All work
02Enterprise

A 90% CSAT score. And still the escalations kept coming.

Five million support conversations a year. Surveys collected from eight to ten percent of them. A score that looked strong. A business that kept getting escalations from the ninety percent nobody was measuring.

5M

Conversations measured

Not just the 8–10% survey sample

90%+

Model accuracy

At deployment, up from 60% at start

Escalations

Proactive outreach replaced reactive firefighting

The situation

The CSAT score was tracking at 89 to 90 percent. By any benchmark, that's strong. But the Senior VP of Global Customer Support had a problem he couldn't solve with that number: escalations kept arriving from customers who had never filled in a survey. Nine out of ten conversations — more than four million a year — generated no feedback at all. The business had no idea how they went.

The insight

The transition from entry state to exit state is the real measure of support quality. Not whether the customer fills in a form — but whether they left the conversation better or worse than they arrived.
What was built
01

A machine learning sentiment model trained on conversation transcripts — voice and chat — that measures the emotional arc of every interaction from beginning to end

02

A training approach built on the conversations that already had CSAT responses, giving the model ground truth to learn from before being applied to the full five million

03

A personalised agent development layer — using the same conversation-level data to identify which agents were strong at which problem types, and building learning paths targeted to actual gaps

What changed

The support team could identify — in near real time — which interactions had gone poorly and reach out proactively before frustration became a complaint. The overall satisfaction score held and improved directionally. Agent development became personalised by conversation-level performance data rather than generic training curricula.

Advanced Data ScienceML ModelCustomer SupportEnterprise
The MiraDoor take

What made this project work wasn't the machine learning. It was starting from first principles: why does anyone contact support, what state are they in when they do, and what does a good outcome actually look like? The technical solution followed from that thinking.

Does this sound familiar?

Send us the three questions you have been trying to answer.

Get in touch