AI assistant
How a paint manufacturer stopped keeping a technologist on incoming dealer requests
Paint and coatings manufacturing
TLDR
Dealers wrote to the plant technologist every day - 330 requests per month. Product selection, consumption calculation, dilution questions, and compatibility checks. 65% of them were standard: the answers exist in TDS and process cards. Standard requests alone took about 54 hours of the technologist's time per month. BackendFlow developed an AI assistant based on the plant's full product documentation.
Result: the assistant handles 75% of requests autonomously. The technologist works on non-standard cases - the ones that require real expertise.
Context
Dealers call and write to the plant technologist every day. Product selection for a specific substrate, consumption calculation for a site, dilution questions, compatibility, application technology. About 330 requests per month.
65% of them are standard. The answers are in TDS, process cards, and instructions. But it is easier for a dealer to message the technologist than search through documentation.
For the technologist: at a rate of about $7.30/hour, standard requests alone took around 54 hours per month. Non-standard requests - formulation for unusual substrates or selection for specific conditions - waited in the queue for 1-2 days. These are exactly the requests that require his qualification and create value.
For dealers: the answer to a simple question came after several hours, sometimes the next day. This slowed their work with customers on site.
Task
Handle standard dealer requests without involving the technologist - quickly and with correct answers. Route non-standard requests to the technologist with prepared context. The technologist should work on what truly requires expertise.
Solution
BackendFlow developed a RAG-based AI assistant trained on all product documentation of the plant: TDS, process cards, tinting bases, and the archive of several years of correspondence between the technologist and dealers.
Workflow:
- A dealer writes a question in Telegram or a web form
- The assistant classifies the request as standard or non-standard
- For a standard request, the answer is generated from the knowledge base within 1 minute
- For a non-standard request, it is routed to the technologist with prepared context: what was asked, which product, and what was found in the knowledge base
- The technologist answers a non-standard request in 40-45 minutes instead of 90
For dealers, only the format changed: they write to the same channel and receive an answer immediately. The system is deployed on the plant's servers.
Insights
The knowledge base was the most labor-intensive part of the project.
Building the assistant took less time than collecting and marking up documentation. TDS files for different products were in different formats, and part of the technologist's answers existed only in messenger correspondence. The quality of the knowledge base directly determines answer quality: garbage in, garbage out. Structured documentation became a standalone asset.
The assistant revealed gaps in documentation.
Some dealer requests could not be answered by the system not because they were non-standard, but because the answer did not exist anywhere in the documentation. This became a signal for the technologist: he added answers to frequent questions that previously existed only in his head.
The technologist's load decreased - and this was visible even without metrics.
In the first weeks, the technologist noted fewer interruptions. He stopped answering the same questions several times a day. Data for precise financial verification is still accumulating, but the subjective effect was visible immediately.
Results
| Metric | Before | After |
|---|---|---|
| Standard requests without the technologist | 0% | 75% |
| Response time for a standard request | several hours | < 1 min |
| Technologist load on standard requests | ~54 h/month | ~13 h/month |
Assumptions: 330 requests/month, 65% standard, adoption 75%, technologist salary about $900/month. Conservative scenario. Full verification after a quarter of operation.
Want the same kind of calculation for your production?
Book an audit ->