Synesis FinOps guide
Ask about Azure and AI spend, security and reliability in this chat. Everything below is typed in the
Chat tab. Type menu there for the same commands as buttons.
Get started
- Open the Chat tab and type
login. Sign in with the Microsoft work account you use for Azure. - Ask a question in plain language, for example "What is our total spend this month?"
- Your daily brief arrives each morning in this chat. Type
turn on alertsto also be told within 15 minutes when something unusual happens.
Ask
Spend questions without a narrower scope cover every Azure subscription you can read, plus every AI provider your organization connected.
Azure and AI, per subscription and provider.
Top services across all subscriptions.
What changed and which resources drove it.
Idle and oversized resources, orphaned disks and IPs, Advisor.
Logs, metrics, activity log and resource health.
Once a provider is connected.
Name a subscription or resource group to narrow it down. reset starts a fresh
conversation.
Alerts and daily brief
- The daily brief is on for everyone: each morning, one card with cost and month to date, resources
created, security to attend to, out-of-support versions and savings, across Azure, Google Cloud and your
organization's AI providers, with the parts you can see.
turn off briefstops it;brief nowsends today's right away;brief azure,brief gcporbrief aione source only. turn on alerts: alerts within 15 minutes of resource-creation bursts, expensive new resources, cost spikes, budget overruns, new high-severity vulnerabilities, open management ports, privileged role grants and health incidents, in Azure and Google Cloud.alert settings: see and change thresholds, or just ask: "alert me above 20 new resources an hour", "only Azure in my brief", "send my brief at 9".check now: run every check now.turn off alerts & briefstops both.
Alerts, checks and the brief are part of every plan.
Connect AI providers
An owner or admin connects a provider once for the whole organization; then the people they allow can ask about it.
- In the Chat tab type
connect(orconnect openai,connect anthropic). - Press the button on the card. A Teams window opens.
- Pick the provider, paste the key, and choose who can ask about it: owners and admins only (the default), named people, or everyone.
- The key is checked with the provider and stored in Key Vault. You get a confirmation in chat.
audit log.| Provider | What to create | What you get |
|---|---|---|
| OpenAI API organization | An Admin API key (sk-admin-…), created by an organization owner
at platform.openai.com → Organization → Admin keys. Project keys do not work. |
Usage and cost by project, model and user |
| Anthropic API organization | An Admin API key (sk-ant-admin…): Console → Settings →
Organization → Admin keys. Regular API keys do not work. | Usage and cost by workspace, key and model |
| Claude Enterprise seats | An Analytics API key, created by a Primary Owner in claude.ai Admin settings. | Per-person usage and seat cost |
| ChatGPT Enterprise | No key: a workspace owner exports Workspace analytics → Users as CSV and drops the file in this chat. | Per-person usage |
Google Cloud
Google Cloud works like Azure: with your own access, read-only. Type login gcp in the Chat tab
(or use Sign in with Google in the Connections tab) and sign in with your Google account once. Detailed cost
comes from your Cloud Billing export in BigQuery. Type gcp export (or use Find or set up billing
export in the Connections tab): I look for an existing export in your billing accounts' projects and an owner or
admin switches it on with one button. If there is none and your Google account has the rights, I create the dataset
with you after you confirm, then give you the one console step Google requires, and tell you when the data arrives.
The same gcp export can add the savings recommendations and Security Command Center findings
exports. When you're signed in to Google, your alerts also cover Google Cloud: privileged role grants, public
access, firewalls open to the internet, new service account keys, logging changes and new security findings.
Ask "who changed IAM in my project this week?" to search the audit logs.
You need Billing Account Viewer, or BigQuery Data Viewer on the export plus Job User
on a project. logout gcp signs out of Google only.
Managing connections
The Connections tab at the top of this app shows everything in one place: owners and admins add, rotate and remove keys there; everyone else sees the connections shared with them. The chat commands below do the same.
connections: what is connected and who can ask about each.rotate key for <name>: replace a key (same Teams window).remove connection <name>: delete the key from Key Vault (asks you to confirm).
AI access: who has what, and is it approved
For owners and admins. Every night I read the connected providers' users, seats and API keys, and during your daily brief I read Entra with your own access: which of those people still exist and are enabled, which enterprise apps are AI tools, who is assigned to them and whether they use single sign-on. I compare it with the approved-AI register and tell you about leavers who still have access, keys owned by personal email addresses, unused keys, tools outside SSO and tools nobody approved. Nothing is blocked: you act in Entra or the vendor console.
ai access: the inventory and what was read ·ai findings: the checks ·ai access refresh: collect now.register: the approved-AI register (everyone can see it) ·approve tool cursor owner pat@contoso.com,restrict tool …,ban tool …,remove tool …. The Connections tab edits how a tool is recognised (Entra app ids, domains, names). Restrict conditions (who, which accounts, spend cap, SSO, models, review date) are set in the portal.history tool <name>: every change to a tool and what it replaced ·undo decision <name>: put the entry back as it was, within 30 days (after that, the portal's Re-apply previous state). Every change is kept in an append-only ledger, copied daily to locked storage; an entry changed outside Synesis shows as drift in its history and in the brief, and is never changed back.- Shadow AI radar: AI apps people consented to or signed into with Entra that are banned or not decided. It only sees apps people sign into with Entra; browser use of consumer AI sites needs Defender for Cloud Apps. The consent check says when people may consent to apps themselves and points to Entra's admin consent workflow.
actions tool <name>: after you restrict or ban a tool, what Synesis could change for it on your LiteLLM gateway, in Entra and in Anthropic and OpenAI, each before and after. Small changes (T1) are applied in chat after you confirm; changes that can cut off people now (T2) are made in the portal with the tool's name typed and an MFA sign-in. Synesis never changes Entra itself: it names the setting and links the app, andcheck handoff <id>records it once Entra has it.undo action <id>undoes one change within 30 days.- Allow Synesis to make changes: Anthropic and OpenAI connections are read-only until an owner turns
this on in the Connections tab (after an MFA sign-in, in the portal only). Then the action panel also offers
to deactivate a restricted or banned tool's Anthropic API keys (Undo reactivates them) and to remove people
the conditions don't name from its Anthropic workspaces or OpenAI projects (Undo adds them back with their
role). Choking an OpenAI project's rate limits is a hand-off: Synesis names the numbers to set in the
project's Limits tab and
check handoff <id>reads them back.stop changes <name>makes a connection read-only again at once. - Today's spend (estimate): for a restricted tool with a monthly spend cap, Synesis prices today's token usage at list price through the day, for OpenAI and Anthropic, and the nightly figure replaces it. The LiteLLM gateway teams' spend today is shown apart (never added in) and stays an estimate: there is no nightly figure for it. Both are always labelled estimates.
- Kill switch: on a capped tool, an owner or admin picks in the portal what should happen when the
estimate crosses the cap (deactivate its Anthropic keys, block its gateway keys, hand off an OpenAI choke or
disabling Entra sign-in) and confirms with the tool's name and MFA. It then runs once by itself, every change
with Undo, and the owners are told in Teams. Without one, the owners get a card proposing what to do.
kill switch <name>shows it;revoke kill switch <name>ends it. - AI usage visibility: how people appear in AI usage, everywhere (the portal, Teams, the brief, alerts
and answers): named for owners and admins (the default; members see pseudonyms for everyone but themselves),
pseudonymous, or business-unit totals only.
usage visibilityshows it; owners show fewer names withusage visibility pseudonymousorusage visibility units, and more names in the portal (Settings > AI usage visibility) after a multi-factor sign-in. In pseudonymous mode owners and admins canreveal Person 1a2b3c4d; each reveal is in the audit log and the owners are told. - Cost map: owners and admins say which business unit each Azure subscription, resource group or
chargeback tag, Google Cloud project, OpenAI project, Anthropic workspace, LiteLLM team and Entra department
belongs to.
cost mapshows it with suggestions from people's departments (never applied by themselves);map <kind> <key> to <unit>andunmap <kind> <key>change it after a confirm, andundo action <id>undoes a change within 30 days. - Chargeback:
chargeback(orchargeback 2026-10) shows the month's spend by business unit, as your own Azure and Google access and the connections you may see hold it. Every dollar is under a unit or listed as Unmapped, currencies are never added together, and the LiteLLM gateway's own figure is shown apart. The CSV download is in the portal (Spend > Business units). - Top users:
top users(ortop users openai 2026-10) shows who used one AI tool in a month and what it cost, adding up to the connection's total, with a note where a tool can't name people. People show as your organization's AI usage visibility allows. - Right-sizing:
right-sizing(owners and admins, also on the Savings page) lists heavy use of a premium model with short outputs and the estimated monthly saving at list price on a smaller model of the same vendor. They are candidates, never added to the savings total: without prompts Synesis can't know the task, and usage APIs carry no latency. - PTU vs pay-as-you-go:
ptu(orptu vs payg; also on the Savings page) reads each Azure OpenAI deployment's last 30 days of token counts from Azure Monitor with your own Azure access and says, at Azure list prices, whether provisioned throughput (PTU) would cost more or less than pay-as-you-go: "at current traffic PTU breaks even at N% utilisation; you are at M%", with the monthly difference. A deployment with too little traffic or history, or whose prices Synesis isn't sure of, says why instead of advising. It is advice, never added to the savings total. You need Reader or Monitoring Reader on the Azure OpenAI account. - Asking for an AI tool (the paved road): anyone can ask for a tool your organization approved or
restricted with
request tool <name>or in the portal (AI governance > AI tool requests): what it is for, your team, the expected spend a month and the kind of data. Your owners get it in Teams; an owner or admin approves or declines it (never their own) after seeing exactly what approving sets up. Per tool, owners and admins choose in the portal what an approved request gets: a key on your LiteLLM gateway under the team's gateway team (you collect it once, in the portal, never in a chat), a place in the tool's Anthropic workspace or OpenAI project, and its Entra group or access package. A restricted tool's conditions take the person in first. Everything it sets up is in the tool's history with Undo, andundo request <id>undoes all of it at once (keys and teams are blocked, never deleted). If Synesis can't take someone out of the group as you (Entra group changes are off, or your sign-in lacks the permission), the undo hands that step to an admin and still takes the rest away. Once a tool's kill switch has fired this month, no request for it is approved and no key is collected until an owner reviews the switch. A new workspace or project per team, or more than $200 a month, is approved in the portal with the tool's name typed and an MFA sign-in. - Ask in plain words: "who has access to OpenAI?", "which leavers still have AI keys?", "is Cursor approved?".
Roles
| Role | Can |
|---|---|
| Member | Ask questions, use alerts and the brief, ask about connections shared with them, and see their own activity in audit. |
| Admin | Everything a member can, plus connect, rotate and remove providers, the AI access inventory and
the approved-AI register, and the whole organization's audit log. |
| Owner | Everything an admin can, plus make admin <email>,
make owner <email> and revoke <email>. |
Type people to see who the owners and admins are. Your organization's first owner is set by Betahub
when you sign up; roles in Azure or Entra do not grant any role here.
All commands
| Type | What it does |
|---|---|
login · logout | Sign in or out with your Microsoft work account |
menu | Every command as a button |
help | Commands and examples in chat |
plan | What your organization's plan includes |
brief now · check now | Today's brief · run every check |
brief azure · brief gcp · brief ai | Today's brief for one source |
alert settings | See and change alerts, thresholds and what the brief covers |
turn on alerts · turn off brief | Start alerts · stop the daily brief |
connections · connect | AI providers: list them · add one (owners and admins) |
people · audit | Owners and admins · recent activity (the organization's for admins) |
ai access · ai findings · ai access refresh | AI access inventory · its checks · collect now (owners and admins) |
register · approve tool <name> | The approved-AI register · decide a tool (also restrict, ban, remove tool) |
history tool <name> · undo decision <name> | A tool's changes and what each replaced · undo its latest decision within 30 days (owners and admins) |
request tool <name> · requests | Ask for an approved or restricted AI tool · your requests (and, for owners and admins, the ones to review) |
approve request <id> · decline request <id> · undo request <id> | Decide a request after seeing what approving sets up · undo everything it set up (owners and admins) |
reset | Start a fresh conversation |