Glossary
›
Interaction Mining

Interaction Mining

Glossary
Also known as:
Customer Interaction Mining

What is Interaction Mining?

Interaction Mining is a method for building AI agents from an organization's own customer interactions: recorded calls, chats, emails and supporting documents are analyzed at scale to work out how top performers handle each kind of conversation, and the result is compiled into a structured playbook that an AI agent follows in live conversations.

The name marks a difference from ordinary analysis. Analytics tools index what was said so people can search and report on it. Mining captures how a conversation was conducted: the order of the questions, the points where a choice was made, and the response that settled each objection. A recorded renewal call in an archive is a record. Once mined, it becomes a set of decisions an agent can make on the next call.

How Interaction Mining works: five stages

The method moves from raw history to a working agent in five stages.

StageWhat goes inWhat comes out
IngestCall recordings, chat and email threads, and the product sheets, policies and scripts those conversations rely onOne body of material per conversation type, with the result of each conversation attached
AnonymizeThe collected materialText with customer identifiers and payment details removed before any analysis begins
ExtractConversations ranked by their resultsThe structure of the ones that succeeded: question order, decision points, objections and the responses that worked
CompileThat structure, plus the organization's own rulesA playbook an agent can run: conversation states, permitted moves, required statements and escalation points
LearnResults of the conversations the agent conductsUpdated rankings of which move works in which situation, released as reviewed versions

The third stage carries the most weight. What experienced people say they do and what the recordings show them doing rarely match, and the differences that matter are small: a qualifying question moved earlier, an objection acknowledged before it is answered. Details like these surface only when the conversations themselves are the source.

Why Interaction Mining matters

In most teams that talk to customers, a few people produce a large share of the results. Their skill lives in their own conversations and is available only when they are. Everyone else, and every channel with no person on it, works from a script.

Research suggests that skill of this kind can be captured from records of work and passed on. In a study of roughly 5,000 customer support agents summarized by the National Bureau of Economic Research, an AI tool that based its suggestions on the working style of the company's most productive agents raised the number of issues resolved per hour by 13.8 percent. An agent with two months' tenure who used the tool performed as well as one with six months' tenure who worked without it.

Most organizations do not need to collect anything new to begin. Years of recorded calls and message threads already exist for quality and regulatory reasons, and few have been studied for how the conversation was won.

Where Interaction Mining is used

The method applies wherever conversations are recorded in volume and results differ from one person to the next.

  • Sales and applications. Insurance quotes, loan and mortgage applications, telecom plan sales, hotel group bookings.
  • Renewals and retention. Conversations with customers whose term is ending or who have said they may leave.
  • Collections. Calls where the order of questions and the options offered decide whether a payment promise is kept.
  • Onboarding. Verification and setup, where a few specialists lose far fewer new customers than the rest.
  • New channels. Carrying what works on the phone into chat, web forms and phone menus.

Simpler approaches are sometimes the right choice. A team with little recorded history has nothing to mine and is better served by a carefully written flow. Questions with one correct answer, such as opening hours or order status, need a knowledge base. And where results barely vary between people, there is no top-performer method to find.

What good Interaction Mining looks like

  • Sources ranked by outcome. The conversations studied are chosen for what they produced, not for who held them or how long they ran.
  • Structure as the output. The deliverable is a set of states, decisions and rules an agent can run, not a summary report.
  • Anonymization first. Identifiers are stripped before analysis, and raw transcripts never travel into the agent.
  • Documents included. Policies, product terms and scripts are mined with the conversations, so the playbook reflects the rules as well as the practice.
  • Review before release. Business owners and compliance approve the playbook before it reaches a customer. A habit that converted well but breaks a rule is taken out.
  • Traceable decisions. Each move the agent makes can be tied to the playbook step that allowed it, and regulated statements are delivered exactly as approved.

How to measure Interaction Mining

Judge the method by the agent it produces, against a baseline recorded before launch.

  • Outcome against baseline. Completed applications, renewals or kept payments per conversation, compared with the same conversation type handled the old way.
  • Distance from the top. How close the agent's results come to those of the people it was built from.
  • Playbook coverage. The share of live turns the playbook anticipated, as opposed to turns that ended in a handoff.
  • Time to first live conversation. The elapsed time from handing over the recordings to an agent taking real conversations.
  • Improvement by version. Whether each playbook release beats the previous one on similar traffic.

Rules that apply in the United States

Interaction Mining reuses records made for another purpose, and the agent built from them speaks for the organization. The US position is outlined here; other countries regulate recording and data reuse differently, and regulated industries add more.

  • Consent to record. Federal law allows a call to be recorded when one party agrees. Some states, California and Florida among them, want every party to agree, so how the original recordings were announced decides whether they can be used.
  • Purpose of use. State privacy laws such as California's limit using personal information for purposes customers were not told about, so the privacy notice should cover building AI systems from recordings.
  • Deletion risk. The Federal Trade Commission has required companies to delete models and algorithms built from data they had no right to use.
  • Payment, health and financial data. Card security codes may not be stored under payment card industry standards. Health information held by providers and insurers falls under HIPAA, and financial institutions' customer records fall under the Gramm-Leach-Bliley Act.
  • Employee notice. The other voice on each recording belongs to an employee. Some states, New York among them, require written notice to staff that calls and messages are monitored.

Frequently asked questions

How is Interaction Mining different from conversation analytics?

Conversation analytics reports on what happened: topics, sentiment, quality scores. Interaction Mining produces something that acts: a playbook an AI agent runs in its next conversation.

How is Interaction Mining different from prompts or retrieval?

A prompt describes how someone thinks a conversation should go. Retrieval supplies documents to answer from. Interaction Mining starts from how successful conversations actually went and turns that into the sequence and decisions an agent follows.

What data does Interaction Mining need?

Recorded conversations of the type the agent will handle, the result of each one, and the documents that govern them, such as policies and product terms. Many imperfect recordings are more useful than a few ideal ones.

How is a top performer identified?

By what their conversations produced, measured across enough of them to rule out luck and corrected for how promising their leads or accounts were. Seniority and job title are poor guides.

Does the agent keep or repeat raw transcripts?

It should not. In a sound implementation, identifiers are removed before analysis, and the agent runs the compiled playbook, not the conversations it was derived from.

Your top performers already wrote the playbook.

Interaction Mining is Encore AI's own method, backed by two granted patents.

Scale your best to be the best.

Book a Demo