Recommendations
When an AI Optimizer run finishes, each comparison model is checked against the baseline on quality, cost and latency. The results appear on Gauge AI → AI Optimizer → Recommendations.
A model is only recommended when its average judge score is at least as high as the baseline's. Savings never come from a model that scored lower.
How a model qualifies
The platform compares each comparison model that completed with the baseline, using the evaluation scores and the cost and latency of the same traces.
| Recommendation | Condition |
|---|---|
| Cost-optimized | Average score is greater than or equal to the baseline's, and average cost per trace is lower than the baseline's. |
| Latency-optimized | Average score is greater than or equal to the baseline's, and p95 latency is lower than the baseline's. Not produced for batch runs. |
| Already Optimal | No comparison model met the condition. The baseline stays the best choice for that dimension. |
A model that failed, or that produced no valid cost or score, is never marked as optimized. When several models qualify, the one with the largest cost saving (or the largest latency reduction) is shown as the recommended model.
The baseline's cost and latency come from the cost and latency keys in the agent's trace metadata. If those keys are missing, the baseline cost or latency is zero and no model can beat it. See Before you start.
The Recommendations tab
- Date filter: Last 7 days (default), Current Month, Last Month or Custom Range, based on when each run was last updated.
- Cost Optimization and Latency Optimization tabs, each with summary cards and a table.
- Table columns (cost tab): Agent, Current Model, Recommended Model, Quality Score, Cost Comparison, Projected Monthly Savings, Projected Annual Savings, Jira, Slack, Linear and Job Details. The latency tab shows Latency Comparison and Improvement instead of cost and savings.
- Runs that are still processing show live progress per model.
Click View Details on a row for the recommendation details page, with Cost, Latency and Score analysis and the traces behind it (input, output, original model, original cost and original latency).
How projected savings are calculated
The Recommendations tab uses a fixed volume so that runs of different sizes are comparable:
The 100,000 requests per day stand for 100 users making 1,000 requests each. They are not your actual traffic, so scale the figures to your own volume before you budget with them.
The Analytics tab totals the projection stored with each run, which multiplies the saving per trace by the number of traces in that run and by 30 for a month (365 for a year). Treat both as estimates and use Cost Dashboard figures for actual spend.
Acting on a recommendation
| Column | What it does | Setup |
|---|---|---|
| Jira | Creates a Jira issue with the current model, recommended model, costs, savings and quality score, then shows the issue key and status. | Connect Jira and map the agent to a Jira project and board under Settings → Integrations → Jira. See Jira. |
| Slack | Opens a dialog to post the recommendation to Slack, then links to the thread. | Connect Slack and map the agent to a channel under Settings → Integrations → Slack. See Slack. |
| Linear | Creates a Linear issue for the recommendation. | Connect Linear and map the agent to a Linear team under Settings → Integrations → Linear. See Linear. |
When an integration is not set up for the agent, the column shows Configure, which opens Settings → Integrations.
Switching models is up to you. The optimizer does not change your application; if your traffic goes through Budget AI, you can move it by changing the model behind the model group your application calls.
ServiceNow risk records
When the project's ServiceNow integration has Enable Risk Records turned on, optimizer results also create risk records:
| Risk record | Created when |
|---|---|
| Cost Optimization | The best cost-optimized model saves more than 10% per trace. |
| High Latency | The best latency-optimized model cuts p95 latency by more than 20% (not for batch runs). |
| Error Rate Alert | Every evaluation for a comparison model failed. |
Next steps
- Model Replay: how comparison outputs are produced.
- Evaluation & Credits: how quality is scored.
- Models & Pricing: keep prices current so savings are right.