days on the new system, against the 30 days immediately before it:
| Before | After | |
|---|---|---|
| Conversations | 250 | 307 |
| Ended in “could not find an answer” | 89 | 44 |
| Failure rate | 35.6% | 14.3% |
| Questions per conversation | 3.22 | 3.24 |
That is a 60% reduction in the share of customers who came to Sally and left without an answer. August, the first full calendar month on the new system, closed at 13.9%, against a baseline that had run between 34% and 45% from April to June.
Two details matter for reading those numbers honestly. Conversation volume went up across the two periods, not down, so the improvement is not the result of quieter or easier traffic. And questions per conversation stayed flat, so customers are not simply giving up earlier.
One caveat, stated plainly rather than buried: the retrieval repair and the change of engine reached production within 48 hours of each other, and nothing in the data separates their individual contributions. This improvement is the result of the whole programme, not of the model change on its own.
The challenge
Sally was answering, but it was too often answering about the wrong thing. Three problems sat underneath that:
Salus model codes are alphanumeric and densely similar: RT510, RT510RF, RT510TX, SQ610, SQ610RF. To a semantic search these are near neighbours, so a general question returned results drawn from six different products at once.
Only 57% of the indexed content carried a product code, and the gap fell hardest on the PDF manuals, which is exactly the material a technical customer needs. Where codes existed they were written inconsistently, the same product recorded five different ways.
Because searches were not scoped to the customer's product, the correct manual often ranked below the cutoff and never reached the assistant. Separately, around 9% of the published Romanian knowledge base was never indexed at all.
The result was a support assistant customers could not fully trust, in a product category where a wrong wiring instruction is not a trivial error.
The solution
Svennis Cloud Solutions took the whole system: the infrastructure, the retrieval layer, the conversational engine, the compliance layer and the management dashboard.
The work began with a diagnosis rather than a rebuild. Svennis reproduced the customer complaints against the live search index, established that the widely suspected cause was not in fact the problem, and identified the real one: the assistant received a mixed bag of products on every search because nothing constrained retrieval to the product the customer had selected.
The repair had three parts. Every article and every attached manual was matched against the full Salus product master and stamped with the codes it actually describes, lifting coverage from 57% to 98.5%, with manual attachments inheriting the codes of their parent article. Retrieval was then scoped to the customer's product, inferred from the conversation itself, with a deliberate fallback so general questions are not over-restricted. Finally the indexing pipeline was fixed so newly published articles are stamped correctly on arrival, and short articles stop being silently skipped.
Why Claude
The move to Claude was tested rather than assumed. Svennis built a 68-case benchmark from real Salus conversations and ran both engines through it with the same test suite and the same judge.
In the cases that had previously ended in "I could not find that", Claude recovered more of them than the alternative engine.
When a customer asks for an installation guide, a user manual or a wiring diagram, Claude reliably surfaces the right document rather than a plausible neighbour.
Sally advises and points to documents. It does not invent a procedure, a ticket or a status, and it answers in Romanian regardless of how the question is phrased.
Claude also works behind the scenes
Claude does more for Salus than answer customers. It runs two monthly analyses that make the assistant better without anyone writing new code.
Each month Claude reads every Sally conversation, isolates the ones it could not resolve, and checks each against the entire knowledge base to establish whether the answer genuinely does not exist. In the first pass, covering roughly a thousand conversations, Claude identified around 20 verified content gaps accounting for about 88 customer conversations.
In the same pass Claude reports the opposite failure: questions where the answer already existed but Sally did not surface it, naming the article that should have been used. It is a monthly feedback loop in which the assistant gets better at finding its own content, with no new content required.
Both analyses land on a management dashboard built for Salus, with an all-months overview, per-month detail and export, so the team can see precisely where to spend its content effort.
Compliant by design
Sally speaks to Romanian consumers, so the assistant was built to meet EU and Romanian obligations rather than have them retrofitted.
Sally identifies itself to every customer as an automated virtual assistant before the conversation starts.
National identification numbers, IBANs, email addresses and phone numbers typed into free text are detected and masked before the conversation is stored or sent to the model.
Conversation records are deleted automatically after 18 months, and a specific customer's conversation can be exported or permanently erased on request.
Conversation data is held within the European Union. Sally stays advisory: it can suggest opening a ticket but makes no warranty, refund or credit decision.
How it works
A chat widget on the Salus Controls Romania website, answering in Romanian at any hour, with no login required.
Claude Sonnet 5, through Anthropic's platform, deciding how to handle each question and which document to answer from.
Semantic search over several hundred published articles, manuals and wiring diagrams, scoped to the customer's own product with a fallback for general questions.
Every article and attachment stamped with the product codes it describes, kept current by a fixed indexing pipeline rather than by manual repair.
Dedicated Svennis infrastructure: Google Cloud hosting, Neon managed Postgres in the EU, Elasticsearch for retrieval.
The impact
About two-thirds of conversations are resolved without escalation, on live Romanian traffic.
Retrieval accuracy rose from 52% to 87%, measured as the share of the top five retrieved documents that are relevant, across a fifty-question Romanian benchmark.
Product coverage rose from 57% to 98.5%, with 1,508 manual and attachment sections inheriting the codes of their parent article.
Roughly 9% of the published Romanian knowledge base was invisible to the assistant and is now indexed, including wiring diagrams and compatibility articles customers actually ask for.
A conversation that had previously failed, an SQ610RF pairing question answered with another product's procedure, now returns the correct steps with the correct sources.
9,127 Romanian conversations in the twelve months to July 2026, within 36,467 across all six European regions.
What comes next
Romania is where this platform was designed, built and proven. Salus operates Sally across six European regions, and the same architecture, the same compliance layer and the same monthly Claude analyses are being extended to the rest of them.
About Salus Controls Romania
Salus Controls Romania is the Romanian arm of Salus Controls, a manufacturer of heating controls, thermostats, underfloor heating systems, gateways and connected smart-home devices. Its product range runs to well over a thousand distinct item codes, sold to installers and to homeowners, and supported through a Romanian-language knowledge base of several hundred published articles, manuals and wiring diagrams.
About Svennis
Svennis Cloud Solutions designs and delivers AI-assisted business systems built on Claude, and is currently going through onboarding to the Claude Partner Network, Anthropic's partner programme.
For Salus Controls Romania, Svennis delivered the entire Sally platform: the infrastructure, the chat experience, the retrieval and indexing layer, the Claude integration, the EU AI Act and GDPR compliance layer, and the management dashboard, from diagnosis through to live production and ongoing support.
Wondering what an assistant like this would find in your own knowledge base?
Describe your business in one sentence and the Mirror shows you where AI would pay off first. No form, no sales call.