Sanaf, the AI employee

An AI customer service agent

Support questions arrive faster than anyone can answer them, and the ones that wait are the ones that turn into public reviews. This job works the queue continuously, answers the routine questions from your own material, resolves what it has been trained to resolve, and pulls a person in the moment a customer needs one.

What it does

Concretely

  • Works the support inbox and help desk continuously rather than in bursts.
  • Answers the routine questions from your help documentation, not from guesswork.
  • Resolves the common issues it has been given a procedure for.
  • Keeps every conversation moving instead of letting it go quiet.
  • Escalates anything complex with a summary of what has already been tried.

The damage is done by the slow ones, not the hard ones

A hard question answered on day three is usually forgiven. An easy question answered on day three is what people write reviews about, because the customer can tell it was easy. Most support reputations are decided by the boring half of the queue, which is exactly the half that can be worked continuously without anyone losing anything.

It answers from your material or it does not answer

The rule this job is hired under is that it never invents an answer. If your help documentation does not cover it, it says so and gets a person, rather than producing a confident paragraph that a customer will act on. An automated support agent that guesses is worse than no automated support agent, because the mistakes arrive faster.

Where it stops is the whole design

Refunds, exceptions to your policy, and any customer who is genuinely unhappy go straight to a person. Those are the conversations where being handled by software is itself the complaint, and no amount of accuracy fixes that. What a person receives is the thread, the history and what has already been attempted, so the handover is not a restart.

Questions

Frequently asked

Is this a chatbot on our website?
No. It works inside the support inbox and help desk your team already uses, which means the reply arrives in the same thread by the same route as a colleague’s would. A widget that only exists on one page is a different product with a different failure mode.
How does it know when to get a person?
Three ways: the question is outside what your documentation covers, the request is one you have told it a person must handle such as a refund or a policy exception, or the customer reads as genuinely unhappy. Any of those and it stops, hands over, and tells the customer a person is coming.
What stops it from telling a customer something untrue?
It answers from your help material and says plainly when it has nothing to answer from. That is a narrower job than a general assistant, and the narrowness is the point: the cost of a wrong support answer is paid by a customer who acted on it.