The challenge
A programmer needs the exact lighting conditions for imaging a part before placement on a PCB. A technician mid-repair needs to know which grease to use, and how much. An operator staring down a stalled line needs an answer now, before the delay turns into real downtime.
All three reach Fuji’s manuals through the same members-only customer portal, which holds machine and software manuals, upgrade reports and specification guides, refreshed regularly with minor revisions. A keyword search across a library that size could only ever get each of them partway there. Users needed to enter the appropriate keywords to locate relevant information, and the people searching were the least able to guess at them, because they are mid-repair or mid-outage.
Finding the page took around twenty minutes of scrolling dense documentation. Then most people checked the answer somewhere else anyway, because nothing showed where it had come from.
The solution
FUJI Smart Doc is a set of AI chat agents organized by machine model line, grounded only in Fuji’s own documentation, so an answer can never come from anywhere else.
Three languages, Japanese, English and Chinese, were each held to the same bar Fuji set rather than shipped as translations of the English.
The corpus keeps itself current: documents uploaded to Fuji’s portal flow into the knowledge base without a separate re-upload step, so approximately 135,000 documents stay up to date without hand maintenance.
The whole thing runs inside Fuji’s own members-only portal and secure logon, serving customers, sales partners and internal staff from one system rather than a separate tool each.
The capability that decided it
Every answer opens the exact page it came from.
Every machine documentation agent has citations enabled. An answer comes back with a citation, and clicking it opens PDF Viewer at the exact page the answer was drawn from, the relevant text highlighted and any diagrams shown in place.
Fuji evaluated multiple vendors on answer quality and on this one capability, which no other vendor offered. PDF Viewer existed nowhere else, so CustomGPT.ai and Fuji worked in close, ongoing dialogue to get an initial version running before anything else was built.
For a manufacturer whose maintenance and safety decisions turn on the details, that is what lets a technician act on an answer rather than go and check it somewhere else. It remains Fuji’s non-negotiable requirement wherever a wrong answer could lead to a maintenance or safety issue.
The results
FUJI Smart Doc has answered 157,701 questions across 43,898 conversations, drawing on approximately 135,000 indexed documents. Current volume runs 10,000 to 14,000 queries a month, individual user numbers have doubled over the past year, and the time to find an answer has gone from around twenty minutes to near-instant.
Ninety percent of customer-facing conversations now end in a usable answer, and 84% across every chatbot Fuji runs, including the internal and partner-facing ones.
Fuji sees the remaining 16% as room for improvement. Conversations that end in a fallback get reviewed, and either the person who asked gets a follow-up, or the source content gets fixed so the same question does not fail twice.
Fuji Corporation describes the effect this way: “We have had customers come up to us unprompted at trade shows saying how much they appreciate having FUJI Smart Doc. Because the information needed for troubleshooting and maintenance can now be accessed immediately, field engineers and customer support teams can respond significantly faster, which has contributed to improved customer satisfaction. It has been a popular addition to the services we provide.”
Why it worked
Verification, alongside the answer. PDF Viewer is what won permission to deploy in a setting where a wrong answer could lead to a maintenance or safety issue.
A knowledge base that keeps itself current. Manuals flow from Fuji’s own portal straight into the knowledge base, so approximately 135,000 documents stay up to date without extra work.
Refined through real use. PDF Viewer kept improving on Fuji’s ongoing feedback, including through the months it took to bring Japanese to the same bar as English.
Built and tested in the right order. The deciding capability came first, the first agents were tested for months and released to sales partners before any customer saw them, and the gate opened last.
Operators, technicians, and programmers can now locate relevant information more efficiently than through keyword searches alone. Usage has increased since the customer launch, and the knowledge base keeps growing as new documentation is published.
| Measure | To date |
| Questions answered | 157,701 |
| Conversations | 43,898 |
| Technical documents indexed | ~135,000 |
| Languages served from one system | 3 |
| Monthly queries, current | 10,000 to 14,000 |
| Time to find an answer, before and after | ~20 min, now near-instant |
| Answer rate, customer-facing chatbots | 90% |
| Answer rate, all chatbots | 84% |
Figures from CustomGPT.ai product usage records and from Fuji Corporation.