CorpusIQ

Forty or more connectors reach commerce, finance, advertising, analytics, CRM, email marketing, files, calendars and databases, from Shopify, QuickBooks and Stripe to PostgreSQL, SQL Server and MongoDB, and one Google sign in covers Gmail, Drive, Sheets, GA4, Google Ads and Search Console. Every connector tool is annotated as retrieval only and writes nothing back. Google offers its Ads permission only in a read and write form, and CorpusIQ states that its Ads tools only report on campaigns, keywords and spend.

Agents can read across the whole estate. They cannot send, edit, pay or update anything in it.

Sourcecorpusiq.ioread 2026-10-08

Each question runs through a fixed sequence. CorpusIQ picks one skill, reads the connectors that skill needs at the same time, matches the same customer or product across systems, validates the numbers and hands the result to the assistant. A question about an overdue invoice reads QuickBooks, Gmail and HubSpot in one pass.

The customer does not design, branch, schedule or chain these steps, and there is no workflow builder or runtime for several agents working together.

Sourcecorpusiq.io/securityread 2026-10-08

There are two retrieval paths. The direct path fetches live records for each question and builds no index. Optional indexed search keeps embeddings, chunk IDs and minimal metadata in a separate namespace for each user until the connector is revoked or the account deleted, and the deep search endpoint queries that archive across files and email with a time range.

Above both, a metrics registry and a source of truth engine decide which system answers each KPI, and entity resolution matches the same customer, product or company across systems.

Sourcecorpusiq.io/securityread 2026-10-08

The guardrails act on what reaches the assistant. A validation engine checks every number before the AI sees it, and when a figure cannot be verified the answer says so instead of filling the gap. Anti drift protection stops an assistant from redefining a KPI, inventing a metric or changing a saved rule.

Writes are confined to CorpusIQ's own settings, and no approval step, policy screen or spending limit is documented for those control plane writes.

Sourcecorpusiq.ioread 2026-10-08

Admins decide what the AI can reach: each connected tool is set to read only or off, owners add contributors on multi user plans, and the Enterprise plan adds single sign on. DEKRA issued CorpusIQ a CASA Tier 2 certificate after an OWASP Top 10 assessment (certificate 151c1b05, valid through 5 March 2027), and the certificate is published as a PDF.

SOC 2 is described as aligned, with a readiness program and quarterly control checks, rather than as an attestation. The page also lists AES 256 encryption at rest with keys rotated every 90 days, TLS 1.3, an annual independent pen test and credentials held in a dedicated vault.

Sourcecorpusiq.io/securityread 2026-10-08

Every answer names the records it came from, down to the Shopify order range or the QuickBooks entry, and shows the rule used when a number was calculated, so a reader can trace a single figure to its source.

Behind the answers, each tool call is logged with the raw query text, the tool parameters and a bounded summary of the result, kept for 30 days in an Azure Log Analytics workspace on an immutable audit stream. No customer view, export or SIEM feed for that log is documented.

Sourcecorpusiq.io/securityread 2026-10-08

A customer's business rules carry over between sessions and between assistants: declared facts, decisions, metric specifications and source manifests, saved once through separately named control plane tools and used by every AI the customer connects. Revenue counted as Shopify orders minus refunds comes out the same in ChatGPT, Claude and Perplexity.

People and their assistants maintain this store. Nothing describes a memory the agent builds for itself with a stated scope or lifetime, and the API carries follow up turns by passing back a conversation ID.

Sourcecorpusiq.io/securityread 2026-10-08

The standard product is a managed service on Microsoft Azure, and the subprocessor table places hosting, model inference and analytics in the USA with no region choice.

The enterprise page describes a reasoning layer that runs inside the customer's own Azure tenant. The pricing comparison answers no to self hosting, and no deployment guide or tenant option is published beyond a request for a technical walkthrough and deployment plan.

Sourcecorpusiq.io/enterpriseread 2026-10-08

A browsable catalogue holds more than 150 skills, each a packaged job a customer runs by asking a plain question. Ad Spend Truth Report reconciles ad platform claims against GA4 and Shopify revenue. Cash Recovery Engine ages receivables in QuickBooks, drafts collection emails for each bucket and estimates what can be recovered this month.

Churn Prevention, CRM Pipeline Health, Customer Health Scorecard, Executive Snapshot, SEO Audit and AI SEO each do a job of their own, and every skill page names the connectors it reads.

Sourcecorpusiq.io/llms.txtread 2026-10-08

Questions reach CorpusIQ through channels it runs itself: the ChatGPT app, Claude and Perplexity over MCP, its own Slack and Microsoft Teams apps, where people ask in any channel or direct message, and the REST API for programs.

Every one of them waits for a person or an agent to ask. The security notes rule out background data extraction and limit the product to actions a user invokes, and no schedule, webhook or event that starts work on its own is described.

Sourcecorpusiq.io/securityread 2026-10-08

The answer depends on the surface. CorpusIQ's own inference, including the embeddings behind indexed search, runs on OpenAI, the only model provider in its subprocessor table. In ChatGPT, Claude and Perplexity the assistant belongs to the customer and calls CorpusIQ as a tool.

In the Microsoft Teams app the customer picks the model that writes the answer: ChatGPT, Claude or Perplexity, or their own key for Gemini, Grok or a local or self hosted model. The pricing page and llms.txt still list Teams as coming soon, while the Teams page and the MCP hub say it is live.

Sourcecorpusiq.io/teamsread 2026-10-08

A public REST API at api.corpusiq.io/v1 uses bearer token auth and ships an OpenAPI 3.0.3 spec. The query endpoint returns an answer with its sources and a conversation ID for follow ups, and the deep search endpoint queries the indexed archive with a time range. Rate limits of 60 and 30 requests a minute, error codes and code samples in cURL, JavaScript and Python are published.

Agents, CI runners and IDE plugins reach the hosted MCP server through an OAuth 2.0 device authorization grant, and the only tools that write act on CorpusIQ's own definitions and settings.

Sourcecorpusiq.io/docs/apiread 2026-10-08

A sample company lets a buyer try questions on synthetic data before connecting a single tool, and the sample connects nothing and sends nothing.

No way to test a change is described: no saved test questions, scored runs or before and after comparison when a metric definition or a skill is edited, and no review of answer quality over time.

Sourcecorpusiq.ioread 2026-10-08

CorpusIQ reaches every system through APIs and user authorized OAuth. The security page states that it uses no device agents and no browser extensions, and nothing describes the product or an agent operating a browser, a desktop or a computer session.

Sourcecorpusiq.io/securityread 2026-10-08