Rogo

Writes go back into outside systems, and the connector catalog is large and named. Senior users "read and write back to their CRM", authenticated action in an outside system stated plainly. Rogo is "embedded directly into your firm's systems and data, from SharePoint and CRM to the financial data platforms your team relies on".

Named this year: Affinity, Microsoft Teams, Moody's, Daloopa, Granola, Dropbox and Slack as expanded connectors; SharePoint, OneDrive and OneNote made "turnkey for all customers with no enterprise agreement required"; and virtual data rooms through VDR providers "including but not limited to SS&C Intralinks", kept in sync so the deal context stays current.

The custom route is open: Custom MCP lets a firm "connect any MCP server, including ones you build yourself, to extend Rogo with your own tools and data", with Rogo calling the customer's server. The Snowflake managed MCP integration has the same shape, with OAuth 2.0 and the customer's existing governance and masking rules preserved. Exports to PowerPoint, Word and PDF are outputs, not integrations.

Sourcerogo.com/news/may-product-update, rogo.com/news/whats-new-november-2025 and rogo.com/news/ai-in-financeread 2026-09-12

Felix turns one instruction into a finished deliverable through multi-step execution across models and agents. Felix is "built to generate full deliverables across decks, models, memos, and dashboards from a single prompt", producing "client-ready PowerPoint decks, Excel models, Word documents, dashboards, and sourced research". Rogo agents "don't just answer questions. They understand financial workflows and execute end-to-end work across deals and investments."

The orchestration is named and architectural: "by orchestrating across leading models from OpenAI, Anthropic, and Google, Felix is able to handle multi-step workflows that would typically require stitching together multiple tools". Rogo calls Felix its "agent harness", a runtime, and describes the system as "agentic end-to-end" for financial workflows.

Multi-participant execution is documented separately: agents "can then be deployed to run components of the process rather than isolated tasks", and Rivanna provides "a collaborative workspace within Rogo for deal teams and their agents". Supporting surfaces compose into it: the Sheets Agent, Code Interpreter for Excel powered by Python, and the Slides Annotator, where a user marks up a deck and lets Felix handle the revisions.

Sourcerogo.com/news/may-product-update and rogo.com/news/rivannaread 2026-09-12

A standing knowledge layer spans the firm's own corpus and external financial data, with grounding enforced at the output. Rivanna extends Projects to "integrate full data rooms" and provides "a centralized, up-to-date repository of facts" that a whole deal team draws on, "querying tens of thousands of files with fine-grained citations". Centralized, kept current and queried repeatedly over a transaction's life, it is a maintained structure rather than context assembled per run.

It spans the firm's corpus and external data. Felix "operates on top of Rogo's underlying data foundation", and Rogo is embedded in SharePoint and CRM plus "market data, filings, research, and proprietary sources", with SharePoint, OneDrive and OneNote turnkey for all customers and VDR contents kept in sync. The Plux acquisition "expands the data and monitoring infrastructure that powers our agents" across UK and European filings, lender updates and court documents.

Grounding is enforced at the output: "every cell in a Rogo-generated spreadsheet now cites its original source", and the platform is built to trace every figure to its source. The knowledge layer is held above the model: Rogo "keeps that layer independent of any foundation model, so a firm can adopt whichever model leads next while its encoded standards, precedents, and agents carry over intact".

Sourcerogo.com/news/rivanna and rogo.com/news/ai-in-financeread 2026-09-12

Verification of output is designed in; approval of action is not. Every figure traces to its source, in-cell citations mean each spreadsheet cell names its origin, and Rivanna cites at a "fine-grained" level across tens of thousands of files, so a reviewer can check the work rather than trust it. The Slides Annotator is an explicit review loop: "mark up your deck directly in the Rogo interface and let Felix handle the revisions, with every iteration trackable". Outputs are framed as "first drafts worth iterating on", and the value claim is closing the gap from "roughly right to client-ready".

No approval gate, review step before execution, policy engine, per agent risk calibration or administrator surface for deciding who signs off on what is documented.

That matters because Felix is sold as autonomous and now runs unattended ("an AI that doesn't wait to be asked", with "scheduled tasks and monitoring, set once and left to work in the background") and writes back to the CRM. Nothing published sits between a background run and that write.

The agent library's "multiple rounds of review by other practitioners" is Rogo's people reviewing Rogo's catalog before release, not the customer's oversight mechanism.

Sourcerogo.com/news/may-product-update, rogo.com/news/whats-new-november-2025 and rogo.com/news/agent-libraryread 2026-09-12

The attestation set is strong and first party, while customer facing access controls go undocumented. SOC 2 and ISO 27001 sit on the security page's compliance strip alongside CCPA, GDPR and the EU AI Act, and a dated announcement states that "Rogo has achieved ISO/IEC 42001:2023 certification, the first international standard for AI management systems", described as "a formal, independent validation". A SafeBase trust center runs at trust.rogo.ai, and a Security Advisory Board has its own page.

On access, "Rogo employs zero-trust, least privilege, and strong authentication" describes Rogo's own internal posture, not controls the customer operates, and the "Full data visibility" line names no mechanism. No SSO, SAML, SCIM, OIDC, RBAC, roles or admin console is published. Granular permission controls, role based access, comprehensive audit trails and customizable governance policies are not described on Rogo's own pages.

Sourcerogo.com/security and rogo.com/news/rogo-achieves-iso-iec-42001-certificationread 2026-09-12

Every answer traces back to its sources, but no record of what the agents did is published.

Provenance is the product's central discipline: outputs are citation-grounded across the corpus, in-cell citations for Sheets mean "every cell in a Rogo-generated spreadsheet now cites its original source", and Rivanna queries tens of thousands of files "with fine-grained citations".

Tracing every figure to its source is real reconstruction of how an answer was reached. The Slides Annotator keeps revision history, "with every iteration trackable", and the credit center gives "administrators the controls to manage" spend.

Citations reconstruct the answer, not the agent. There is no audit log of what an agent did, when and under whose authority, no run trace, no attribution distinguishing one user's agent from another's, no retention period, and no export or SIEM path. The security page's "Full data visibility" line ("control data access and usage with full insight into your operations") names no mechanism.

Sourcerogo.com/security, rogo.com/news/whats-new-november-2025 and rogo.com/news/may-product-updateread 2026-09-12

Memory of a user's preferences and positions now persists across sessions, but no scope, lifetime or controls for it are documented. A product update announces it under its own Memory heading: "Rogo now has memory that tracks your preferences and long-running opinions across your work".

The memory is limited in three ways. What is stored is the user's preferences and standing positions, not commitments the agent recorded from its own runs. Per-user scope is implied by "your work" but not stated as an isolation boundary. And no lifetime, expiry or purge path is documented, and no way to list, inspect, edit or delete an individual memory is published.

The Offset acquisition post describes "agentic systems that develop memory for how financial models are built, updated, and maintained over time", but states "Offset's technology will be integrated into the Rogo platform"; that is a plan, not a shipped capability. The Rivanna repository of facts is a deal team knowledge store, not agent memory.

Sourcerogo.com/news/may-product-update, against rogo.com/news/rogo-acquires-offsetread 2026-09-12

What is published is tenant isolation, not a deployment option. Under "Private data stays private", the security page says "data is stored in siloed environments, isolated from other customer data". Logical separation inside the vendor's own cloud is how multi-tenant SaaS is built; it gives the buyer no choice about where or how the platform runs.

No region is named, and no region selection, EU residency option (despite a London office and European expansion), on premises or air gapped mode, private VPC or single tenant tier is published. Single tenant deployment, zero data retention, bring your own key and AWS Nitro Enclaves isolation are not described on Rogo's own pages.

"Every Rogo deployment is bespoke. Our operating model pairs enterprise-grade security with white-glove partnership, led by ex-finance professionals who act as true change-management partners." That is a services engagement, not a deployment topology, and a bespoke rollout is not a customer environment.

Sourcerogo.com/security and rogo.com homepageread 2026-09-12

The Agent Library is a named product surface with its own page: "hundreds of purpose-built agents across investment banking, private equity, private credit, and more". Adoption is evidenced: "the shared library has been executed more than 430,000 times". Units that run hundreds of thousands of times are working assets, not brochure entries.

The production process describes discrete units: "every agent begins with a practitioner who has performed the workflow professionally... they encode the process, financial logic, checks, and judgment required", then "the agent is tested repeatedly against live companies" and undergoes "multiple rounds of review by other practitioners" before release. A comparable-company screen and a credit agreement review are unrelated jobs, and neither depends on the other.

A customer-authored tier sits on top: Rogo Agents lets a firm "build and deploy custom agents in Rogo to encode your firm's workflows, formats, and expertise". Rogo's forward deployed team also "builds customer-specific agents tailored to each firm's workflows"; that is work inside an engagement rather than a packaged asset.

Sourcerogo.com/news/agent-library and rogo.com/news/may-product-updateread 2026-09-12

Work reaches Felix without anyone invoking it, and requests arrive across a broad set of channels. A product update documents the seam under an Autonomous agents heading: users "ask Felix via email to run scheduled tasks and monitoring, set once and left to work in the background". Schedules and monitoring, configured once, run unattended, and Rogo frames the shift the same way: "Felix represents the next generation: an AI that doesn't wait to be asked."

The monitoring infrastructure was acquired for the purpose. The Plux team built systems "monitoring filings, lender updates, court documents, and company disclosures to surface signals", and Rogo states this "expands the data and monitoring infrastructure that powers our agents". Rivanna adds proactive diligence over full data rooms.

Channels are broad. Felix has an email address and tailors output to the sender's role; senior users "ask for an answer by email" and work "through email or on a phone", and Felix runs "across Rogo's core surfaces, including email", with an iOS app shipped and Slack and Microsoft Teams among the connectors. In Rogo's words, "you can delegate work asynchronously and produce consistent outputs regardless of where the request is initiated".

Sourcerogo.com/news/may-product-update and rogo.com/news/expanding-rogo-coverage-across-european-marketsread 2026-09-12

Work is routed across three named model providers, with administrators in control of the models. Rogo runs on "Model Broker, its model router, which sends each task to the frontier model best suited to it across Anthropic, OpenAI, and Google", with a published rationale that "70% of enterprise queries should use a lightweight, cheaper model" and a worked cost range: "the same task can run at $1.26 on a top-tier model or as low as $0.02 through optimized routing".

Customer and admin control is stated three ways. "Routing across Anthropic, OpenAI, and Google makes the switch a configuration change rather than a rebuild." A firm can "adopt whichever model leads next while its encoded standards, precedents, and agents carry over intact".

And administrators hold the controls: the questions that count are "whether the platform keeps spend efficient and gives administrators the controls to manage it, which is where Rogo's Model Broker and the credit center come in". Model-agnostic architecture posts shipping Claude Opus and GPT 5.5 into the same harness corroborate it.

Sourcerogo.com/news/ai-for-investment-banking, rogo.com/news/introducing-model-broker and rogo.com/news/gpt-5.5-now-available-in-rogoread 2026-09-12

No interface for calling Rogo itself is published. There is no REST or GraphQL API, SDK, developer portal (open or gated), API reference, webhook or A2A agent card, the site has no Developers section, and no docs or developer subdomain exists.

Custom MCP runs the other way. It lets a firm "connect any MCP server, including ones you build yourself, to extend Rogo with your own tools and data": the customer stands up a server and Rogo calls it, with Rogo as the client reaching outward. That is an integration and says nothing about whether Rogo is callable. The Snowflake managed MCP integration runs the same direction, Rogo consuming a governed dataset. Rogo Agents ("build and deploy custom agents in Rogo") is an in-product builder, not external callability.

Sourcerogo.com navigation and footer, rogo.com/news/may-product-updateread 2026-09-12

Rogo tests a great deal, but none of it is an evaluation surface the customer can use. Rogo runs the "Big Finance Benchmark, an internally curated set of real financial tasks designed by our ex-finance team" to evaluate incoming frontier models; it is rigorous, and it tests models for Rogo, not the customer's configuration.

The agent library has a real release gate (each agent "is tested repeatedly against live companies until it performs consistently" and undergoes "multiple rounds of review by other practitioners before release"), but the reviewers are Rogo's practitioners and the gate is Rogo's QA of its own catalog.

Nothing customer-facing is published: no sandbox or dry run before a custom agent runs on live deal material, no evaluation harness a firm can point at its own agents, no scored test cases, no regression check when a firm edits an encoded workflow, and no A/B comparison. The asymmetry is worth noting: customers can "build and deploy custom agents in Rogo to encode your firm's workflows", and Rogo documents an exacting review process for its own agents while publishing no equivalent for the ones the customer builds. Verifying an output through citations is not testing a change.

Sourcerogo.com/news/agent-library and rogo.com/news/gpt-5.5-now-available-in-rogoread 2026-09-12

Rogo's architecture runs entirely on programmatic interfaces, so there is no browser or computer control. No hosted or local browser, desktop session, or remote or local computer control is documented.

Rogo reaches outside systems through named connectors to Affinity, Teams, Moody's, Daloopa, Granola, Dropbox and Slack; SharePoint, OneDrive and OneNote; VDR providers including SS&C Intralinks; OAuth 2.0 into Snowflake's managed MCP server, preserving the customer's own governance and masking; and Custom MCP for anything else. A vendor that answers every reach-out problem with a connector or an MCP server is not driving screens.

Three things that might look like computer use are not. "Code Interpreter for Excel... powered by Python" is code execution inside the platform's own runtime. The iOS app and the web app are Rogo's own clients. And the Slides Annotator lets a user "mark up your deck directly in the Rogo interface", Rogo's own surface, not PowerPoint being driven.

Sourcerogo.com/news/may-product-update and rogo.com/securityread 2026-09-12