Ask your company anything. Every answer shows its source.
Grounded answers over your SharePoint, Drive, wikis and tickets, delivered on whichever assistant your people already use: Microsoft 365 Copilot, Claude, OpenAI or a fully self-hosted stack. Permissions respected, citations mandatory, EU residency available on every path.
The 1.8 hours figure is a McKinsey estimate for knowledge workers. The other three are design rules of every rollout on this page, whichever stack carries it.
Your team keeps re-discovering what the company already knows.
The knowledge exists: buried in SharePoint, scattered across Drive, locked in the heads of people who left six months ago. What is missing is a sanctioned way to ask.
Hours lost to searching
McKinsey estimates employees spend 1.8 hours daily hunting for information. Across a 200-person company that is 360 hours a week, burned on CTRL+F across scattered PDF drives.
New hires lost for weeks
Onboarding takes 2-3x longer because institutional knowledge lives in people, not systems. When someone leaves, their context walks out the door.
Public AI is a compliance risk
Pasting contracts into consumer chatbots violates GDPR, NIS2 and your NDAs. Your legal team knows it. Your employees do it anyway, because the sanctioned way to ask does not exist yet.
One question, four ways to deliver it.
The stack is a delivery detail; the discipline is the product. Switch paths below and watch the same question land on Copilot, Claude, OpenAI and a self-hosted deployment, with the citation that makes the answer defensible.
Illustrative conversation over a fictional document set. Real rollouts ground in your corpus, behind your permissions, and get judged on your golden questions.
The deflection rule
“You don't need a bigger support team. You need to automate the work that team is doing.”
The question that eats your week.
Knowledge assistants earn their keep one recurring question at a time. These are the four we see most, and the pilot usually starts with whichever one made you wince.
Third day of the new hire, fourth Teams ping before lunch.
Policy answer with the source page, in the channel they asked in. HR reads the escalations, and only the escalations.
A renewal lands and nobody remembers which agreements deviate from the template.
Clause-level answers across the whole contract folder, each one citing document and page.
The auditor wants the procedure as it stood in March, and the wiki only shows "current".
Versioned sources mean the answer cites the right revision, and audit prep stops being archaeology.
The correct answer exists in three closed tickets and one wiki page nobody reads.
Grounded on tickets and docs together, so the next reply is consistent with the last one.
We built this for ourselves first.
The retrieval discipline on this page comes from building DECKLOG, our Microsoft 365 knowledge operations product, and from the unglamorous work of making its answers hold up.
Prototyped honestly
DECKLOG's retrieval core started as a zero-dependency local prototype on Ollama: no vendor accounts, no cloud bill, just proof that grounded answers over our own docs were worth building.
Then measured
The prototype graduated to hybrid retrieval (BM25 + vectors + reranking) on pgvector, judged by a golden-question regression suite that runs like a test suite. Quality is a number here, and it has to keep passing.
Now a product and a service
DECKLOG ships as the self-hosted path of this solution, and the same evaluation harness travels to every Copilot, Claude and OpenAI rollout we do.
What searching costs you today.
Slide your own numbers in. The waste is already on your payroll; the question is whether it stays there.
Your numbers
McKinsey estimates knowledge workers lose 1.8 hours a day to finding information; 20 minutes is a deliberately conservative slider default.
Monthly cost of searching
€58,667
1,467 working hours a month, spent re-finding what the company already knows.
Per year
€704,000
Engagements are scoped to your estate, so the price comes after the free scan, and the scan already tells you where the hours are going.
Run the free AI readiness scanAudit, roll out, then keep it true.
Fixed fee or retainer, agreed before work starts. No hourly billing at any tier.
Grounding Readiness Audit
Two weeks. You learn what your knowledge estate can support.
- Source inventory: SharePoint, Drive, wikis, tickets, file shares
- Permissions and oversharing scan before any AI touches the corpus
- Stack recommendation: Copilot, Claude, OpenAI or self-hosted, with reasons
- Golden-question set seeded from questions your teams actually ask
- Written report + 45-minute walkthrough
Outcome
A defensible answer to "which path, on what data, at what risk".
Assisted Knowledge Rollout
The chosen path, wired, evaluated and piloted with a real team.
- Copilot agents + Graph connectors, or RAG on Claude / Azure OpenAI, or DECKLOG
- Permission-trimmed index: answers respect who is asking
- Citations mandatory: every answer shows its source or declines
- Golden-question evals wired into the release process
- Pilot group rollout with adoption measurement and runbooks
Outcome
A grounded assistant a named team uses weekly, with quality you can prove.
Managed Knowledge Layer
We keep the answers true as the documents change.
- Index health: new sources, re-syncs, permission drift checks
- Monthly eval runs with quality and usage reporting
- Model and stack swaps as the market moves
- New departments onboarded onto the same discipline
- Single Slack channel to the engineer who built it
Outcome
The knowledge layer stays current, cited and used, quarter after quarter.
Honest answers to the questions buyers actually ask.
We already pay for Microsoft 365 Copilot. Why not just switch it on?
+
Because Copilot answers with whatever the asking user can technically access, and most tenants over-share far more than anyone realises. Our rule from the Copilot rollout service applies here verbatim: we will not deploy Copilot into a tenant with known oversharing. The readiness audit finds it first; then Copilot becomes a great path.
Do we have to buy DECKLOG for this?
+
No. DECKLOG is one of four paths, and it wins when you need a self-hosted or air-gapped deployment with zero third-party data plane. If Copilot, Claude or Azure OpenAI fits your estate better, that is what we recommend, and the audit says so in writing.
What document types and sources can be covered?
+
PDF, DOCX, PPTX, XLSX, Markdown, HTML, plus SharePoint, Google Drive, Confluence and ticketing systems through connectors. Custom sources (ERP exports, proprietary formats) are handled in the rollout when they carry answers people actually need.
Where does our data live?
+
It depends on the path, and it is decided before the build: in-tenant for Copilot, EU endpoints for Azure OpenAI, EU routing via Bedrock for Claude, or fully self-hosted for DECKLOG (Hetzner EU or your own infrastructure, air-gapped if required). Your documents never train public models on any path.
How accurate are the answers?
+
Every answer cites its source, document and page, or says it cannot find one, rather than improvising. Quality is measured against a golden-question suite built from your real questions, and the suite has to pass before rollout and after every significant change.
Who actually does the work?
+
One senior engineer, the same one you meet on the discovery call. No account managers, no offshore hand-off, no junior rotation. You get a single Slack channel and a direct line to the person tuning your retrieval.
Built from three standing services.
DECKLOG Implementation
The productized private-RAG route: hybrid retrieval, citations built in, self-hosted option.
View serviceAI Readiness & Copilot Rollout
The Copilot path done safely: oversharing scan, Purview labels, wave rollout, adoption.
View serviceCustom AI Integration
Claude and OpenAI builds with the eval harness, abstraction layer and observability.
View service