AI in OneStream

Put AI on your financials, without letting go of the data.

An independent guide to OneStream's native SensibleAI and its Finance Agentic Layer, and to the Claude, OpenAI and Microsoft Copilot plans that keep sensitive financial data protected at the lowest list cost.

Book an assessment See capabilities

Prices and terms as of October 2026 · sources linked throughout

Quick answer: lowest-cost protected plan LIST PRICE · OCT 2026

On Microsoft 365	Copilot Chat for all, Copilot seats for power users
FP&A team on Claude	Claude Team, $20/user/mo · Enterprise for audit logs
Standardized on ChatGPT	ChatGPT Business, $20/user/mo · Enterprise for audit logs
AWS, Azure or GCP commit	Pay-per-token API in your own cloud
Federal (civilian)	SensibleAI Forecast & Studio (FedRAMP High) + gov-cloud assistant

annual billing · details and caveats below

Two ways AI reaches your OneStream data

Keep AI inside OneStream, or bring the assistant your people already use.

OneStream now offers two routes into the same governed data. Both are administered and logged in one place, the SensibleAI Control Tower, and they aren't exclusive: you can use both.

Path A

Native SensibleAI

AI that runs inside OneStream's own Azure environment.

SensibleAI Forecast: machine-learning forecasting with models built for your data.
SensibleAI Studio, plus AI plugins for reconciliation anomaly detection, System Diagnostics, Cube Commentary and AI Narrative Analysis.
SensibleAI Agents, generally available since May 19, 2026: Finance Analyst (plain-language cube queries and variance analysis), Search, Deep Analysis, and a Forecast Agent bundled with Forecast, which OneStream's September 2026 upgrade guide still calls Beta. The agents work in OneStream and in Microsoft 365 (Excel, Teams).
Where it runs: Forecast and Studio run on Xperiflow, a single-tenant Azure stack with resources not shared across customers, “exclusive to OneStream Azure SaaS customers.” Forecast and Studio use Azure OpenAI models, and OneStream states the prompt “never leaves the Azure ecosystem.”
What governs it: Control Tower roles and per-agent access mapped to OneStream security groups, per-user and per-company credit limits, and an admin switch to turn generative AI off. OneStream “does not use your data to train our own models,” and support staff get access only when you grant it in a ticket.

Sources: Xperiflow architecture FAQ · GA announcement · AI portfolio · Xperiflow 4.4 guide

Path B

Bring your own assistant

The Finance Agentic Layer lets Copilot, Claude, ChatGPT or Gemini query OneStream.

How it connects: the open Model Context Protocol (MCP). Every request passes through OneStream's Agentic Gateway, “the mandatory entry point for every agentic request, regardless of source,” with rate limiting and budget controls built in.
What it does today: OneStream has demonstrated searching and running existing Cube View reports and pulling slices of cube data from Claude. Workflow-based triggers are described as coming.
What you need: the Agentic Finance Toolkit, available from Platform 9.3, with the assistant connected from the Control Tower's Integrations page.
What governs it: each call is “authenticated against the user's OneStream identity” and enforced against their existing role-based permissions, period locks and approval gates. “If a user doesn't have access to a data set in OneStream, neither does their agent.” Admins see usage, and MCP calls show in Credit History.

Sources: Agentic Layer · Splash 2026 recap · Agentic Layer for developers · Agentic Finance Toolkit

The data boundary, stated plainly

OneStream says your data “stays within your governed OneStream environment,” and that third-party agents “receive governed outputs.” So the source data stays put, but the answers your people pull into Claude, ChatGPT or Copilot are then held under that vendor's plan terms. That's why the plan you choose matters as much as the OneStream setup.

Use native SensibleAI when…

– you want processing to stay inside OneStream's Azure environment
– you need machine-learning forecasting
– you're a federal civilian agency: only Forecast and Studio are in OneStream's FedRAMP High authorization (DoD still under review)

Use the Finance Agentic Layer when…

– your people already work in Copilot, Claude or ChatGPT and want one assistant across OneStream, email and documents
– you accept that query results go to the assistant vendor under its commercial terms
– you want to build custom agents or register your own MCP client

OneStream's process

How a OneStream customer actually turns AI on.

Built from OneStream's own documentation as of October 2026. Steps tagged Documented are what OneStream publishes; steps tagged We recommend are our practice. OneStream doesn't publish a formal readiness or pilot program.

00
Decide scope before you buy
We recommend

Inventory candidate use cases: variance commentary, close-status Q&A, reconciliation triage, forecasting. Classify the data each one touches (unreleased results, payroll, personal data). Choose Path A, Path B or both, pick the assistant plan, and brief your external auditor early.
01
Confirm hosting and version
Documented

SensibleAI is “exclusive to OneStream Azure SaaS customers.” Minimum platform versions: 9.3.0 for SensibleAI Agents, 9.3 for the Agentic Finance Toolkit, and 9.3.1+ for the Xperiflow 4.3 and 4.4 upgrades. The Xperiflow 4.3 upgrade “is not self-service”: it needs a support request and a booked downtime window.

Release notes · Xperiflow 4.3 guide
02
Contract for an AI Services environment
Documented

“You must complete a separate contract outside of the standard OneStream application environment,” with “limited availability to OneStream environments enabled with AI Services.” SensibleAI tiers are sized to data volume. OneStream publishes no AI prices. SensibleAI Forecast, Studio and Agents are also transactable on the Microsoft Marketplace.

Forecast setup · Microsoft partnership
03
Support ticket and install
Documented

Installation starts with a support ticket: OneStream sends a download link and a meeting invite, installs the data-science engine and guides the solution install. From Solution Exchange you load the AI Dependency Bundle (“required for all AI Services solutions”), the SensibleAI Agents Configuration Bundle and Xperiflow Administration Tools.

Solution Exchange · Agents overview
04
Build infrastructure
Documented We recommend

In the Control Tower: Build Infrastructure, select the default application, fill in Roles and Agent Access, then Complete Setup. Setup creates a wernicke.server service account that runs scheduled jobs (a member of Administrators), and OneStream's docs tell implementers to “obtain their explicit approval to proceed.” We recommend recording it as a privileged account in your IT general controls inventory and reviewing its activity.

Infrastructure setup
05
Map security
Documented We recommend

Control Tower roles map to OneStream security groups: SAI Control Tower Viewer (read-only analytics, credits, chat reviews), SAI Administrator (adds credit limits and settings) and SAI Support (adds Data Explorer, audit logs and the upgrade portal). Finance Analyst, Search and Deep Analysis access each map to groups; Xperiflow 4.4 adds Forecast, Studio and All Agents access, which must be assigned at Application or Global scope. We recommend dedicated AI groups, SAI Support limited to named admins, and agent access by role rather than for everyone.

Control Tower settings · Forecast security practices
06
Sync users and set credit limits
Documented We recommend

OneStream's setup syncs users and offers “Split Evenly” to distribute credits. Credits refresh quarterly; admins set company and per-user limits, and Credit History breaks usage down by user and agent. We recommend explicit per-user limits during a pilot instead of an even split.

Credits
07
Configure Finance Analyst per application
Documented

Setup is required “for each desired application.” Connect Cubes comes first; Set UD Types, Create Personas and Set Accounting Period follow in any order. Optional context (dimension defaults and members, business and cube context, report row templates, Cube Views) is optional. “Finance Analyst is not available to end users until all required steps are completed.”

Finance Analyst setup
08
Connect external assistants (Path B)
Documented
Everything starts on the Control Tower's Integrations page.
- Claude: copy the connection details, add a custom web connector under Organization Settings › Connectors, and upload OneStream's skill file as an organization Skill.
- ChatGPT: create the app with advanced OAuth settings, raise a OneStream support ticket to allowlist the callback URL (“ChatGPT generates random callback URIs”), add the MCP connector and upload the skills file.
- Microsoft Copilot: create an OAuth client in the Microsoft Developer Portal (“My organization only”, PKCE on), enter the reference ID in OneStream, and upload the app package in the Microsoft 365 admin center.
- Gemini and custom MCP clients: supported too; admins can create custom client connections themselves.
AI platform connections
09
Pilot read-only, then widen
We recommend

One application, a small named group, read-only questions, which is what OneStream's demonstrated MCP tools do anyway. Score answers on the Control Tower's Evaluations page, which tests whether responses are accurate, repeatable and explainable. Measure value with your own timings, not vendor claims.

Evaluations
10
Gate any write-back
Documented We recommend

The latest SensibleAI Agents release adds Commentary Write-Back: “every proposed write” is shown for review before saving, and it respects OneStream security, workflow locks and read-only settings. We recommend keeping anything that affects journals out of scope until the controls in the checklist below are designed and tested.

Release notes
11
Monitor and upgrade
Documented We recommend

Audit logs capture time, action, entity, user and “the before and after state of the record.” Upgrades run through an “AI Services Upgrade” service request and the Upgrade Portal; “while a port is running, the system locks.” We recommend scheduling ports outside close windows.

Environment upgrade

Compare plans

Claude vs OpenAI vs Microsoft Copilot, for financial data.

Business tiers only. Consumer plans don't belong near your financials (see below). List prices and terms as of October 2026. Confirm with the vendor before you buy.

Scroll sideways to compare →

Claude vs OpenAI vs Microsoft Copilot business plans: list prices and terms as of October 2026
Plan detail	Claude Anthropic	OpenAI ChatGPT · API · Azure OpenAI	Microsoft Copilot Microsoft 365
Business plans & list prices	Team Standard $20/user/mo annual, $25 monthly Team Premium (5× usage) $100 annual, $125 monthly Enterprise $20/seat/mo annual + usage at API rates API per 1M tokens in/out: Haiku 5.5 $0.10/$0.50 (prompts up to 100K tokens) · Sonnet 5.5 $2/$10 · Opus 5.5 $4/$20 · Fable 5.1 $10/$50	ChatGPT Business $20/user/mo annual, $25 monthly Enterprise contact sales API per 1M tokens in/out: GPT-6 Luna $0.10/$0.50 · GPT-6.1 Sol $2/$10 · GPT-6 Astra $10/$50 Azure OpenAI Global Standard at the same rates; Data Zone +10%	Copilot Chat no added cost with an eligible Microsoft 365 plan Copilot Business (≤300 users) $21/user/mo yearly, $18 promo for the first year through Dec 31, 2026 Microsoft 365 Copilot add-on $30/user/mo yearly M365 E7 (includes Copilot) $99 Agents $0.01/message pay-as-you-go; Copilot Studio $200 per 25,000 credits/mo
Minimum seats	Team: 2 (max 150). Enterprise: 20 self-serve, 50 sales-assisted	Business: 2. Enterprise: not published	Copilot Business: up to 300 users, on an eligible Microsoft 365 Business plan. Add-on: a qualifying Microsoft 365 plan
Trains on your data?	Team, Enterprise, API: no by default. Exception: chats users submit as feedback are kept 5 years and may be used. Admins can turn “Rate chats” off	Business, Enterprise, API: no by default. Azure OpenAI: prompts “NOT available to OpenAI,” not used for training	“Aren't used to train foundation models” under Enterprise Data Protection
Retention controls	Enterprise: custom retention (minimum 30 days). Team: no custom control. API: zero data retention by agreement, with exclusions (Batch, Files, Claude for Excel, the chat apps)	Enterprise: admin-set retention; compliance logs held 30 days (export them). API: abuse logs up to 30 days; zero retention needs approval	Microsoft Purview retention and eDiscovery apply. Keep the optional “Anthropic models with Data Retention” setting off
SSO / SCIM / audit	Team: SSO; no SCIM or audit logs. Enterprise: SCIM, audit logs, Compliance API, IP allowlisting	Business: SAML SSO; no SCIM or Compliance API. Enterprise: SCIM, role-based access, Compliance API	Entra ID identity; prompts and responses in the unified audit log. E5: 10-year advanced audit
Data residency	First-party API: US-only option at 1.1×; no EU inference. EU via Amazon Bedrock EU profiles or Google Cloud “eu” (+10%)	API: US and EU processing (EU needs zero retention or another approved retention arrangement), +10% on newer models. ChatGPT Enterprise: data-at-rest residency. Azure: regional or Data Zone	EU Data Boundary service, but Anthropic models are excluded from it
Customer-managed keys	Not documented for Claude plans; ask Anthropic	ChatGPT Enterprise EKM; API per project; Azure customer-managed keys	Not covered in the sources we reviewed; ask Microsoft
Compliance	SOC 2 Type II, ISO 27001, ISO/IEC 42001, HIPAA BAA. FedRAMP High only via Claude for Government, Bedrock GovCloud or Vertex Assured Workloads	SOC 2 Type 2; ISO 42001 for ChatGPT Enterprise; HIPAA BAA for eligible API customers; FedRAMP 20x Moderate (ChatGPT Enterprise, API). Azure OpenAI in Azure Government (FedRAMP High)	ISO 42001; HIPAA supported (except web search). Available in GCC, GCC High and DoD, without Anthropic or OpenAI-operated models
Cloud-hosted options	Amazon Bedrock (no Anthropic operator access); Claude Platform on AWS (counts toward AWS commitment); Microsoft Foundry “Hosted on Azure” (MACC-eligible); Google Cloud	Azure OpenAI in Microsoft Foundry; OpenAI Private Link over Azure (through sales)	Runs in Microsoft 365; Copilot Studio can call Foundry models (billed separately)
Connects to OneStream?	Yes: organization connector + skill	Yes: ChatGPT MCP app; OneStream ticket to allowlist the callback	Yes: OAuth client + admin-center agent. SensibleAI Agents also run in Teams and Excel
Fits	FP&A teams that want Claude's models and Excel add-in; AWS or GCP shops that want API control	Teams standardized on ChatGPT; Azure shops that want OpenAI models in their own tenant	Companies already on Microsoft 365 that want governed AI at the lowest added license cost

Matching tiers cost the same per token: Claude Sonnet 5.5 and GPT-6.1 Sol are both $2 in / $10 out per million tokens, and Haiku 5.5 and GPT-6 Luna both $0.10 / $0.50. At those tiers the choice is about fit, controls and where you already have contracts, not token price. Token counts also differ: Claude 4.7 and later models produce about 30% more tokens for the same text.

Sources: Claude pricing · Claude Enterprise · Claude API pricing · Claude training policy · ChatGPT plans · OpenAI API pricing · OpenAI trust · Azure OpenAI privacy · Copilot pricing · Copilot data protection · Anthropic in Copilot · full list

Which plan fits

The lowest-cost plan that still protects your financial data.

Pick the scenario closest to yours. Each gives the cheapest plan we'd sign off on, why, how to set it up, and what to watch.

A · Standardized on Microsoft 365

Copilot Chat for all, Copilot seats for power users

Plan: Copilot Chat at no added cost for everyone; full seats only for heavy Excel and analysis users: Copilot Business at $21 ($18 first-year promo through Dec 31, 2026) if you have 300 or fewer users on Business plans, otherwise the $30 add-on.
Why: Copilot runs under Microsoft's data-processing terms, respects existing permissions and sensitivity labels, and lands in Purview audit and eDiscovery. Since July 1, 2026, Copilot Chat in E3/E5 and Business plans adds Word, Excel and PowerPoint agents.
Set it up: Connect OneStream through the Agentic Layer's Copilot path. Decide per group whether Anthropic and OpenAI-operated models are allowed, keep “Anthropic models with Data Retention” off, and add Purview DLP for Copilot.
Watch for: The Excel, Word and PowerPoint agents use only Anthropic models and stop if Anthropic is disabled. OpenAI-operated models were switched on by default on July 24, 2026 unless you chose otherwise. If you turn on agent pay-as-you-go, set a budget. E5 shops rolling out broadly should price M365 E7 at $99, versus E5 $60 + Copilot $30 + Agent 365 $15 = $105 bought separately (illustrative list-price arithmetic, not client data).

B · Analyst-heavy FP&A team

Claude Team, or Enterprise when you need audit evidence

Plan: Claude Team Standard at $20/user/mo billed annually (2–150 users), with $100 Premium seats only for the heaviest users. If you need audit logs, SCIM or custom retention for SOX evidence, the floor is Claude Enterprise: $20/seat/mo plus usage at API rates, minimum 20 seats self-serve or 50 sales-assisted.
Why: Team doesn't train on your content by default and has SSO and admin controls. It lacks SCIM, audit logs, the Compliance API and custom retention.
Set it up: Enforce SSO, turn off “Rate chats” so feedback isn't kept five years, and connect OneStream with the organization connector and skill.
Watch for: Claude for Excel doesn't inherit custom retention, isn't in Enterprise audit logs and isn't zero-retention eligible; Anthropic advises against it for “audit-critical calculations without verification.” If you're standardized on ChatGPT instead, ChatGPT Business at $20 is the equivalent, with the same gaps.

C · AWS, Azure or GCP commitment

Pay per token, inside your own cloud

Plan: Mid-tier models (Sonnet 5.5 or GPT-6.1 Sol) through your cloud, connected to OneStream as a custom agent or custom MCP client: for example, automated variance commentary.
Options: AWS: Claude on Amazon Bedrock (no Anthropic operator access, retention “none” enforceable org-wide, PrivateLink). Azure: Azure OpenAI in Foundry (prompts not available to OpenAI, Data Zones, private endpoints) or Claude “Hosted on Azure” (prompts stay in Azure under Anthropic's data-use terms; MACC-eligible). GCP: Claude on Google Cloud.
Watch for: Batch processing is 50% cheaper but isn't zero-retention eligible at Anthropic or OpenAI. Fable-class Claude models require AWS's 30-day review retention on Bedrock.

D · Regulated: federal, HIPAA, SOX-heavy

Pick tiers that produce evidence

Federal: OneStream's FedRAMP High authorization covers SensibleAI Forecast and Studio only, approved for civilian agencies, with DoD “under review.” Copilot is available in GCC, GCC High and DoD without Anthropic or OpenAI-operated models; OpenAI has FedRAMP 20x Moderate for ChatGPT Enterprise and the API; Claude reaches FedRAMP High only through Claude for Government, Bedrock GovCloud or Vertex Assured Workloads. Defense contractors: Anthropic's government page describes a Department of War supply-chain designation under litigation. Check with counsel.
HIPAA: Claude API BAA (self-serve) and Claude Enterprise; OpenAI BAA for eligible API customers; Copilot supported except web search.
SOX: Claude Enterprise (audit logs, Compliance API), ChatGPT Enterprise (Compliance API: logs kept 30 days, so export to your SIEM) or Copilot (unified audit log; E5 10-year audit).

E · Small team on a budget

Copilot Chat at no added cost, a few paid seats

Plan: On Microsoft 365 Business: Copilot Chat for everyone plus Copilot Business for one to three power users. Otherwise Claude Team or ChatGPT Business, both $20/user/mo billed annually with a two-user minimum and no training by default.
Watch for: Neither Claude Team nor ChatGPT Business has audit logs or SCIM. Compensate with MFA in your identity provider, OneStream-side audit logs and read-only scope. Pay annually: monthly is $25 on both.

F · EU-based or EU data subjects

Keep processing in the EU on purpose

Options: Copilot is an EU Data Boundary service, but Anthropic models are excluded and off by default in the EU, EFTA and UK. Azure OpenAI offers an EU Data Zone. The OpenAI API's EU region requires zero-retention or similar approval. Claude's first-party API has no EU inference. Use Bedrock EU profiles or Google Cloud's “eu” region.
Watch for: Ask OneStream which region processes its own AI. Most EU AI Act rules apply from August 2, 2026, and Annex III high-risk rules from December 2, 2027. Review any HR or payroll-decision use with counsel.

Plans to keep away from financial data

– Claude Free, Pro and Max. Training is opt-out, and if a user allows it, chats may be kept de-identified for up to five years under consumer terms.
– ChatGPT Free, Go, Plus and Pro. OpenAI's plan matrix marks “No training on API or business data by default” as unavailable on Plus and Pro.
– Personal seats on corporate cards. No SSO, no admin controls, no audit trail.
– Specific features on regulated data: ChatGPT's “local computer access” if you need zero retention, and Copilot's “Anthropic models with Data Retention” unless you accept Anthropic's terms and 30-day retention.

Sources: Claude consumer retention · ChatGPT plans · ChatGPT Work security · OneStream FedRAMP · FedRAMP scope · Claude for Government · EU AI Act timeline

Keep costs down

Ten ways to spend less without loosening controls.

01
Tier licenses by role. Copilot Chat for everyone, full seats for heavy users; Claude Team lets you mix Standard and Premium seats. For 40 finance users on E3: 40 × $30 = $1,200/mo, versus Copilot Chat for all plus 8 × $30 = $240/mo (illustrative list-price arithmetic, not client data).
02
Pay annually. Claude Team and ChatGPT Business are $20 annual vs $25 monthly; Copilot $30 vs $31.50; Copilot Business $21 vs $25.20.
03
Use pay-as-you-go with budgets. Copilot agent pay-as-you-go is $0.01/message, off by default, with budget alerts. Employee-facing agents used by Copilot-licensed users are no charge under fair use.
04
Spend existing cloud commitments. Claude in Foundry is MACC-eligible; Claude Platform on AWS counts against your AWS commitment; OneStream SensibleAI is transactable on Microsoft Marketplace.
05
Right-size the model. 500 commentary runs a month at about 30K input and 1.5K output tokens cost about $37.50 on Sonnet 5.5 or GPT-6.1 Sol, versus about $187.50 on Fable 5.1 or GPT-6 Astra (illustrative list-price arithmetic, not client data; token counts are assumptions).
06
Cache repeated context. Cache hits on Claude Opus 5.5 and Sonnet 5.5 cost 0.05× base input; GPT-6.1 Sol cached input is $0.10 per million tokens.
07
Go regional only when required. Regional or US-only processing typically adds 10% (Anthropic 1.1×, OpenAI on newer models, Azure Data Zone).
08
Don't reserve capacity early. Azure provisioned throughput starts at 15 units billed hourly whether used or not: about $10,950/mo, or about $3,315/mo on a one-year reservation (illustrative list-price arithmetic at $1/unit-hour and $2,652/unit-year, not client data).
09
Start read-only with OneStream credit limits. Credits refresh quarterly and can be capped per user; the Agentic Gateway adds budget controls.
10
Audit usage-based billing. Some Copilot services bill credits on top of the seat, and “auto-apply new services” is on by default in spending policies.

Sources: Copilot meters · Copilot credit rates · Claude in Foundry · Claude on AWS · Claude residency · Azure provisioned throughput · Copilot usage billing

Security checklist

The controls we put around AI and OneStream.

Mapped to the OWASP Top 10 for LLM and agentic applications, the MCP specification and NIST's AI Risk Management Framework.

ControlRisk it addressesHow

Per-user identity, no shared super-user

Risk: Privilege abuse; excessive agency

Use the Agentic Layer's per-user authentication. Never give an assistant an admin's personal access token: it “takes on the identity and access rights of the creating user.” Keep non-interactive access roles tightly assigned.

Approve and watch the AI service account

Risk: Privileged access

Explicit approval for wernicke.server (OneStream's own instruction), an entry in your ITGC inventory and periodic activity review.

Least-privilege role mapping

Risk: Sensitive information disclosure

Dedicated groups for each Control Tower role and agent; SAI Support (Data Explorer, audit logs) only for named admins; agent permissions at the right scope.

Read-only first; human approval for writes

Risk: Excessive agency; unreviewed postings

OWASP calls for minimal scopes and “user approval for high-impact actions”; the MCP spec says “there SHOULD always be a human in the loop.” OneStream's write-back shows every proposed write for review.

Commercial tiers, retention set

Risk: Vendor retention and training

Commercial tiers only: Claude Team or Enterprise, ChatGPT Business or Enterprise, API, or Copilot under Microsoft's data-processing terms. Set Claude Enterprise and ChatGPT Enterprise retention; turn off Claude's “Rate chats.”

Model-provider governance in Microsoft 365

Risk: Data leaving Microsoft's boundary

Decide Anthropic and OpenAI subprocessor access per group and keep “with Data Retention” models off. Models hosted on Azure by Microsoft mean “data does not leave Microsoft.”

Data loss prevention and web-search limits

Risk: Sensitive information disclosure

Purview DLP for Copilot can block labeled files, sensitive prompts and web search. It “can't scan the contents of files that you upload directly into prompts,” so set a policy on uploading OneStream exports. Copilot web queries go to Bing outside the data-processing terms.

Prompt-injection defenses

Risk: Prompt injection; goal hijack

Treat workbooks and documents as untrusted input, keep external content separate and require approvals. OWASP notes that retrieval and fine-tuning “do not fully mitigate” injection.

MCP hygiene

Risk: Tool poisoning; shadow servers

Allow only OneStream's MCP server and vetted connectors, review tool descriptions and pin versions. The MCP spec requires OAuth 2.1 with audience-bound tokens; token passthrough is “explicitly forbidden.”

End-to-end logging into your SIEM

Risk: Insufficient audit; SOX evidence

OneStream Control Tower audit logs and MCP calls in Credit History, plus the vendor side: Claude Enterprise audit logs, the ChatGPT Enterprise Compliance API, the Purview unified audit log or Bedrock CloudTrail.

Private networking and keys

Risk: Data in transit and at rest

Bedrock PrivateLink, Foundry private endpoints, OpenAI Private Link (through sales), Claude Enterprise IP allowlisting; Azure customer-managed keys and OpenAI EKM where offered.

Spend and rate caps

Risk: Unbounded consumption

OneStream credit limits, Gateway rate and budget controls, Copilot pay-as-you-go budgets.

Change control over prompts, skills and settings

Risk: Control drift; vendor changes

Version the skill files you upload to Claude and ChatGPT, watch vendor default changes, and keep beta features off until reviewed.

Human review of anything that touches the financials

Risk: Confabulation; auditability

COSO's generative-AI guidance (summarized by Deloitte) calls for human oversight of outputs affecting journal entries, reconciliations, estimates or management review controls, with logs of the model, prompts, inputs, outputs and approvals.

Test before and after go-live

Risk: Accuracy and regression

OneStream Evaluations (accurate, repeatable, explainable) and adversarial testing per OWASP's GenAI red-teaming guide.

Auditor and counsel alignment

Risk: Audit evidence; unreleased results

PCAOB Release 2024-007 on technology-assisted analysis (effective for audits of fiscal years beginning on or after December 15, 2025) emphasizes controls over IT and information in electronic form. Have counsel review vendor terms before unreleased results flow to an outside AI service.

Sources: OWASP LLM Top 10 · OWASP agentic Top 10 · MCP authorization · MCP security practices · NIST AI RMF · NIST AI 600-1 · Purview DLP for Copilot · Deloitte on COSO GenAI · OneStream access tokens

AI architecture review

A clear answer on where AI goes, which plan to buy, and how to control it.

No discovery theater. We look at your OneStream environment, your licensing and your controls, and you leave with a decision you can take to your CFO and your auditor.

You get

– Target architecture diagram and decision memo (path and plan)
– Plan and cost comparison workbook at list prices
– Security configuration checklist and risk register mapped to OWASP and NIST
– RACI and a read-only pilot plan with exit gates

Book an assessment

What we review

01 Current state: hosting and version, the path to 9.3, applications, security groups, existing and “shadow” AI use, Microsoft 365 / Claude / OpenAI licensing, cloud commitments
02 Use cases mapped to process, data sensitivity and the financial-reporting controls they touch
03 Path per use case: native SensibleAI, Finance Agentic Layer, or both
04 The lowest-cost plan that meets your controls, with license tiering
05 Security design: roles, service accounts, MCP allowlist, vendor tenant settings, logging
06 Controls narrative and an evidence pack for your external auditor

Typically 3–4 weeks; optional pilot support after. A typical range, not a promise. It depends on environments and access.

Implement

Enable SensibleAI and the Finance Agentic Layer.

Optimize

Right-size licenses, credits and security on an existing AI setup.

Rescue

Fix an over-permissioned or stalled AI pilot.

FAQ

Questions finance and IT leaders ask us.

If we connect Claude, ChatGPT or Copilot, does our financial data leave OneStream?

Source data stays in OneStream, but third-party agents “receive governed outputs.” Those answers are then handled under the assistant vendor's terms, so use a commercial plan with retention controls.

Can the AI see data a user can't?

OneStream says no: access is “enforced against their existing role-based permissions,” and “if a user doesn't have access to a data set in OneStream, neither does their agent.” That also means agents faithfully expose whatever access users already have, so clean security comes first.

Which OneStream version do we need?

9.3.0 or later for SensibleAI Agents, Platform 9.3 for the Agentic Finance Toolkit, and 9.3.1 or later for the Xperiflow 4.3 and 4.4 upgrades. SensibleAI is for OneStream Azure SaaS customers.

Do the AI vendors train on our data?

Not by default on commercial tiers: Claude Team, Enterprise and API; ChatGPT Business, Enterprise and API; and Microsoft Copilot. Consumer tiers differ. Anthropic's exception is chats users submit as feedback, which admins can switch off.

What does OneStream's AI cost?

OneStream doesn't publish SensibleAI, credit or AI Services pricing. It requires a separate contract, and credits refresh quarterly. In a review we model the assistant-license side at list prices and get OneStream's quote alongside it.

Can AI post journals or change cube data?

OneStream's latest release adds commentary write-back, with every proposed write shown for review. OneStream hasn't publicly documented write operations for third-party assistants. We recommend starting read-only.

Is OneStream's AI FedRAMP authorized?

SensibleAI Forecast and Studio were added to OneStream's FedRAMP High authorization, approved for civilian agencies, with DoD under review. Agents and the Agentic Layer aren't part of that announcement.

Can we switch generative AI off?

Yes. Admins can “turn off the Generative AI capabilities” in Xperiflow Administration Tools, and assistant access can be removed by disconnecting the integration and removing agent-access groups.

Sources

All accessed October 23, 2026. Vendors change plans, prices and defaults often. Confirm with the vendor before you buy. This page is guidance, not legal advice.