When three companies become one, their knowledge doesn't merge with them. This company was built from three businesses across three continents, each with its own products, its own teams, and its own way of documenting how things work. The support organization had to keep three legacy systems running while building toward a single unified product, and answers agents needed were scattered across separate Zendesk knowledge bases and an internal wiki nobody officially owned. The company turned to Implicit to consolidate that knowledge into one governed, source-backed layer, without pouring unverified content into an AI it couldn't trust.
The challenge
The external knowledge was in decent shape. Each of the three businesses ran its own Zendesk knowledge base, kept current by active technical writers and updated alongside product and feature releases. The internal knowledge was the problem. Years of institutional know-how lived in Confluence with no dedicated owner, no update cadence, and no governance, which made it genuinely hard to tell which internal content could still be trusted.
The company had recently hired its first dedicated content and knowledge manager to bring order to this, sitting inside the technical support team. The near-term goal was pragmatic: get the most-used articles current and keep them that way, while the broader organization worked toward sunsetting all three legacy systems.
At the same time, the support model was shifting from self-serve toward high-touch, white-glove service, which put internal agent enablement front and center. Agents leaned on Slack to ask each other questions, and Zendesk's native search struggled with the detailed, multi-article questions that came up most, the ones where the answer was buried deep in a long article or spread across several. Early experiments with generic AI hit the real blocker fast: unreliable, unsourced answers. If an agent couldn't trace where an answer came from, they couldn't trust it, and an AI answer nobody trusts is worse than no answer at all.
Legacy systems to unify
Three merged businesses, each with its own products, teams, and documentation, consolidating into one platform.
Articles in the first Implicit workspace
External help content alone in the initial pilot scope, with internal content still being scoped.
Languages in use
Support runs across multiple languages globally, so answers have to hold up across all of them.
The solution: a governed, source-backed knowledge layer
Rather than point a chatbot at every file it could find, the company used Implicit to build one curated knowledge layer per product, feeding it only content it had verified. Implicit's GraphRAG engine structured that content into a queryable knowledge base where every answer traces back to its source. From ingestion to answer, the flow runs through a governed pipeline:
-
1
Connect the real sources
Implicit ingests each product's Zendesk help-center content directly from its root URL and pulls in internal material from Confluence and Google Drive, unifying knowledge that used to live in separate silos.
-
2
Isolate by product
Separate workspaces keep each product's knowledge cleanly walled off, so the AI serving one legacy product never pulls from another's docs, and content can be added incrementally as it's verified.
-
3
Structure it into a graph
GraphRAG auto-generates a taxonomy of concepts and keywords and maps the relationships between them, so retrieval reaches the right passage even when the answer spans multiple articles or sits deep inside a long one.
-
4
Keep it live
Sources sync automatically from the Zendesk backend on a daily cadence, so the knowledge base stays current as technical writers ship updates, with no manual re-upload.
-
5
Answer with citations
Every response links back to the exact source, with the AI's reasoning visible, and when the material can't support an answer, the system says so rather than inventing one.
-
6
Close the feedback loop
A thumbs-down prompts agents for context, then routes the signal by type, bad source data versus a knowledge gap, feeding gap detection, drafting, and an approval workflow before anything re-enters the knowledge base.
-
7
Deliver it where agents already work
A resizable Chrome sidecar brings the knowledge base into any browser tab, an MCP layer exposes the same knowledge to AI agents, and in-context translation serves the multilingual team.
Where native search fell short, governed retrieval held up
Zendesk's native search keyword-matched one article at a time. It struggled with exactly the questions agents brought most often: detailed, multi-article questions where the answer spanned several documents or lived deep inside a long one. Implicit's governed retrieval reads across the full curated set and returns a cited, source-linked answer.
Native knowledge-base search
One article at a time
Keyword-matched a single article at a time, and struggled with detailed questions that spanned multiple articles or lived deep inside one.
Implicit governed retrieval
Across the full curated set
Graph and vector retrieval across the full curated set, with every answer cited and source-linked, reaching high accuracy in the pilot with minimal tuning.
Key capabilities
Walled-garden curation
The AI only ever sees content the team has verified. Knowledge is added deliberately, workspace by workspace, instead of scraping entire sites and hoping for the best.
Cited answers and visible reasoning
Every answer links to its source and shows how it was reached, so agents can validate it in seconds. This was the feature that turned AI from a risk into something the team would actually rely on.
Multilingual delivery
In-context translation serves a team working across five to six languages, so agents get trusted answers in the language they work in.
Governed access and roles
Role-based permissions and per-workspace scoping give the knowledge team control over who can add, approve, and edit, with approval workflows on the roadmap.
The results
Implicit gave the support organization one place to find trusted answers across a messy, mid-merger knowledge landscape, without giving up control of what the AI was allowed to say.
Answer accuracy in pilot
Reached with minimal configuration or tuning, even while the pilot team was temporarily short-staffed.
Sources syncing live
Help-center content syncing from Zendesk and auto-updating daily in the initial workspace, so answers stay current on their own.
Daily sync volume
Trust as the unlock
Citations and visible reasoning cleared the trust barrier that had blocked earlier AI experiments, turning "I can't verify this" into "I can see exactly where this came from."
Built for a moving target
Daily Zendesk sync and workspace isolation let the team keep answers accurate across three legacy products while building toward one unified platform, instead of freezing knowledge in place.
Why a generic AI wasn't enough
Plenty of tools will bolt a chatbot onto a pile of documents. For a support team living through a three-way merger, that was exactly the wrong approach: it meant trusting an AI with unverified, ungoverned, sometimes-stale internal content and no way to check its work. Implicit went the other direction, layering four things that turned scattered knowledge into a source of truth agents were willing to stand behind:
-
Curated, workspace-isolated ingestion kept the AI grounded in verified content instead of scraping everything in sight.
-
GraphRAG made deep, multi-article answers retrievable, even when the answer spanned several documents.
-
Live Zendesk sync kept every answer current as technical writers shipped updates.
-
Citations on every answer made the AI checkable, the difference between AI that sounds confident and AI a support team can actually trust.