On her first day, a new hire on your support team doesn’t start from nothing. She gets the returns policy, a folder of tickets the team handled well, the exceptions everyone knows by heart, and a colleague who tells her which cases go to a manager. She is useful in her first week because the team hands over what it already knows.
Most AI gets no such welcome. A new assistant starts with none of it, so your team explains the same things again, one correction at a time, and the corrections rarely outlast the session. The knowledge was there all along, in reviewed tickets, approval threads, policy documents and runbooks, and the AI sees it only when someone pastes it in or a search happens to find it.
Duplex can start the way a good new hire does. It learns what your team already trusts before it serves, so it is useful from the first request, and it goes on learning from every result your team confirms.
Start with what your team already knows.
Start with what you already know.
Most teams already have the material, scattered across the tools they use. There are reviewed examples and decisions they stand behind, policies and playbooks, tool runs that worked, and methods they trust: the way a refund is approved, the route a billing dispute takes, the checks before a contract goes out.
Today that material reaches AI in a few familiar ways, and each has its use. A prompt is quick, but it has to travel with every request, and it grows until nobody dares edit it. A document search finds the passage that looks most like the question, which helps when someone knows what to ask. A fine-tune changes the model itself, at the price of a training project that is fixed the day it ends.
Pretraining Duplex is different. It learns from the material before its first request, the same way it learns from live work: it understands the decision behind each example, where it applies and where it doesn’t. Your team places each piece in the work area it belongs to, so the support playbook guides support replies and never the finance team’s approvals.
And Duplex knows where each piece came from. An answer it gives on day one comes with the examples behind it, and each example leads back to the decision or document it rests on, so your team can check the answer before trusting it.
Useful from the first request, current from then on.
On day one, Duplex takes on the work it was pretrained on from the first request: the returns question that comes in every morning, the refund that needs a manager’s approval, the dispute that belongs with billing. It settles those in milliseconds, with no model call, and everything else goes to your AI with what Duplex knows about the case attached. Your team stops explaining the basics, and your AI spends its time on what is new.
Pretraining is a head start, and Duplex is useful without it. Started empty, it learns from live work alone, one confirmed result at a time, which suits a team with little material to begin from. Pretrained, it begins with what your team already knew, instead of discovering it result by result.
A head start is only the start. Policies change, tools change and people find better methods, so a snapshot of what your team knew on day one goes out of date with the first policy change. For an AI model, pretraining is where learning ends: it learns the world, is released and stays as it is. For Duplex, pretraining is where learning begins.
When a policy changes, Duplex follows the new one from the next request, and everything else it learned, pretrained or live, stays as it was. The head start stays a head start, instead of becoming something your team has to work around.
Started empty, Duplex learns from live work alone, one confirmed result at a time.
Clear beats big.
The temptation is to pour in everything: every ticket, every thread, every version of every policy. More feels safer, and archives are easy to export.
But a large archive mixes clear examples with outdated, conflicting and doubtful ones: the policy from two versions ago, two managers who decided the same case two ways, the reply that was sent and never checked. Learn from all of it, and the doubt comes along with the knowledge.
A small set of clear, current examples teaches more than a large archive of uncertain ones. A clear example is one your team would give again today, for a reason it could explain. Your team reviews what goes in and chooses the decisions it stands behind. Include the cases that should go to a person, and Duplex learns to hand those over instead of guessing.
Day one isn’t final, either. If an example turns out to be wrong, or the policy behind it changes, your team can correct or remove that one lesson, and everything else Duplex learned stays as it was.
A good new hire is useful in the first week because of what the team handed over, and good in the first year because of everything learned on the job. Duplex works the same way. Pretraining is the handover, and every result your team confirms after that is learning on the job.
A large archive mixes clear examples with outdated, conflicting and doubtful ones.