The answer exists, in a PDF nobody opens
A sales assistant that answers out of the manufacturer’s own datasheets. A technical catalogue where a wrong answer gets buried under a concrete slab. We built an assistant that answers from the company’s own documents, refuses when the documents do not cover it, and gets corrected by the service team rather than by developers.
The problem
This is a catalogue where being confidently wrong is expensive. "Will this membrane work below grade?" answered vaguely means the wrong product goes under a slab, fails, and becomes a five-figure remediation and a lost customer. That is why a generic chatbot was never on the table: something that guesses about chemical compatibility is not a convenience, it is a liability. The knowledge to answer properly already existed and was almost never read, and the small customer-service team covering the whole technical range could not be the only route to it.
What we built
- Built the assistant on retrieval over the manufacturer’s own published material rather than a general model: product data and live pricing from the commerce system, PDF datasheets and safety sheets, guides, FAQs and 69 how-to videos
- Made the video library searchable the same way as the text: captions where they exist, automatic transcription where they do not, and frame-by-frame description for silent clips, so an answer can point at the right minute of the right video
- Gave every source document a fingerprint so the nightly refresh skips what has not changed, which is what makes re-reading everything every night affordable
- Scored retrieval confidence before answering, so that below a threshold the assistant says plainly that it does not have documentation to answer and brings in a person instead of guessing
- Engineered the handoff as a feature: it tracks who is actually at a desk, answers what it can before escalating, collects an email and sends the team a written brief when nobody claims the conversation, and re-alerts on repeat requests rather than silently deduplicating them
- Built a correction loop where a service agent thumbs down a reply, writes a plain-language note, and an approved fix is live in retrieval seconds later with no developer involved
- Put commerce on a leash the client controls: product cards only for items recommended by name, carts validated against what the visitor was actually shown, and a discount ladder the ecommerce manager defined rather than the bot volunteering the best price
Result
- Technical questions answered in seconds, around the clock, against an email baseline measured in hours to days
- The service team reserved for the conversations that actually need a person
- A knowledge base that maintains itself: edit a product page at 4pm and the assistant knows by morning
- Corrections owned by the people who spot them, applied by the service team in one click
- A bot that talks a ready-to-buy customer out of the wrong product, which is the reason the client trusts it with the right ones
- No per-seat or per-conversation licensing: the client owns the system and the knowledge base
Automation running here
- A support agent grounded in real product dataAn automated assistant answering from a manufacturer’s own catalogue and documentation rather than a generic model, so the specification questions that used to interrupt the sales team get answered correctly at 9pm.
Product data nobody can actually build on
The catalogue of record is a print document, a spreadsheet, and someone’s memory, and the three disagree. Nothing downstream can be better than the data underneath it, so the storefront, the search filters and the ERP all inherit the same mess.
Read the case →Orders nobody can price at checkout
A large share of your volume ships LTL, and nothing can be quoted until someone looks at the pallet. So the order gets taken by phone and typed in again, and the website quietly becomes a brochure.
Read the case →