Know what your AI workflows cost, and how they behave.
For energy retailers, networks, generators and utilities using AI in customer service, field support and engineering. See captured spend, look into unusual activity and keep a record of governance changes.
Customer service
Billing and outage enquiries assistant
US$2,940
41,200 conversations this month
Field support
Asset-history and procedure summaries
US$610
3,850 requests this month
Engineering
Research agent over design documents
US$1,280
920 tasks this month
Synthetic month, for illustration.
Different workflows, different costs, clear ownership.
A customer assistant, a maintenance-document search and an engineering agent need different measures. See each one's AI activity separately, with the context its owner needs.
Demand rises. Is the cost rise expected?
A storm or a billing change brings a surge of enquiries and legitimate use. Repeated calls raise cost without answering anyone else. Read spend beside recorded activity before deciding what needs attention.
Useful assistants can still repeat themselves
An assistant fetching asset records or preparing a summary can call a model over and over. The team that owns it needs to know early, with enough detail to look into it.
Provider bills do not explain each workflow
A provider total does not say what customer service, field support or engineering research cost. Spend needs to land with the team and workflow behind it.
Follow the AI activity behind a field-support request.
The model calls, their cost and what was detected, for one run of an instrumented workflow.
- 1
A technician asks for an asset-history summary
Your field-support assistant opens a run for the request.
- 2
The application retrieves the records
From your own systems, outside SAVI.
- 3
The assistant calls its models
Through the SAVI SDK, tagged with the team, the workflow and the run ID.
- 4
SAVI records usage, cost and detections
Tokens and cost for every call and the personal-data types found in the prompt. An agent making more than 20 calls in 30 seconds is paused.
- 5
The technician reviews the answer
Under your existing procedures. Your application can report the result to SAVI.
One run, as SAVI records it
Synthetic example
- Workflow
- field-support-assistant
- Team
- Network Field Operations
- Model calls
- 6
- AI cost
- US$0.21
- Data types found
- Person ×1 (value not sent to SAVI)
- Result reported
- Accepted after correction
SAVI sees the AI calls your application makes through the SDK. It does not check maintenance advice, and it does not replace your approval procedures.
Cost visibility and governance for the AI layer.
See which workflows drive AI spend
Spend by provider, model, team and workflow, so customer service, field support and engineering each see their own. Calls without a team are grouped and counted, so gaps show.
Where your application reports results, compare spend with completed work, such as a processed document or a closed support case.
Cost dashboardGive a workflow a budget to track
Attach a workflow's runs to a budget pool. Its owner is warned at 70% and 90% and told at 100%. By default nothing is stopped.
Blocking is a separate choice: after 7 days in shadow mode, an administrator can set the pool to refuse the next call at the budget. Weigh that against the workflow's continuity needs first.
Budget poolsCatch repeated calls before they add up
An agent making more than 20 calls in 30 seconds is paused for 120 seconds, in every workspace. You set the limit and the length. The pause is recorded with its trigger and reason.
The pause refuses the agent's next calls where your application runs the pre-call check. Calls already sent finish, and the agent's process keeps running.
Loop detectionMake AI policy changes reviewable
Restrict models and MCP tools for a team. Every change records who made it, which setting, from what to what, and when, and the record is signed.
JSON exports of the audit trail carry a signature a reviewer can check with SAVI's public key. A pass shows the file is unchanged, not that nothing was left out.
GovernanceChoose a measure that fits the work.
Where your application reports a result for each run, SAVI sets AI cost against that unit of work.
Workflow
Example measure
What needs defining
Customer-service assistant
AI cost per conversation
Where a conversation starts and ends
Document processing
AI cost per accepted document
What accepted means; retries inside a run add to its cost
Field-support assistant
AI cost per completed request
The completion status your application reports
Engineering research
AI cost per completed task
Where a task starts and ends, and who marks it complete
Example set-ups, not built-in metrics. A unit cost depends on the results your application reports and the AI calls SAVI captures. It does not show service quality or savings.
Keep operational authority where it belongs.
What SAVI is
Visibility and configured controls for the AI calls your applications make: what was used, what it cost, who owns it, and what happened when activity needed attention.
What SAVI is not
A grid-control, plant-protection or safety-assurance system. Operational decisions, customer escalations and safety procedures stay within your own approval processes.
Before you turn on any refusal, decide what the application does when a request is refused: a manual procedure, a handoff to a person, or a retry.
Bring AI incidents into your existing response process.
Alerts carry the workflow, what triggered them and the spend so far, so the owning team has somewhere to start.
Which AI activity does SAVI see?
Calls your applications make through the SAVI SDK for OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, Google Vertex, Cohere and Mistral, or send by OpenTelemetry, plus usage pulled from OpenAI, Anthropic, Azure and Bedrock accounts. Connectors for Microsoft 365, Google Workspace, Entra ID, Okta, GitHub Copilot seats and Claude Enterprise list the AI tools people use. Systems that do not call AI through these routes are outside what SAVI sees.
What happens to customer and asset data in prompts?
The SDK scans prompts inside your application and sends SAVI the data types and counts it finds, never the text. Your provider receives the prompt unchanged: detection does not redact it. Prompt text reaches SAVI only if you turn on content capture, which is off by default and stored encrypted.
Can a budget stop a workflow?
Only if you choose to. Pools alert by default. Blocking needs 7 days in shadow mode first, an administrator and a reason, and it works through the pre-call check, which refuses the run's next call at the budget. Calls already in flight can take spend past it.
What does the pre-call check do if SAVI cannot be reached?
It lets calls through by default, so an outage at SAVI does not stop your workflow. You can set it to refuse instead. Applications that do not run the check are never refused by SAVI.
Where do alerts go?
Slack, email, PagerDuty, a signed webhook, or your SIEM's HTTP event endpoint in CEF format. Loop alerts carry the agent, the reason, spend so far and a 24-hour projection.
Start with one AI workflow.
Bring a customer-service, field-support or engineering use case. We'll walk through what SAVI can capture, how its spend is attributed, and which alerts or controls fit your application.