You call a company, explain your issue to the bot, get transferred, and explain it again. Then you message them on WhatsApp the next day and start from zero.
The fix isn't a smarter model. It's shared memory: past interactions, decisions and customer context stored in one place, available to every agent and every human with permission to see it. When a customer says "can we pick up where we left off?", the next conversation should already know where that was.
This turned out to be one of the hardest parts of our platform to get right, especially with permissions.
Have you seen a company do this well? Which product made you feel remembered?
Support our launch - https://www.producthunt.com/prod...
@istiakahmad That's a great question, Istiak.
At that scale, quality can't depend on manual reviews alone. AI-powered call evaluation can assess conversations against defined parameters like script adherence, accuracy, compliance, customer experience, and outcomes.
It also flags exceptions for human review, so teams can monitor quality consistently without listening to every call.
congrats on the launch ajay. 10M calls in 4 months is a real number, not a demo. i would be more curious about one case where an ai employee got it wrong and how you caught it, that tells more than the success story
@agshin_rajabov That's a fair question, Agshin.
Here's a scenario we pay close attention to: imagine an AI employee handling payment follow-ups and marking a customer as having paid based on an ambiguous response. That's exactly the kind of situation where it shouldn't assume.
With defined checks and escalation rules, the discrepancy can be flagged and routed to a human for verification before the workflow moves forward. For us, catching these edge cases is just as important as measuring successful interactions.
the payment case is a good example, an ambiguous yes is where people also make mistakes. do the escalation rules come ready or each team writes their own?
@agshin_rajabov It comes with pre-defined rules, however, business can modify and define their own rules and limits as well.
@agshin_rajabov Exactly, Agshin. That’s why we don’t believe escalation should be one-size-fits-all. We provide the flexibility to configure the rules, but each team defines what counts as an exception for its own process. In a payment workflow, for example, an ambiguous response shouldn't automatically be treated as a confirmed commitment. The AI employee can be instructed to seek clarification or escalate to a human instead.
For escalation rules, I’d test calibrated abstention separately from task success. Build a set of deliberately ambiguous leads, policy conflicts, and missing-data cases, then measure whether the employee escalates to the correct owner with the missing evidence clearly stated. Otherwise a high completion rate can reward confident guessing.
@keesan12 That's a really important distinction, Keesan. A high completion rate doesn't mean much if the AI is confidently making the wrong decisions. Testing how well it recognises uncertainty, captures the missing context, and routes issues to the right person is just as important as measuring task completion. Great point!
What really caught my attention was the 10M+ calls for Bajaj Capital within just 4 months. That is not a pilot; that is production load.
@hamza_aafzal Absolutely, Hamza. That's what makes the scale meaningful. Taking an AI solution from a successful pilot to handling millions of real customer interactions requires reliability, consistency, and the ability to perform under production conditions. That's the kind of execution we believe in building at UnleashX.