Platform
See what your AI is actually doing
What your AI systems attempted, what they called, what they read and what it cost — across every connected system rather than one dashboard per vendor.
Across the estate
Systems
9
Actions · 24h
240
One view rather than one dashboard per vendor. Six vendor dashboards cannot be added together, because they do not count the same thing.
Why this exists
Every vendor ships a dashboard. None of them can see the others.
The individual dashboards are fine. The problem is that nobody in the organization can answer an estate-level question, because the six views do not share a subject — one counts tokens, another seats, a third API calls.
OpsAI can answer them because every action passes through one point. That is not a feature so much as a consequence — if you are already deciding every action, you already have the only complete count of what was attempted.
The refusals are part of it, which no vendor dashboard has. A view that reports what succeeded is a usage report; a view that reports what was attempted is an operational one.
- What did AI attempt?
- Across every connected system, in one number, with the refusals included. Six vendor dashboards cannot be added together because they do not count the same thing.
- What did it read?
- Which sources, how often, and whether anything reached for something it was not granted. The reads are usually invisible in per-vendor views because nothing was written.
- Which model did it use?
- Per action, not per month. A model pinned in one place and called from four agents is a dependency you can see the shape of rather than a line on an invoice.
- What is changing?
- The rate of attempts, the held rate, the refusal rate. Direction matters more than absolute value here — a held rate that halved after a threshold change is the useful signal.
What it looks like
Decision volume, and what is behind it.
The shape matters more than the total. This estate is weighted to business hours, which is what you would expect of support and finance workloads — and a flat curve on the same workload would be the interesting finding.
- Decisions · 24h
- 240
- Authorized
- 215
- Held
- 20
- Denied
- 5
- Median decision
- 17 ms
- Inference calls
- 0
Decisions over time
AuthorizedHeld or denied
AI system inventory11 systems, every one owned
| AI system | Answers for it | Autonomy | State |
|---|---|---|---|
| refund-resolver | Priya Nair | L3Act with approval | Acting |
| order-lookup | Priya Nair | L1Observe | Acting |
| ap-invoice-agent | Rahul Menon | L2Advise | Acting |
| payout-runner | Rahul Menon | L4Act autonomously | Watched |
| vendor-onboard | Anita Rao | L3Act with approval | Acting |
| inventory-sync | Anita Rao | L2Advise | Acting |
Actions attempted
240
24 hours
Authorized
215
ran inside their bound
Refused or held
25
invisible in vendor views
Median decision
17ms
policy evaluation
Illustrative240 decisions from the OpsAI sample estate over 24 hours, not a customer deployment or a published benchmark.
Where the activity lands
Which systems your AI is actually reaching.
Derived from the decisions themselves rather than declared. The gap between what an agent was granted and what it actually reaches is usually the more interesting number, and it only exists if you can see both.
Activity by system
9 systems reached of 12 connected| System | Category | Actions attempted | Share of activity | Agents reaching it |
|---|---|---|---|---|
| ZendeskZendesk | Support | 55 | 100% | 3 |
| RazorpayXRazorpayX | Payouts | 33 | 60% | 2 |
| SalesforceSalesforce | CRM | 30 | 55% | 2 |
| BoxBox | Documents | 27 | 49% | 1 |
| GmailGmail | 27 | 49% | 3 | |
| NetSuiteNetSuite | ERP | 27 | 49% | 1 |
| PostgresPostgres | Database | 22 | 40% | 1 |
| Zoho BooksZoho Books | Accounting | 10 | 18% | 1 |
| RazorpayRazorpay | Payments | 9 | 16% | 1 |
Granted but never reached is worth as much as reached often.
A system an agent may touch and never has is a grant that can be withdrawn — the cheapest reduction in exposure available, and one nobody finds without a view that spans the estate. The same is true of a credential nothing derives from.
Models actually called
6 of 7- Claude Haiku2026-04-113private endpoint · approved by Priya Nair
- Claude Sonnet2026-04-112private endpoint · approved by Priya Nair
- GPT2026-03-282vendor API · approved by Rahul Menon
- Gemini Pro2026-02-192vendor API · approved by Sneha Iyer
- Llama2026-01-301self-hosted · approved by Kabir Sen
- Mistral2026-03-051self-hosted · approved by Anita Rao
What to watch
Direction, not absolute value.
The useful signals here are rates that moved. An estate's absolute numbers say more about its size than its health, and the second derivative is where the operational information actually is.
Decisions over 24 hours
authorized against held or deniedDecisions over time
AuthorizedHeld or denied
Three rates worth an alert.
- Refusal rate rising
- Something is reaching for what it was not granted. Usually a deploy, occasionally worse.
- Held rate near zero
- The thresholds were probably set where nothing reaches them, which means the gate is not load-bearing.
- Attempts flat overnight
- For a business-hours workload, an unattended agent that works the night shift at the same rate is worth a look.
From a number to an action
Traces
Any figure here opens into the six stages behind a single decision.
When a rate moves sharply
Incidents
A pattern rather than an instance, and the investigation starts at the trace.
Whether it is improving
Posture
The same estate scored against what good looks like, rather than counted.
Where to start
Ask how much your AI attempted last week.
If the answer requires opening more than one vendor console, the number does not exist yet — and the refusals almost certainly do not appear in any of them.