Grounded, citable answers
Every answer carries [n] citations that open the source excerpt and page. If the answer isn't in your knowledge base, AMAS says so instead of guessing.
AMAS Enterprise AI Knowledge Governance PlatformEvery answer cited.Every access governed.
Put your enterprise knowledge to work — with AI you can govern. AMAS turns documents, meetings and support history into permission-aware, citable, auditable knowledge, so employees can ask in plain language and IT can run AI under real audit, internal-control and privacy requirements. Source files and indexes stay on your premises; you choose how it's deployed.
Tested on your own documents · on-site, online or at your office
What is the hotel limit for domestic business trips, and do I need approval first?
Hotel limits vary by job grade and city; anything above the limit needs your department head's approval before the trip . Submit receipts within the reimbursement window after you return .
- [1]Domestic Travel Policy.pdfp. 3 · “…amounts above the limit require prior approval…”
- [2]Expense Claim Procedure.docx§4.2 · “…attach original receipts…”
Searched only documents you can accessSource: internal
Logged to audit trail
- [n] citationsOpen the source excerpt and page
- 4 confidentiality levels× permitted roles × departments
- Cloud ⇄ on-premSwitch in 30 s · no silent fallback
- Audit trailWrite-protected at the database layer
01 — Challenges & value
AI is already inside your company. Governance is not.
Connecting AI to a knowledge base is the easy part. Keeping it trustworthy for years — under audits, internal controls and privacy law — is what AMAS is built for.
Today
- 01
Knowledge walks out the door
Files live in inboxes, drives and chat groups; senior staff's judgment lives only in their heads. When people move on, teams start from zero.
- 02
AI answers you can't verify
General-purpose AI sounds confident but can't show its sources. When it's wrong, no one can trace which document or page it came from.
- 03
Shadow AI and a governance gap
Employees already paste company data into outside AI tools — with no access control and no audit trail of who saw or sent what.
With AMAS
Access control built in, every step on record
Retrieval stays within what each user is allowed to see. Four confidentiality levels, departmental isolation and separation of duties — with key access written to the audit log.
Deploy your way
Start hybrid without buying GPU servers up front; switch to fully on-premises inference when requirements tighten. If on-prem inference fails, AMAS never silently falls back to the cloud.
Knowledge that compounds
Good answers are reviewed by people and promoted into the knowledge base, effective on approval. Support cases, meetings and research notes all feed the same governed store.
02 — Architecture
One hub. One set of rules — from data to every application.
Whether a request comes from chat, the support desk or an AI advisor, it passes through the same permissions, confirmations and audit trail. Existing systems connect through APIs and connectors — no rewrite required.
Entry points
- AI assistant
- @AI advisors
- Support center
- Workspace
- In-depth reports
- Sales copilot · Security compliance
- Open API
- 01
Data sources
- Company documents (PDF, Office, images, scans)
- Meeting recordings, transcripts, support cases
- External sources (allow-listed, scheduled refresh)
- 02
Parsing & indexing
- Text, tables and OCR
- Visual understanding of charts and diagramsCloud by default · on-prem option
- Semantic + keyword index; knowledge graph with page & quote
- Completeness checks & human review gates
- 03
Understand the question
- Work out intent and where to look
- Colloquial terms, abbreviations, mixed languages
- Confirm before any external search
- 04
Permission-scoped retrieval
- Search only what the asker may see
- 4 levels × roles × departments
- Rerank candidate passages
- 05
Generate with citations
Question + permission-filtered passages
- Inference: one cloud provider you choose, or your own GPU
- Verify the answer; add [n] citations and source type
- Say so when evidence is missing
- 06
Audit
- Q&A, system actions and security events logged
- Write-protected at the database layer
- Admins can't view their own records
Governance across every stage
- Identity: local, AD/LDAP, Entra ID
- Roles & capabilities
- Review · versioning · recycle bin
- Sanitization of external input
- Confirmation before writes & external queries
03 — Core modules
Eight modules on one governed knowledge base.
Start with what you need and enable more as you grow. Every module inherits the same permissions, citations and audit trail.
Document intelligence
Understands the documents others can't — scans, flowcharts, merged-cell tables and slide charts — and turns them into citable knowledge.
- PDF, Office and images; automatic OCR for scans; batch upload with duplicate and new-version detection.
- Steps and branches from flowcharts, values from charts, hierarchy from org charts, title blocks and tolerances from drawings.
- Gaps are flagged, OCR-inferred text is labeled, and flowcharts need human approval before reaching Q&A.
Cited AI assistant
Understands the question, decides where to look, searches only what you're allowed to see — and answers with sources.
- Six modes, auto-detected or chosen: general chat, internal knowledge, enterprise systems, web search, internal-vs-external comparison, in-depth reports.
- Finds colloquial terms, abbreviations and mixed Chinese/English; model numbers and document codes get exact-match priority; custom glossaries.
- Web search is limited to allow-listed sites and confirmed first; answers label their source type — internal, external or general AI knowledge.
Enterprise systems, web search, comparison and in-depth reports are part of the Enterprise plan.
Knowledge governance & access
Clear rules for who can see, who approves, and how to recover.
- Four confidentiality levels × permitted roles × department scope. Users without access can't even tell a document exists.
- Separation of duties: IT admins see only structural metadata of content not opened to them; the owning department decides.
- Review before publishing, versioning, 30-day recycle bin; 20 built-in roles and 27 capabilities, customizable with live preview.
@AI advisors
Summon sales, support, HR or finance advisors with an @ mention — without widening anyone's permissions.
- Administrators define each advisor's persona, knowledge scope and reasoning depth.
- Picks the right skill per question — knowledge search, enterprise-system lookup, data analysis (calculations, statistics, charts) or reports.
- Uses only data the requester may see; sensitive fields such as salary stay masked by role.
Support center & flywheel
AI researches and drafts; agents finalize and send. Approved answers become knowledge the very next case can use.
- AI never replies to customers on its own. Drafts show whether they're grounded, carry citations, and switch tone.
- Simplified Chinese, English, Japanese and Korean messages are translated; replies go back in the customer’s language with a back-translation.
- Case views, kanban and analytics, Excel import of legacy cases; a dashboard of top-cited and never-cited knowledge.
Email / LINE / WeCom automated intake: built, pending on-site acceptance.
Workspace: meetings, research, reminders
Meetings become summaries and action items; research becomes a shareable notebook; overdue work surfaces on its own.
- Create meeting records from transcripts, text, audio, public URLs or live recording; summaries and action items trace back to the original lines.
- Research notes combine personal files, company knowledge and web clippings, with mind maps and knowledge graphs; personal libraries stay isolated.
- Daily checks for overdue and stalled items; say “remind me to submit the report next Wednesday” to add a task.
Sales copilot & external intelligence
AI organizes, people confirm — meetings become action items, and external intelligence stays current.
- Four kinds of items extracted from transcripts and written only after confirmation; opportunity board and pipeline views.
- Seven external sources refresh on schedule: patents, regulations, grants, research, market intel, tenders and CVEs.
- Optional one-way opportunity sync to Salesforce (off by default).
Regulation and tender sources currently focus on Taiwan public data.
Security compliance & operations
Day-to-day ISO 27001 compliance, vulnerability intelligence and system operations — in your IT team's hands.
- ISO 27001 recurring forms scheduled with an ICS calendar feed; daily CVE updates, alerting only on categories you use.
- Software bill of materials exportable as CycloneDX; continuous answer-quality monitoring.
- One-click upgrades with backup and rollback; remote support needs your authorization, is time-limited and cannot see documents or conversations.
One-click upgrades require the vendor upgrade channel to be enabled.
04 — Knowledge flywheel
Every well-answered case makes the next one easier.
All eight stages are live. AI researches and drafts; people decide what gets sent and what becomes knowledge.
- AI
- People
- System
- People + system
- 01
AI drafts a reply AI
Researched from the support knowledge base, with citations.
- 02
Agent finalizes People
Edits before sending; AI never replies to customers directly.
- 03
Capture on close People + system
The question and final answer become a knowledge candidate.
- 04
Manager review People
Similar entries side by side; approve, edit or return.
- 05
Live on approval System
Searchable for the very next question.
- 06
Reused automatically AI
Cited in future drafts, so suggestions keep improving.
- 07
Dashboard tracking System
Most-cited and never-cited knowledge.
- 08
Continuous revision People
Update and resubmit; the current version stays in use during review.
Status of related capabilities
- Pending on-site acceptanceAutomated intake from Email, LINE and WeCom (built; today cases are created manually or imported from Excel)
- Built · off by defaultAnswer templates by category
- PlannedCase knowledge graph and cross-case FAQ synthesis
05 — Deployment
Keep your knowledge in-house. Borrow the AI compute you need.
Two deployment models — choose by sensitivity. We are precise about what crosses the boundary, because your auditors will be.
On-premisesYour network
- Source filesStored on your network; downloads always permission-checked
- Parsing & indexText, tables, OCR, semantic + keyword index
- Permission filterRetrieval only within what the asker may see
- Audit trailWrite-protected at the database layer
Data boundary
Question + permission-filtered passages
Generated answer
Files & index never leave
Cloud or on-prem GPUAI inference
- HybridCloud LLM from the one provider you choose
- Fully on-premYour own GPU server — nothing sent out for inference
- Switch in 30 sNo restart. No silent fallback to the cloud.
| Item | HybridRecommended start | Fully on-premisesHigh sensitivity |
|---|---|---|
| Source files, index, access control | On-prem | On-prem |
| AI inference | Cloud LLM from the provider you choose | Your own GPU server |
| What goes to the cloud | The question + permission-filtered passages | No cloud inference |
| Chart & diagram vision | Cloud by default, switchable to on-prem | On-prem model; files never leave |
| Up-front investment | No up-front GPU purchase; pay per use | Requires an on-prem GPU server (e.g., NVIDIA DGX Spark) |
| Best for | Teams starting small to prove value | Organizations that must keep data inside their network |
Guaranteed in both models
- Switch within 30 seconds
Admins switch between cloud and on-prem in settings — no restart, no change for users.
- You choose the provider
AMAS uses only that one and never switches vendors on its own.
- No silent cloud fallback
If on-prem inference is down, AMAS shows an error instead.
- Cloud spend you can see
Usage and estimated cost by department and user, exportable to CSV.
- Permission-checked downloads
Downloads always pass a server-side check — no public or shareable links.
- Allow-listed lookups only
External lookups are limited to allow-listed sites.
Runs on a single server — pilot with one department, then scale company-wide; enable modules as you need them. External knowledge updates and online upgrades need outbound HTTPS 443 only.
06 — Trust & governance
Trust you can verify — users check the answers, auditors check the system.
Controls are built into the product, not bolted on: who can see what, who approves, and what gets recorded.
Permissions follow the person
Every answer is filtered by the asker's access; citations, advisors and deep research never bypass it.
4 levels × roles × departments
Public, internal, confidential, restricted — combined with permitted roles and department scope.
Existence hiding
Users without access get the same response whether or not a document exists.
Separation of duties
IT admins see only structural metadata of content not shared with them; the owning department decides.
Three-tier administration
Super-admin is CLI-only with two-person presence; break-glass access is time-limited and logged.
Audit trail
Five record types; write-protected at the database layer; retention templates (5 years general, 7 years finance/healthcare).
Admins can't view their own audit records
No self-oversight.
Humans at the critical points
Confirmation before writes and external queries, review before documents go live, no automatic replies to customers.
Remote support on your terms
Authorized by you, time-limited, revocable, with no view of documents or conversations.
- Layered protection for sensitive fields: salary shown as actual values, a relative index, or denied, by role
- External input such as inbound support messages is sanitized before it reaches the AI
- SBOM (CycloneDX) with CVE matching, plus answer-quality monitoring
- Secure SSO defaults: LDAPS enforced, at least one local admin always retained, admin roles never granted via AD groups
07 — Integrations
Connects to what you run — with the status stated plainly.
We separate what ships today from what is built, designed or still a vision — so you can plan with confidence.
Available now
- AD/LDAP, Microsoft Entra IDSingle sign-on; groups map to roles and refresh at every login.
- ERPRole sync and instant deactivation on offboarding; data queries integrated case by case.
- HR / payrollLeave balance, payslips and expense status in chat; actions like leave requests need confirmation.
- SalesforceOne-way opportunity sync (off by default).
- Open APIOther systems retrieve answers and citations; credentials, source-IP allow-list, every call logged.
- Calendar & tasksCreate events and tasks from chat; ISO calendar subscribable in Outlook/Google.
Built · pending acceptance
- Email, LINE Official Account, WeComAutomated support intake with agent-mediated replies.
Planned (designed)
- Standard MCP compatibilityAMAS as an MCP server, so external AI tools can query enterprise knowledge within user permissions.
- Cloud document platformsRead-only access to Google Drive, SharePoint and Confluence via MCP.
Vision
- Multi-agent workflowsRetrieval, analysis, writing and review chained into workflows — with human checkpoints and audit at every step.
- Two-way ERP/CRM, Teams notificationsRequires vendor cooperation and use-case assessment.
Integration scope is assessed against your existing systems.
08 — Use cases
Where AMAS fits.
Typical scenarios built on live capabilities. These are illustrative use cases, not customer case studies.
Manufacturing: product & equipment support
Query spec sheets, engineering drawings and service manuals; turn good answers into team-wide FAQs; translate inbound messages automatically.
Public sector & regulated organizations
Compare internal rules with current regulations, set audit retention to statutory periods, run fully on-prem.
R&D-intensive companies: research & patents
Research notes combine internal and external sources, with auto-updated patent and paper libraries plus mind maps and knowledge graphs.
IT & security teams: day-to-day ISO 27001
Scheduled recurring forms, CVE alerts filtered to what you run, SBOM for auditors.
HR & administration: employee self-service
Cited answers on leave, travel and expense policies; check leave balance and payslips in chat.
B2B sales teams: from meetings to pipeline
Action items from meetings, tender and grant intelligence, opportunity board and pipeline views.
09 — Roadmap
Adopt agentic AI one governed step at a time.
Let AI prove itself where people review its work — then expand.
- Stage 1Live
Enterprise knowledge brain
Cited Q&A, permission filtering, internal-vs-external comparison, in-depth reports, knowledge governance.
- Stage 2Live
Data-analysis copilot + conversational actions
Optional fully on-prem inference; advisor data-analysis skill (calculations, statistics, charts); query and act on enterprise systems with confirmation.
Enterprise-system data queries are integrated case by case.
- Stage 3PlannedVision
Open & collaborative
Two-way MCP compatibility and read-only document platforms, then multi-agent workflows with human checkpoints and audit at every step.
MCP: planned (designed). Multi-agent: vision (no timeline).
Does our data go to the cloud? Can it run fully on-premises?
Source files and indexes stay on your network, and retrieval and permission filtering run on-premises. In cloud mode, only the question and the permission-filtered passages are sent to the AI provider. Chart and diagram pages use cloud vision by default and can be switched to on-prem. You can switch to fully on-premises inference at any time, and AMAS never falls back to the cloud if on-prem fails. External knowledge updates and online upgrades need outbound HTTPS 443 only.
Could colleagues see documents they shouldn't through the AI?
No. Every answer is filtered by the asker's permissions first, and citations, AI advisors and deep research never bypass them. Without access, users get the same "no information found" response whether or not a document exists. Even system administrators can't see restricted content that hasn't been shared with them.
Will the AI make things up?
Every answer carries [n] citations you can click through to the source, and labels which kind of source it used. When evidence is insufficient, AMAS declines or marks the answer as ungrounded. Answer quality is monitored continuously, including the share of answers given without supporting sources.
Can we choose which AI model to use?
Yes. In cloud mode you choose the provider — for example Anthropic Claude, OpenAI GPT, or another OpenAI-compatible service — and AMAS uses only that one. You can also run models on your own GPU server. Admins switch in settings, effective within 30 seconds.
Can it integrate with our existing AD, ERP, CRM and HR systems?
AD and Entra ID are configured in the web console. ERP role sync and offboarding are live, with data queries integrated case by case. Opportunities can sync one-way to Salesforce, HR systems support leave, payslip and expense lookups, and an open API lets other systems call AMAS. Standard MCP support is in design, extending to platforms such as Google Drive and SharePoint.
Will the AI reply directly to our customers?
No. AI only drafts — your agents always finalize and send. Once a manager approves a good answer, it's searchable for the very next question; when existing knowledge is revised, the current version stays in use until the update is approved.
Next step
See it on your documents.
Book a working session on-site, online or at your office. We will show answers, citations and permissions on material you choose.