Skip to main content

Optimization

Find room to lower AI costs.

Keep useful results at the center of cost decisions.

Compare spend, retries, and rework to find changes that may help. Savings depend on your workflow.

Join the waitlistSee the cost record

AgentOps can turn your own graded runs into a public benchmark. An organization admin previews the summary before publishing. Signed reports show sample sizes, uncertainty, work types, and weekly trends, while small groups and private source details stay out. Sharing is off by default and can be revoked.

The invoice stops before the work does.

Model spend is only one input. The accepted result also carries context preparation, tool use, retries, review, waiting, coordination, and rework. Cut one line without seeing the rest and the total can rise.

growth in weekly ChatGPT Enterprise messages over a year
8×
growth in average OpenAI reasoning-token use per organization over a year
320×
OpenAI, 2025

Ship the governed customer import

Accepted

Completion cost by accepted run

Indexed cost per accepted run · run 1 = 100

0255075100
−45% vs Run 1

Run

Optimization events

  1. Run 3 · −8Less preparation in this example
  2. Run 6 · −9Lower model spend in this example
  3. Run 8 · −6Less rework in this example

Changes in the illustrative cost breakdown

Run 1Run 10

Model
30 → −15
Context preparation
18 → −11
Retries
12 → −8
Rework
8 → −4
Coordination
10 → −3
Tools
10 → −2
Review time
12 → −2
Total
100 → 55−45

Fictional example — the 100-to-55 curve is not measured savings or a customer result. Measure your own workflow to establish a baseline.

Completion cost values: Run 1: 100; Run 2: 94; Run 3: 86; Run 4: 82; Run 5: 78; Run 6: 69; Run 7: 66; Run 8: 60; Run 9: 58; Run 10: 55. Optimization events: Run 3: Less preparation in this example; Run 6: Lower model spend in this example; Run 8: Less rework in this example. Illustrative reduction by component: Model 30 → 15; Context preparation 18 → 7; Retries 12 → 4; Rework 8 → 4; Coordination 10 → 7; Tools 10 → 8; Review time 12 → 10.

Review cost alongside the result it buys.

Usage reports and completed-work records answer different questions. Use both to judge whether a change is worth making.

Usage view

What did the model calls cost?

  • A usage total alone does not tell you whether the team could use the result.

Completed-work view

What did it take to finish useful work?

  • Compare recorded cost, review effort, and the result against the same requirements.

A lower price is useful when the work still meets your needs.

One task. See where the $38 went.

A sample refund task makes the cost visible: the work accepted, the context sent again, and the effort spent on retries and rework.

Illustrative example

Task 4821 · Reduce refund cycle time

AI cost in USD

Cost falls across these three sample runs. All three pass the same checks; future results can differ.

Tasks per month1,200

At your example volume

Run 1

$45,600

Illustrative AI spend in USD; human time is excluded. These amounts and quality results are sample assumptions, not measured customer results or promised savings.

Context rebuilt at three handoffs. 2 retries · 8 premium calls

Accepted model output
$8.40 · 22%
Context sent again
$20.50
Retries and failed runs
$4.60
Repeated actions and rework
$2.70
Automated checks
$1.20
Tool calls
$0.60
AI cost to completion
$38.00

A small sample does not prove future quality. Review live results and keep a way to reverse the change.

Five opportunities to explore.

These are places to look for improvement, not savings forecasts. Test changes against your own work and quality requirements.

  1. Context: spend less effort repeating the brief.

    Before

    Context rebuilt at every handoff.

    • Request, agent run, review: each one missing pieces.

    Opportunity with Aestus

    Keep useful decisions with the work.

    • Start with the shared brief, past decisions, and their sources. Add detail when it is needed.

    Potential value: less repeated explanation and preparation.

  2. Model choice: check what the task needs.

    Before

    One expensive choice for different kinds of work.

    • Routine steps may use more costly resources than they need.

    Opportunity with Aestus

    Compare alternatives against the same requirements.

    • Review cost, time, and quality before choosing a different model, agent, tool, or person.

    Potential value: a less costly choice where the results support it.

  3. Learning: use past work to guide the next attempt.

    Before

    The same mistake needs another round of work.

    • Past decisions and feedback are hard to find when the next task begins.

    Opportunity with Aestus

    Make useful lessons available for review.

    • Use earlier results and feedback to choose a change. Check whether it helps before using it more widely.

    Potential value: fewer repeated mistakes and less rework.

  4. Early review: find a problem before more work depends on it.

    Before

    A missing approval is found at publication.

    • The team has already paid for the draft, review, and follow-up work.

    Opportunity with Aestus

    Review approval needs before the workflow runs.

    • Illustrative example: identifying a missing approval avoids two agent runs and three later steps.

    Potential value: less work spent on a path that cannot proceed.

    For scale, 100 agents taking 2,000 actions a day would produce about 73M actions a year. This is an illustrative activity estimate, not a savings forecast. Earlier reviews do not replace controls during live work.

  5. Coordination: one path to acceptance.

    Before

    Which version is accepted? Who reviewed?

    • Salesforce, Notion, Slack, GitHub: each holds a piece.

    Opportunity with Aestus

    One record holds the accepted version and its reviewer.

    • Supported connections bring updates and results back to the task.

    Potential value: fewer follow-ups to find the current result and reviewer.

Consider the change. Check the result. Keep control.

  • Learn from previous results

    Look at recorded costs, time, reviews, and failures to see where another approach may help.

  • Assess proposed improvements

    Review the proposed change and its supporting results. Check that the comparison is relevant to your own work.

  • Choose what to use more widely

    Set review requirements and limits. Keep a way to stop or reverse a change when the results do not meet your needs.

Review a model change before you apply it.

The refund summary step has a lower-cost option. Check the comparison and keep the choice with your team.

Illustrative example

Cost per summary

Illustrative example

$2.40

Passed the same review

Illustrative example

20 / 20

20 accepted summaries per model

Sample check: 20 of 20 summaries passed the same review criteria in each group.

Example uses the premium model.

A small sample does not prove future quality. Review live results and keep a way to reverse the change. Try the example. Changes stay in this demo.

Measure the cost of a useful, completed workflow.

Model usage is part of the cost. A completed result adds another question: what did it take to produce work your team accepted? Include recorded costs and clearly labeled estimates for effort that is not captured.

  • Input efficiency

    Per call: model usage and other recorded inputs.

  • Workflow efficiency

    Per run: steps, retries, review, waiting, and rework.

  • Completion efficiency

    Per accepted result: cost and effort assessed alongside quality.

What is modeled, and what is measured.

The charts are illustrative, not measured customer savings or forecasts. Your reviewers define what acceptable work means. Compare recorded usage with clearly labeled estimates for time and effort. Results can vary, and costs may rise as work changes.

We have a model gateway and a cost dashboard.

Keep them. Aestus adds the rest of the cost record: tools, retries, time, review, coordination, rework, and whether the final result was accepted.

Measure one workflow's baseline.

Talk to us