Next best shown: 86.5%
RESULTS & EVIDENCE
One foundation.
Proven on real enterprise work.
From Fortune 500 enterprises to startups, Apollo-1 applies each company’s policies to the work it needs done. Explore production deployments, complex calculations and decisions with evidence you can inspect.
PUBLIC BENCHMARK / τ²-BENCH AIRLINE
Unparalleled reliability for workflow execution
at a fraction of the cost.
τ²-bench is a public benchmark that tests whether an agent can handle airline service requests with tools while following the airline’s policies. We ran its 50 tasks four times each.
Next best shown: 78%
5.4× to 30× less than frontier models
The gap widens with every repeat
Chance a task is completed correctly in one trial, and in every one of 2, 3 and 4 trials
pass^k, the chance that a task is completed correctly in one trial and in all 2, 3 and 4 trials. Apollo-1: 90.5, 89.3, 88.5, 88. Claude Opus 5.5: 86.5, 82, 79.5, 78. GPT-6.1-sol: 83.5, 79, 76, 74.
- Apollo-1 (AUI)
- Claude Opus 5.5 (xhigh)*
- GPT-6.1-sol (max)*
- Leading public entries: GPT-5.2 (high), Claude Opus 4.5, GLM-5, Gemini 3 Flash, Qwen3.5-397B-A17B
SELECTED DESIGN PARTNERS
Different industries.
Same Apollo-1.
One universal foundation. Each company supplies its own policies, systems and permissions, without domain retraining.
Automotive retail
Travel insurance
Enterprise software
Consumer AI / Language learning
Ecommerce software
moovitAn Intel company
Urban mobility
PRODUCTION / AUTOMOTIVE RETAIL
Customer conversations.
Completed work.
At Sonic Automotive, Apollo-1 searches inventory, books test drives and answers vehicle-specific warranty questions.
The agent also outperforms human agents on warranty lead generation, with an auditable trace of its decisions.
100%
autonomous execution
within the deployed scope
Inventory search · Test-drive booking
Vehicle-specific warranty questions
Vehicle-specific warranty questions
THE WORK
Support customers as they explore vehicles and take the next step.
THE BUSINESS RESULT
Autonomous execution across the deployed tasks, with stronger warranty lead generation than human agents.
THE RECORD
Inspect the facts, rules and actions behind the agent’s decisions.
THE ECONOMICS OF REAL WORK
Build faster. Run for less.
Stay in control.
Measure the effort to build the agent, the inference needed to run it and the work required to maintain its behavior.
Lower inference cost.
Deployment comparisons span three production migrations, comparing the same customers and workloads across runtimes.
Inference cost only. Savings vary with task complexity, model and integration scope.
A complex-rule agent.
The measured build involved a 1,150-page rulebook, turning requirements into an agent that could be exercised against scenarios.
Agent build time. Enterprise integration, validation and production rollout are separate.
Update the policy.
Keep the foundation.
Change the relevant program logic without domain retraining. Run scenarios again and inspect what changed in the calculations, decisions and actions.
EVIDENCE FROM COMPLEX TASKS
Evidence from
complex work.
Evaluations go beyond fluent answers. They examine repeated determinations, calculations, missing facts and the conditions that permit—or prevent—an action.
Consistent decisions across runs.
Repeated-scenario evaluations compare determinations, figures and policy treatment. The test is whether the decision holds when the scenario is run again.
WHAT TO INSPECT
The same established facts, program version and state. Compare outcomes and calculated values across independent runs.
Amounts that reconcile.
Complex-rule evaluations exercise caps, allocations and adjustments. Calculations can be checked independently against the inputs and formulas recorded.
WHAT TO INSPECT
Source values, intermediate calculations and the final total. Follow an amount back to the rule that produced it.
The right decision to stop.
Gated cases route to review when a controlling condition prevents a standard outcome. A correct referral is part of successful execution.
WHAT TO INSPECT
Which condition took precedence, why the action was blocked and what must happen before work can continue.
Decisions grounded in the record.
Evaluation records connect retrieved information, structured facts and service responses to the decision. They distinguish supplied data from values derived during execution.
WHAT TO INSPECT
Where each fact came from, what was calculated and what the connected service actually returned.
Policy gaps made visible.
Working from complex rulebooks surfaces unresolved instructions. Evaluation materials record those gaps and the readings used when an answer depends on them.
WHAT TO INSPECT
The missing instruction, any declared assumption and the question the policy owner needs to resolve.
These are evaluation capabilities, not a single accuracy score. Consistency, correctness and production performance are measured separately.
YOUR WORK / YOUR SUCCESS CRITERIA
Bring the cases your current agent
struggles with.
Test Apollo-1 against your policies, expected outcomes and exceptions. Inspect the decisions, change a rule, and compare build effort and inference cost on the same work.
- Correct outcomes
- Required refusals & referrals
- Repeatability
- Cost per completed task
PUT APOLLO-1 TO WORK