IT Folder uses AI to help you document, search, and manage your IT environment. Here's exactly how it works, what data is involved, and the controls you have.
Four commitments that govern how AI works in IT Folder.
IT Folder does not train or fine-tune models on your data — we run no training pipeline. We reach external providers through their commercial APIs, whose terms exclude API data from model training by default. Those defaults have exceptions, such as abuse monitoring and an account owner's own opt-in, which we set out below.
Content is scanned before it is sent to an external AI provider. Detected SSNs, credit card numbers, API keys, and labeled passwords are redacted. If the scan fails, nothing is sent.
For new organizations, AI is on by default so BIP works immediately: Jev question classification (TypeSafe AI) and Google Gemini query generation on IT Folder's key. Documentation assistance, Q&A and OpenAI embeddings stay off until turned on. Someone you authorize to manage settings can turn AI off at any time and can switch individual features on or off. Turning it off makes every later external call refuse, including queued background work — hosted local processing that powers ordinary search keeps working unless you also use the master switch.
AI drafts, answers, and search results are presented to you to use, edit, or discard. Queries the agent generates run in a read-only transaction under a reduced database role, so a generated query cannot change your records — the agent itself still writes its own audit, cache and feedback rows. Background helpers like auto-summaries and file descriptions run on models hosted in our own environment, and the API records provenance for generated fields, including whether someone has edited them since. We do not use AI to make automated decisions about people.
What each feature does, which model runs it, and what data leaves our infrastructure.
Generates summaries, templates, and content suggestions from the inputs and document context you supply. Sends the prompt text to your configured AI provider. Generation does not go looking for your other documents on its own — that retrieval happens in Q&A, described below.
Finds relevant results across your documentation using vector embeddings. By default, embeddings are generated on our own infrastructure (all-MiniLM-L6-v2) and search text is not sent to an external provider. If your administrator opts in to OpenAI embeddings, search queries are sent (PII-redacted) to OpenAI for embedding. When you ask a question (Q&A), the most relevant excerpts of documents you are permitted to see are sent (PII-redacted) to OpenAI to compose the answer — your organization's permissions are applied before anything is retrieved, and results come from your tenant database.
Translates natural-language questions into read-only database queries using Google Gemini 2.5 Flash on an IT Folder-managed key or, if your organization supplies its own Anthropic API key, Anthropic (Claude) under that key. What is sent: your question, your recent questions in the same conversation, database schema metadata (table and column names and types), and — if a generated query fails — the database error message so it can self-correct. These are PII-redacted first. The rows a generated query returns are not included in the request we send to the model, though a database error or an example query can still contain a value. Columns we hold encrypted are kept out of the schema the model sees and are rejected if a query names them; contact and other sensitive columns are masked in results. The agent has a monthly fair-use allowance per account; when it is reached, built-in (non-AI) answers keep working and AI-generated queries resume the next month.
When the agent's built-in answers and its generated query still cannot answer a question, it may make one more Google Gemini call on the IT Folder-managed key to reinterpret the question. What is sent: your question as typed, the question after our vocabulary clean-up, the reason it could not be answered, and your earlier questions in the same conversation with the names of the tables they used — never their results. PII is redacted first. The model returns a plan only (for example, "look up the device named …"); it never writes a query itself. The plan runs through the same read-only, permission-checked path as any other question, and its answer is re-checked by the same deterministic rules before you see it, prefixed "I read this as …" so you can tell. It is capped at two cents per question and has its own monthly allowance per account, separate from the agent's; when that allowance is used up, the agent answers without it. It runs when your organization's AI is on and Google Gemini is one of its allowed providers.
Creates vector representations of your documents for search. By default this runs locally (all-MiniLM-L6-v2) and nothing is sent externally. If your administrator opts in to OpenAI embeddings (text-embedding-3-small or text-embedding-3-large), content is cleaned, PII-scanned, and chunked before it is sent to OpenAI.
Builds image and text embeddings for visual search using CLIP, hosted in our own environment. The "description" is CLIP matching the image against ten fixed candidate phrases (for example "a photo of a server rack in a data center") and keeping the closest — it is not a free-form caption. Image pixels are not sent to an external AI provider. If your administrator opts in to OpenAI embeddings, text that belongs to the image — its filename, alt text and caption — is embedded through OpenAI like any other text.
Creates concise summaries of documentation using DistilBART, which runs locally. Long documents are summarized from their opening portion.
The external providers and local models IT Folder uses. Your administrator controls which external features are active.
| Provider | Used For | Key Type | Data Sent |
|---|---|---|---|
| OpenAI (GPT-5.2, text-embedding-3-small / -large) | Documentation assistance, Q&A answers, opt-in embeddings | Your org's API key | Prompts, questions with relevant document excerpts, and — if OpenAI embeddings are enabled — document chunks and search queries (PII-redacted) |
| Anthropic (Claude — models selectable by your administrator) | Documentation assistance, content generation, and the Text-to-SQL agent (natural-language queries) | Your org's API key | Prompts (PII-redacted). For the Text-to-SQL agent: your question, recent questions in the same conversation, schema metadata, and database error messages on a failed query (PII-redacted). No query results. |
| Google (Gemini 2.5 Flash) | Text-to-SQL agent (natural-language queries), and the agent's one recovery call when it cannot otherwise answer | IT Folder-managed | Your question, recent questions in the same conversation, schema metadata, and database error messages on a failed query; for a recovery call, your question and the names of the tables your earlier questions in the conversation used (PII-redacted). No query results. |
| TypeSafe AI (Jev typed classification API) | Query agent question classification: which kind of record, operation and expiration source a question is about. Always used when your organization's AI is on (on by default); if Jev's monthly allowance is reached or Jev is unavailable, built-in routing is used instead | IT Folder-managed | Your question, your previous question when history is on, and a focused document's title and source table (PII-redacted), with a fixed rubric and the allowed record types. No database rows or document content. |
| SentenceTransformers (all-MiniLM-L6-v2) | Text embeddings for search (default) | Runs locally | Nothing sent externally |
| SentenceTransformers (all-mpnet-base-v2) | Query agent question matching | Runs locally | Nothing sent externally |
| Cross-encoder (ms-marco-MiniLM-L-6-v2) | Query agent intent check | Runs locally | Nothing sent externally |
| CLIP (clip-ViT-B-32) | Image embeddings, image search, and candidate-phrase descriptions | Runs locally | No image pixels sent externally |
| DistilBART (distilbart-cnn-12-6) | Document summarization | Runs locally | Nothing sent externally |
| spaCy + Microsoft Presidio (en_core_web_sm) | PII detection and redaction | Runs locally | Nothing sent externally |
| VADER (rule-based) | Document sentiment scores | Runs locally | Nothing sent externally |
For the full list of sub-processors, see our Sub-Processors page.
Technical controls that protect your data in the AI pipeline.
Presidio-based detection runs on content before it is sent to any external AI provider, including the query agent. One enforced code path handles these calls and refuses to send if scanning fails. Detects and redacts:
Detection is pattern-based and cannot catch every possible secret or personal detail, so avoid pasting credentials into documents.
Each tenant has its own PostgreSQL database, and AI processing is scoped to your tenant.
How your API keys and AI traffic are protected:
The text-to-SQL agent has defense-in-depth protections:
AI activity is logged for compliance and troubleshooting:
For new organizations, AI is on by default: Jev question classification and Gemini query generation. Documentation assistance, Q&A and external embeddings stay off until turned on. Users you authorize to manage organization settings can:
settings.update), not by a fixed "admin" account — anyone you grant that permission can change themOnce content leaves our infrastructure it is governed by the provider's terms, not ours. These are their published defaults. We have not independently verified account-specific protections such as zero retention or exclusion from human review.
We have not obtained account-level evidence for zero-retention approvals, modified abuse monitoring, no-human-review commitments, or a guaranteed processing location. Where a provider's default already covers something, we describe the default rather than promising more. API authentication or a paid plan on its own does not establish zero retention, no human review, or a particular processing region.
Processed by models running in our own environment, not sent to an external AI provider — and not on your computer. See Where Processing Happens below.
Three different places, often confused for each other.
Models we describe as local or hosted — MiniLM embeddings, CLIP image descriptions, DistilBART summaries, VADER sentiment, and the PII scanner — run inside IT Folder's own application and worker environment. That means the input is not sent to an external inference API. It does not mean the processing happens on your computer or inside your network, and the inputs, vectors and outputs can still be stored within IT Folder.
Our application and storage infrastructure targets AWS US West. External providers run their own infrastructure, and where they process a request is governed by their terms and our account configuration. We have not verified the effective location of every production service, log sink and backup, so we do not claim that all processing and storage is US-only.
We record metadata for AI calls — provider, model, status, token counts and duration — using retries and idempotent event identifiers. It is not a guaranteed ledger: a database outage can still leave an incomplete record. An "AI-generated" label tells you where content came from; it is not a statement that the content is correct.
A generated database query runs under a 30-second statement timeout, a ceiling of 1,000 rows and a 2 MB response-size budget. The agent answers a question; it is not a way to export a table, and the rest of our API pages results at 100 rows at a time. A query asking for more rows than that is capped at 1,000 and the result is flagged as possibly truncated. A result whose size exceeds the 2 MB budget returns an error asking you to select fewer fields, rather than a partial answer that looks complete. Counts and grouped summaries are unaffected by the row ceiling. Treat any list that reaches the cap as a sample and narrow the question rather than relying on the cut-off.
What these features cannot promise. Please read this before relying on AI output for an important decision.
Results can be incomplete, source information can be out of date, and a technically valid database query can still answer the wrong question. Check important results against the source records — particularly dates and filters — and review generated procedures and configurations before you use them in production.
Our scanner detects structured data such as email addresses, phone numbers, IP addresses, financial identifiers, SSNs and recognised credential formats. It does not reliably detect names, street addresses, internal hostnames, every password format, or deliberately obfuscated secrets. Avoid pasting credentials into documents.
Disabling a feature, removing a permission or deleting a source controls what happens next. It cannot retrieve an answer already shown to a user, or information already sent to a provider. Content you deliberately saved — a generated document, a summary, a description — is your data and stays until you delete it.
Access is decided by your organization's permissions before anything is retrieved or sent — never by the AI model. Sources that changed, were deleted or are no longer permitted are withheld from answers rather than reused from a cache.
Text that reaches a model — a document, a knowledge-base article, an imported file — can contain instructions aimed at the model rather than at you. We mark retrieved material as untrusted, check that cited sources are ones you were actually allowed to see, and enforce the query and permission limits described above regardless of what a document says. We do not claim immunity from prompt injection, and we have not yet completed an adversarial evaluation against the live models.
The behaviour described on this page is verified against our application source code and its automated tests. Two things that source code cannot establish are stated as open rather than settled: the configuration of our providers' accounts — covered under What Providers Do With Your Data — and the effective region of every production service, log sink and backup, covered under Where Processing Happens. Where we have not verified something, this page says so instead of rounding up.
Read our full privacy policy for legal details, or contact us directly.