The longer it runs, the less it asks.
Step 1 of 4: The start
90% fewer model calls, and more accurate.
3,000 customer-operations decisions: refunds, account unlocks, credit limits, payouts. Each has one right action out of about nineteen, or a refusal, and five times along the way the company changed a rule. Every approach used GPT-5.5; Duplex learned as it went.
- 94.6%Accuracy, the highest in the testThe model alone 59.8% · with a cache 77.3%
- −90.3%Model calls−84.8% cost · −81.9% latency
- ~23×Faster once learned≈97 ms against ≈2,231 ms
3,000 decisions, and where each one went.
Answered by Duplex itself, passed to GPT-5.5, or refused.
Step 1 of 4: Arrive
3,000 decisions arrive: refunds, unlocks, credit limits and payouts, in more than 2,000 phrasings.
After a rule changed, who gave the new answer.
Of every 100 decisions that a rule change had touched.
Step 1 of 3: The model alone
GPT-5.5 cannot know a company’s private rule change: it gave the old answer 97% of the time.
The longer the stream, the better the numbers.
The same work on a 300-decision stream and on a 3,000-decision stream.
Step 1 of 4: Accuracy
Duplex was 85.0% accurate on a 300-decision stream, and 94.6% accurate on 3,000.
The bill, and the wait.
GPT-5.5 alone against GPT-5.5 with Duplex, on the same 3,000 decisions.
Step 1 of 3: Per decision
Every 1,000 decisions cost $0.13 with Duplex, and $0.82 with GPT-5.5 alone.
How it was run.
- The work
- Customer-operations decisions: refunds, account unlocks, credit limits, payouts, access and fraud escalation. One right action out of about nineteen, or a refusal.
- The stream
- 3,000 decisions in more than 2,000 phrasings, 68% of them worded in a way not seen before, with new kinds of request arriving midway.
- Rule changes
- Five in-house rule changes, each flipping the right action for a kind of request; 1,073 decisions came after a change.
- Feedback
- The right answer was revealed for about one decision in ten, the same ones for every approach.
- Models
- GPT-5.5, with real API calls, for every approach.
- Compared with
- GPT-5.5 alone, and GPT-5.5 with a semantic cache at its best setting; on the 300-decision stream, also a cache with retrieval and a knowledge graph.
- Kept equal
- The same model, feedback and refusal rules for every approach.
What it means for your work.
Your bill shrinks as it learns.
Duplex answered 86% of the decisions itself: 90% fewer model calls and 85% lower cost.
A rule change reaches every answer.
After the rules changed, Duplex gave the new answer 92.5% of the time; GPT-5.5 alone, 2.3%.
It gets better with volume.
From 300 decisions to 3,000, accuracy rose from 85.0% to 94.6% and the cost per decision fell by two thirds.
See it on your own work.
From your own assistant to every agent your company runs, a pilot puts Duplex on your own work, with the measures agreed before it begins.