When a company shows you a chart with itself on top, the first thing to ask is who wrote the test. A test you design yourself tends to measure what you are good at, however careful you are. You choose the tasks, the data and the scoring, and from the outside nobody can tell how much of the result is the product and how much is the test.
The results worth the most are the ones on someone else’s ground: a test built by people with no stake in your product, answered on the same tasks as everyone else and scored by the people who wrote it. You don’t choose the questions, and you can’t move the goalposts. It is the difference between winning at home and winning away.
World of Workflows is that kind of ground. Its authors built it on ServiceNow and scored it with their own evaluator. Duplex went in against three frontier models and came out ahead at predicting how the system changes, without a single model call.
World of Workflows: a public benchmark on ServiceNow, scored by its own authors.
Someone else’s test, someone else’s scoring.
The benchmark puts AI inside the kind of system where enterprise work happens: incidents, requests and approvals, each with its own rules about what follows. An agent working in ServiceNow has to understand what an operation will do as well as what the request asks for. One of the things the benchmark tests is whether you can predict how the system changes after an operation.
Every task came from the benchmark’s authors, working independently of Vareon. Duplex and three frontier models answered the same ones, on three offline tracks, three runs each, and the authors’ own evaluator scored every run. The company being tested chose neither the questions nor the marking.
That is what makes the result easy to trust and easy to repeat. Anyone can take the same benchmark, run it the same way and score it with the same evaluator, and get a number that means the same thing for everyone who Sat the test.
And the field was strong. GPT-5.1, Gemini 3 Pro and Claude Sonnet 4.5 are frontier models, each among the most capable general models of its day, bringing broad knowledge and careful reasoning to any task they are given. On ground none of them chose, the question was the same for all four: what will this operation change?
The benchmark’s authors built the tasks on ServiceNow, independently of Vareon.
The best at predicting what changes.
Predicting that change is harder than it sounds. Depending on how a team has set the system up, resolving an incident might close related tasks or notify a whole group. Those effects live in the system’s own rules rather than in the words of the request, so a model reading the request has to reason its way to them from general knowledge of systems like it.
Duplex predicted it best. It scored 0.256, against 0.205 for Gemini 3 Pro, 0.204 for Claude Sonnet 4.5 and 0.166 for GPT-5.1, and the lead is statistically significant against all three, on the same tasks.
It did this with no model calls. Duplex made 570 predictions at zero model cost, while 84 predictions on a frontier model cost $13.40. That gap is small on a benchmark and large in production, where an agent asks the same question before every operation, all day.
The difference is in what each one brings. A frontier model is general intelligence, learned slowly, like the neocortex: it knows the world. Duplex is runtime intelligence, fast, like the hippocampus: it learns how this system behaves from what happens in it, so predicting the next change is something it knows rather than something it has to work out.
Your agents can look before they act.
An agent that works in your systems changes real things. It closes incidents, reassigns requests and tells people what happened. The costly mistakes are the ones it didn’t see coming: an incident resolved without the closing note it needs, or a caller told before anything was fixed. By the time anyone notices, the change has already happened.
Duplex learns how your own systems behave: what each action changes and what comes next. So before your agent runs an operation, Duplex can tell it what will change, the status, the note and who is told, and the agent checks that first, the way a careful colleague looks before pressing the button. If the prediction shows the caller being told before the fix is confirmed, the agent can stop and ask.
Prediction isn’t the whole job, and the benchmark shows where the line is. On exact action recovery, which means naming the precise operation with every detail right, Duplex tied GPT-5.1 and Gemini 3 Pro. That is work where a model’s reasoning earns its place, so Duplex answers what it has learned and your AI takes the rest. Every result your team confirms moves a little more of that work to Duplex.
A result on someone else’s ground is one you can plan around. On a test it didn’t write, scored by the test’s own authors, Duplex saw what each operation would change better than three frontier models, for nothing in model costs. With Duplex learning from your own systems, your agents get the same: every operation checked against what it will change, before it runs.
Your agent is about to run an operation: resolve an incident in your system.