goes into searching for answers
56% of employees say some information can only be obtained by asking a colleague. Survey of 12,000 people.
Atlassian, State of Teams 2025ENCLAVE · corporate AI
Give your people AI that knows company documents — inside your perimeter and within their access rights.
ENCLAVE answers employees in plain words from company documents and shows the exact passages of your policies, not the “model’s opinion”. Access rights from source systems are applied before the answer is built. It is a corporate alternative to public AI chatbots that your security team can sign off on — inside your perimeter, air-gapped included.
● on-prem and air-gapped
● SSO OIDC · SAML
● audit log → SIEM
chatanswer
How long do we keep a client’s personal data after the contract ends?
Three years after the contract ends; then the data is deleted with a formal deletion record[1]. Exports from the CRM environment only with sign-off from the personal data processing officer[2].
[1] SharePoint · Personal Data Processing Policy, cl. 5.3[2] Confluence · CRM Handling Procedure
Personal Data Processing Policy.docxavailable to you
Risk Committee Minutes.pdfhidden by access rights
✓ source access rights applied before answeringmodel: local
0unauthorised exposures in 100,000 “employee × document” checks
0.6 smedian response time across 695 questions in 24 hours (test bench, August 2026)
10knowledge sources out of the box, plus a desktop app for Mac and Windows
16model slots — from local to cloud
Interface
This is ENCLAVE at work
Employees ask in their own words — ENCLAVE answers with footnotes to source fragments, shows what was hidden by access rights and honestly says “insufficient data” when there is no basis.
Demo data.
Why now
Your people already ask AI chatbots. The question is which ones — and about what.
Knowledge is scattered across wikis, trackers, drives and chats. Search returns links, colleagues give different versions, and public AI brings leak risks and made-up answers. Banning it doesn’t work: you need a corporate alternative that is easier to use.
mostly use personal AI
Only 20–24% of respondents have corporate access to AI — the rest take work tasks to public services.
RBC School of Management, 8 October 2026of sensitive data being sent to AI
That is the average for a single organisation. And in August 2025, more than 370,000 Grok chats and 4,500 ChatGPT chats ended up in Google search results.
Netskope 2026, ITPro, 12 January 2026How it works
Sources → permission-aware index → answer with footnotes.
01
Connect
Connectors pull permitted content from Confluence, Jira, SharePoint, Git, S3 and network shares into an index inside your perimeter.
02
Permissions
Access rights from source systems come along with the documents. A document with unclear permissions is shown to no one.
03
Search
For each question, ENCLAVE finds the relevant passages and drops everything the employee is not allowed to see — before the answer is built.
04
Answer
The gist in plain words, footnotes [1][2] to exact passages and document cards. No grounds for an answer — an honest “not enough data”.
Capabilities
Not a chatbot for the sake of it — a corporate knowledge layer.
Answers linked to exact passages
Every statement carries a footnote to a specific part of a document: file, system, quote and a link to the original. No precise answer — a list of documents that may help.
Everyone sees only what they’re allowed to
Access rights from source systems are applied before the answer is built. A test of 100,000 “employee × document” combinations: 0 unauthorised exposures and 0 wrongly hidden documents.
Search mode without generation
When you need a document rather than an answer: semantic search returns documents and passages, with no model involved.
Ready documents from an answer
Word, Excel and PowerPoint files built from what was found — a summary, a table, a deck for management.
Any model — your call
16 connection slots: DeepSeek, OpenAI, Claude, Grok, Gemini, local Ollama and more, plus your own model with an OpenAI-compatible API inside your perimeter.
A log that can’t be rewritten
Overwrites are blocked at the storage level; SIEM export in CSV, CEF, syslog and NDJSON. “View as employee” mode for the security team — recorded in the log.
Knowledge map and benchmark questions
A graph of topics and entities shows what the knowledge base covers, and benchmark questions show how answer quality changes over time.
Folders on the employee’s computer
The Mac and Windows app connects an employee’s local folders — with the same access rules and audit.
Air-gapped — verified at build time
Every build checks that there are no outbound connections. An external model is connected only if you choose one.
Who benefits
Employees get answers. Security gets control. IT gets predictable operations.
Employees
Chats and answers with citations, smart passage search, one click to the original source, a desktop app for Mac and Windows.
Administrators
Connectors and sync, the index, models and platform health, answer quality and alerts, licence, updates and single sign-on.
Security and audit
Fail-closed permissions, “view as employee” mode, an immutable log, SIEM export, air-gapped deployment.
Deployment
One install command. Capacity sized to your user count.
- Installation
- A single command,
enclavectl up; the licence comes as a file - Environment
- On-prem, private cloud or an air-gapped environment with no internet access
- Sources out of the box
- Confluence, Jira, SharePoint (Online and Server, NTLM), GitHub, GitLab, S3, WebDAV network shares (including behind your own certificate authority), file upload, websites, custom connectors via SDK; employee folders via the Mac and Windows app
- Sign-in and roles
- SSO via OIDC and SAML (Keycloak, AD FS); roles: employee, administrator, reviewer. In production, single sign-on only
- Audit
- Immutable log; SIEM export — CSV, CEF, syslog, NDJSON
- Capacity (guideline)
- 16–32 CPU cores and 64–128 GB — up to 25 users; 1 GPU with 48 GB — up to 50; 1 GPU with 80–96 GB — up to 200; a cluster of 2–4+ GPUs — over 200. Features are the same; only capacity differs
- On the roadmap
- OneDrive, SMB/NFS, Exchange, Bitrix24, 1C and EDMS (electronic document management systems); sync with AD/LDAP and SCIM
4-week pilot
In a month — a report: what people ask, where the gaps are, who sees what.
WEEK 1
Deploy
Installation inside your perimeter and connection of the first knowledge system.
WEEK 2
Permissions and sign-in
Single sign-on, account mapping and a “what does the employee see” check.
WEEK 3
Real questions
Second and third systems, questions from the pilot group, a review of unanswered questions.
WEEK 4
Report
Demand, quality, knowledge base gaps, a permissions report for the security team — and a decision.
A closed environment adds one preparation week.
Questions
What security and IT teams usually ask
Will an employee see a document they have no access to?
They shouldn’t: access rights from source systems are applied before the answer is built, and a document with unclear permissions is shown to no one. The permission check has been run on 100,000 “employee × document” combinations — zero unauthorised exposures. In the pilot, we separately check “what the employee sees” on your own accounts.
Can it work without internet access?
Yes. ENCLAVE runs in an air-gapped environment with local models, and every build checks that there are no outbound connections. An external model is connected only if you decide so.
What if the answer isn’t in the documents?
ENCLAVE will honestly say “not enough data” and show documents that may help. Such questions are collected into a report — a ready-made list of gaps in your knowledge base.
Do we need a GPU?
For up to 25 users, a CPU server with 16–32 cores and 64–128 GB of memory is enough. More users need a GPU — we will size the configuration to the load.
How accurate are the answers?
Accuracy depends on your documents, so we don’t promise it in words — we measure it in the pilot on your corpus, with benchmark questions and a review of unanswered ones.
Why is this better than public AI or a cloud assistant?
Questions and documents never leave your perimeter, answers rely only on your sources and come with footnotes, permissions are checked before answering, and the log cannot be rewritten.
Does it connect to 1C, EDMS, Bitrix24 or Exchange?
Not yet — they are on the roadmap. For your own system, you can write a connector with the SDK or upload files.
Is it certified by FSTEC (Russia’s information security regulator)?
No. For audit, there is an immutable log, SIEM export and “view as employee” mode, recorded in the log.
ENCLAVE pilot
We’ll show it on your documents, not on a demo bench.
Leave a request — we’ll agree on the first knowledge system, the pilot group and success criteria.
Request sent
We’ll contact you at the email you provided to arrange your ENCLAVE pilot.
Other OpenZTA platforms