Chat + voice. Your manuals, finally usable.

Every “AI for documents” demo answers a hand-picked question on a hand-picked PDF. Real service corpora have something harder: superseded bulletins, model-year scoping, ACL boundaries separating engineering from sales, and three different things called “E-12.” The question is not whether the agent answers a demo question correctly. It is whether it answers the right question for the right machine when the technician is standing next to it.

How it works

The agent ingests your service corpus — manuals, service bulletins, parts catalogues, ticket history — and tags every document at ingest: effective date, supersession chain, applicable model range, applicable serial range, owning team. None of that tagging is optional; it is what makes the retrieval trustworthy.

When a question arrives, the system narrows on metadata before the language model sees candidate documents. A question scoped to a 2022 diesel variant does not surface bulletins written for the 2019 model or the electric successor. The answer comes back with a document-level citation — the source PDF and page, not a highlighted line. The reviewer sees the answer in context: the warning above it, the torque spec below it, the revision date in the header. That context is why document-level citation earns more trust than line-level citation. A paraphrase the model got wrong is visible. A line-level excerpt hides the surrounding page and quietly trains reviewers to stop checking.

Same retrieval runs on chat and voice. Same access-control lists. Same audit log.

When “E-12” means three things at once

Across a single customer’s product lines, “E-12” is a fault code for low oil pressure on the older diesel product line, a battery management warning on the newer electric product line, and the SKU prefix for a family of wiring harnesses in the parts catalogue. Ask a general-purpose LLM what E-12 means and you get a confident answer — usually the one most represented in its training corpus — and it is wrong for the machine in front of the technician.

The agent does not guess between those three meanings. It narrows on product line, model year and document type before selecting candidates, then returns the right answer with the right citation.

The same narrowing applies to language. The agent supports more than 40 languages, including the Nordic and Eastern European mixes where most service organisations have gaps. A Polish technician submits a question in Polish. The relevant manual is in German. The agent translates the query, retrieves against the German corpus, produces the answer, and responds in Polish with a link to the German source page. The technician gets the right answer in their language. The citation still points to the authoritative German document. Nothing gets lost in the routing.

By the numbers

At Nize Equipment, deploying the Knowledge Agent produced a 60% drop in L2 escalations within 90 days, 4.1× faster technician onboarding, and 92% weekly active users among the field team — the last number being the one that matters most, because it means technicians opened it the second time. Full case study at Nize Equipment.

What you control

Access. Role-based and per-document ACLs are applied at retrieval time, before the model sees candidates. Sales does not see engineering drawings. Field technicians do not see commercial pricing. Filtering before generation is the only way to ensure access boundaries hold.

Supersession. The agent refuses to cite a withdrawn bulletin. When a bulletin is superseded, it is tagged; the agent routes to the replacement. A confident answer citing a document pulled eighteen months ago is a trust failure, and it is preventable.

Audit. Every question, every retrieved chunk, every cited source, every model version, every calling user, every timestamp — logged. When a technician disputes an answer six weeks later, the retrieval is replayable.

Deployment. Cloud or on-prem. Regulated customers in sectors with data-residency requirements run the agent in their own environment. The behaviour is identical.

What it connects to

SAP, IFS, Microsoft Dynamics, ServiceNow, IBM Maximo, HubSpot — the agent sits alongside whichever ERP, FSM and CRM your team already uses. Answers can trigger work-order creation, parts lookups and ticket updates without the technician leaving the chat window. Read access is the default; write access is granted per system, with an audit record on every outbound call. When the connector you need does not exist, we build it — typical timeline two to four weeks. Full list at Integrations.

Where to go next

• Construction & Material Handling — how Knowledge Agent fits into dealer service operations, with call-volume patterns and escalation benchmarks.
• Service use case — the service-manager view: how the agent changes L1 / L2 split, onboarding time and technician retention.
• Operational trust — the full argument for why corpus tagging, citation pattern and audit log are the three things that determine whether technicians open the agent the second time.