Organizational knowledge, with the boundaries stated
AI over your organization's knowledge, with the boundaries spelled out
Answering from your own documents, tickets and code is a retrieval and permissions problem before it is a model problem. This page separates what ZeroTwo describes today, what depends on your setup, and what is reference design, then gives you a pilot checklist to test retrieval, no-answer behavior and access yourself.
- 5 itemsDescribed on our public pagesEach item links to where it is stated.
- 5 itemsDepends on your setupVerify these on your own data.
- 6 itemsReference designHow any governed system is built, not a feature list.
Not shown on this page: a retrieval trace, an access-denial result, benchmark scores, connector sync intervals or an on-premise option. Where a claim would need one of those, it is listed as something to verify.
What ZeroTwo describes today
These are statements from our own public pages, not results from a test on your data. Follow each link to the page that states it.
Connect your apps
Connect tools such as Google Drive, SharePoint, Microsoft Teams, Notion and GitHub. The connectors page says ZeroTwo can only see what you can see in the connected app, and that you are shown the access you are granting before you approve it.
- Google Drive
- SharePoint
- Microsoft Teams
- Notion
- GitHub
Read common file types
Connectors retrieve text from common file types: TXT, PDF, CSV, XLSX, PPTX and DOCX. Free allows 1 file upload per message; Plus and above allow multiple uploads per task.
Plan comparisonWrite answers with 60+ models
The generation step can run on any of 60+ models, so you can compare how different models handle the same material.
Model listMemory and projects by plan
Free has limited projects and memories. Pro has expanded memory and unlimited projects. Ultra has maximum memory.
Plan comparisonShared workspaces for teams
ZeroTwo Business is $39.99 per seat per month with pooled team credits and shared workspaces. The connectors page says admins on Business workspaces decide which connectors are turned on.
ZeroTwo Business
What depends on your setup
Five questions that decide whether a knowledge assistant is safe to roll out. We do not publish an answer to these for your data, so each one points to a test or a document.
- Document-level permissions
Can an answer ever draw on a document the person asking cannot open in the source app? ZeroTwo says it sees only what you can see, but no per-document access-denial result is published here.
Run test 4 in the pilot checklist with two accounts.
- Freshness
How soon does an edited or deleted document change the answers? Sync timing is not described on this page and varies by source.
Run tests 5 and 7: edit or remove a document, then ask again.
- Whether your system is covered
If a source is not in the connector list, its content has to be exported or uploaded by hand.
- Retention and training
What is kept, for how long, and whether it is used to improve models can differ by plan and by setting.
Read the privacy policy and get written answers for your plan.
- Where generation runs
If answers must be generated inside your network or region, that is a deployment requirement. This page makes no on-premise or data-residency claim.
Review the enterprise requirements, then ask sales in writing.
A seven-test pilot you can run on your own documents
Use a small folder you understand well, two accounts with different access, and a notes file recording each question, the date, the model, the answer and its citations. The categories follow what enterprise retrieval benchmarks such as EnterpriseRAG-Bench say matters: single and multi-document questions, conflicting information, and recognizing when information is absent.
| Test | Setup | Pass | Fail |
|---|---|---|---|
| 1. Single-document answer | Ask a question whose answer sits in one document you can open. | The answer matches the document, cites it, and the cited passage opens. | No citation, or the citation does not contain the claim. |
| 2. Two-document answer | Ask something that needs facts from two different documents. | Both documents are cited and each claim is attributed to the right one. | One source is missing or claims are blended without attribution. |
| 3. No-answer case | Ask about something that is not in the corpus at all. | It says it cannot find the answer in your documents. | It produces a confident answer anyway. |
| 4. Denied access | Using an account without access to a document, ask a question only that document answers. | Nothing from the document appears, not even its title. | Any fact, title or hint of the restricted document leaks. |
| 5. Stale source | Edit a document, wait the sync time you expect, then ask again. | The answer reflects the edit, or states how old its source is. | The old answer is returned with no sign it may be out of date. |
| 6. Conflicting versions | Keep two documents that disagree and ask about the point. | It surfaces the conflict or says which source is newer. | It picks one silently. |
| 7. Deleted source | Remove or revoke a document, then ask about its content. | It is no longer cited or used. | It still appears in answers. |
Benchmark source: arXiv:2605.05253, abstract read 5 October 2026. The checklist is our own and is not part of that paper.
Six layers every governed system needs
This is a reference design for any retrieval system, ours or anyone else's. It is not a statement about which layers ZeroTwo implements; use the Ask prompts when you evaluate a vendor, including us. For deployment and governance requirements at larger scale, review enterprise deployment and governance requirements.
Sources
Wikis, documents, tickets, chat, email, CRM records, code and call transcripts. Real company data is messy, duplicated and partly out of date.
AskWhich of our sources can it read, and with whose permissions?
Ingestion
Parsing, OCR, table extraction, chunking, deduplication and embedding, keeping source metadata so answers can cite where they came from.
AskWhat is stored, where, and for how long?
Index
Keyword search for exact terms, vector search for meaning, and sometimes a knowledge graph for relations between people, products and contracts. Each catches queries the others miss.
AskWhich of these does it use, and can I see which one found a result?
Retrieval
Filtering by the asker's permissions, ranking with recency, and checking that what was retrieved can actually answer the question before generating.
AskWhat happens when nothing relevant is found?
Generation
A model writes the answer from the retrieved passages and cites them. The model changes style and reasoning, not what the index contains.
AskCan I compare models on the same retrieved passages?
Governance
Access control inherited from source systems, an audit trail, redaction of sensitive fields and a defined data boundary.
AskWhat is logged, who can read the log, and where does generation run?
Six questions worth grounding in your own data
Use these as pilot questions. Each lists what a good answer shows, not a result we obtained. For a single file rather than a standing corpus, start with AI document analysis.
Employee policy
“What is our parental-leave policy in Germany?”
A good answer: Cites the current policy for that country, names its effective date, and says so when no policy is found.
Sales context
“What did we last tell this customer about pricing, and how did they reply?”
A good answer: Points to the specific email or CRM note, using only records the asker may see.
Engineering onboarding
“How does our auth service handle single sign-on?”
A good answer: Cites the design doc and the code, and flags where they disagree.
Support history
“Summarize this customer's history with us.”
A good answer: Lists ticket references and dates without mixing in a different customer.
Audit and compliance
“Where did we document the change-management control mapping?”
A good answer: Names the document and section, or states that none was found.
Research and strategy
“Summarize our last three competitive analyses and where they disagree.”
A good answer: Attributes every claim to its source document.
What it costs to run the pilot
Paid plans include a monthly credit pool and Auto-request allowance, not unlimited use, and Pro 2x ($59.98/mo) doubles Pro's allowances rather than removing them. Prices come from our pricing catalog; the pricing page has the full comparison.
Free
$0/mo
- 100 free credits to start
- +20 bonus credits every day you log in
- Limited image generations
Pro
$29.99/mo
$26.99/mo billed annually
- 3,500 monthly credits
- 2,500 free Auto requests / month
- Expanded memory
Business
$39.99/seat/mo
- 4,000 monthly credits per seat (pooled)
- Pooled team credits & shared workspaces
- 1,250 free Auto requests / seat / month (pooled)
Frequently asked questions
What is organizational knowledge AI?
It is the practice of grounding a language model's answers in your own documents, tickets, chats and code instead of its general training data. The usual technique is retrieval-augmented generation (RAG): find relevant passages first, then have the model answer from them and cite them. How well it works depends mostly on retrieval quality and permissions, not on which model writes the answer.
Is RAG the same as a knowledge graph?
No. RAG retrieves text passages and gives them to the model as context, which suits unstructured documents. A knowledge graph stores entities and the relations between them, which suits questions about connections, such as every customer in an industry who bought a given product. Some systems combine both with keyword search. Whether that helps depends on your data, so test it on your own corpus rather than assuming.
How is company data different from public web data?
It is messier: duplicates, near-duplicates, misfiled documents, conflicting versions and partial information, spread across chat, email, tickets and drives. A May 2026 paper, EnterpriseRAG-Bench (arXiv:2605.05253, read 5 October 2026), was written because existing benchmarks focus on web or other public sources. It describes about 500,000 synthetic documents across nine source types and 500 questions in ten categories, including conflict resolution and recognizing when information is absent.
What sources can ZeroTwo connect to?
The connectors page lists the current apps, including Google Drive, SharePoint, Microsoft Teams, Notion and GitHub, and says connectors retrieve text from common file types such as TXT, PDF, CSV, XLSX, PPTX and DOCX. If your system is not listed, export or upload the content. Check the connectors page for the current list before planning around a specific source.
How do you keep it from making things up?
No system can promise it never will. What you can ask of any retrieval system is that answers cite their sources, that it declines when the corpus does not cover the question, and that high-stakes answers get human review. Tests 1 to 3 in the pilot checklist measure exactly that on your own documents.
How do access controls work?
The connectors page says ZeroTwo can only see what you can see in a connected app, that you review the access you grant, and that admins on Business workspaces decide which connectors are turned on. Whether restricted documents ever surface in someone else's answers is something to verify, which is what the denied-access test is for. For SSO, data-processing or residency requirements, ask in writing through contact sales.
What about freshness?
How quickly a changed document shows up depends on the source and how it is connected, and this page does not state a sync interval. Edit a document, ask the same question again, and see whether the answer changes or says how old its source is. Tests 5 and 7 cover edits and deletions.
What does it cost?
Free is $0 with 100 free credits to start and limited model selection. Pro is $29.99 per month ($26.99 per month billed annually) with 3,500 monthly credits and 2,500 free auto requests / month. Pro 2x is $59.98 per month and doubles those allowances rather than removing them. Business is $39.99 per seat per month with pooled credits. Allowances are monthly pools, not unlimited use; the pricing page has the current comparison.
Related ZeroTwo workflows
- Enterprise AI platform
Deployment, governance and procurement questions for larger rollouts.
- AI document analysis
One-off review of a single file or a small set, without a standing corpus.
- AI business plan generator
Plan generation grounded in your real strategy documents.
- Market research tool
External signals alongside your internal knowledge.
- AI for sales prospecting
CRM-context prospecting with AI grounding.
- Contact sales
Get security and deployment answers in writing.