A customer writes in: “My parcel never arrived, can you send another?” Your assistant searches everything it has seen and finds a near-perfect match from last week: “My parcel arrived damaged, can you send another?” It sends last week’s answer, a replacement, no questions asked. The words were almost the same. The case was not, because a missing parcel starts with a trace at the courier.
The search did its job. It found the closest match, which is what look-up does. RAG finds passages that look like the question, vector databases find what sits nearby, knowledge graphs find what someone linked, caches return what was asked before and memory recalls what was saved. Every one of them finds something that already exists. But similar is not the same as right, and the closer a wrong match looks, the more confidently it is served.
Learning is a different act. It understands the pattern behind your work, compresses it into what generalises, and predicts what comes next. Duplex learns, and your look-up tools become its tools: it calls each one when a step needs it, and decides what the answer should be.
Asked the same way, look-up finds the stored answer, and Duplex knows it too.
Learning carries the pattern to new cases.
The next request rarely matches the last one word for word. Someone phrases it differently, leaves out a detail or adds context that changes nothing. A cache returns what was asked before, so it needs the new request to look like an old one. A search by similarity copes better with new wording, because it finds what sits nearby, but nearby is a measure of the words, not of the case.
Learning works on the case itself. When your team confirms a result, Duplex learns the pattern behind it: why it worked, and where it applies. So it knows the same request in new words, and it knows when a request that sounds the same is asking for something else.
Telling look-alikes apart is where the difference shows most. A lost parcel and a damaged one share almost every word, and so do an order cancelled before dispatch and one cancelled after it. A tool that serves what looks similar answers each pair the same way, with the same confidence. Duplex sees the difference, and brings your AI in, briefed with what it knows about the case.
In a test that ranged from exact repeats to requests from another domain, Duplex answered 7.5% of misleading look-alikes wrongly, against 45.3% for a cache. Across all the requests, 98.98% were handled correctly: answered right, or handed to your AI.
Your look-up tools become its tools.
Look-up tools are good at what they do, and Duplex uses them for exactly that. Retrieval is the right way to find a clause in a policy, a knowledge graph the right way to see how a supplier links to everything else, and a cache the fastest way to return a decision already made. Those facts belong in your systems, current and owned by the people who look after them.
What the tools can’t do is decide. They find, and leave the rest to whoever asked: which source to trust, what to do when two of them disagree, what the answer means for this case. Duplex brings that judgement. It calls each tool when a step needs it, learns which source answers which kind of question, and decides with the current facts in front of it.
A single invoice can need all three: the policy, the supplier’s status and the last decision about that supplier. Most days they agree. The day the supplier goes on hold, the last decision is out of date, and the answer has to follow the newest fact and show where each piece came from.
That is a fair division of labour, and everything you have built stays in service. Your tools go on finding what they were built to find, each more useful now that something decides when to use it. Duplex learns, decides and acts, with every answer traced back to its sources.
To answer one invoice, Duplex needs a policy, the supplier’s status and the last decision.
A learned pattern tells you what comes next.
Look-up is about the past: it finds what happened. Learning reaches forward. Once Duplex knows how a kind of work unfolds, it can predict what an action will change, plan the steps after it and decide whether to answer, hand over or decline.
Support tickets follow a pattern: opened, triaged, fixed, confirmed, closed. Look-up can find a ticket that looks like the one in front of you. Duplex has learned from the confirmed ones, so it knows what this ticket needs next and what should follow. When a ticket takes a turn no earlier one took, Duplex hands that step to your AI, and once your team confirms how it ended, the turn becomes part of the pattern.
On World of Workflows, a public enterprise benchmark built on ServiceNow, Duplex predicted how the system changes after an operation better than GPT-5.1, Gemini 3 Pro and Claude Sonnet 4.5, with no model calls. In planning, it found the shortest route every time without a single model token.
A library tells you what has been written. A colleague who has learned the work knows what the case in front of them means, notices when it is new and gets on with the next step. Your look-up tools are the library, as useful as they have ever been. Duplex is the colleague, and with every result it knows the work a little better.
Confirmed tickets teach Duplex how this kind of work unfolds, one step after another.