How Enterpret Classifies Customer Feedback Under the Hood: The 6 Stages
Enterpret classifies customer feedback once, when it arrives, instead of every time someone asks a question. Each record from 50+ sources is normalized, enriched with signals like sentiment, and classified against a product-specific Adaptive Taxonomy with five levels: three Keyword levels for the product area (L1, L2, L3), then a Theme and Sub-theme for what the customer is actually saying. Each record also gets one of four intent Categories: Help, Improvement, Complaint or Praise. Those record-level predictions are stored and linked to accounts, users and opportunities in the Customer Context Graph, and every dashboard, Wisdom answer, agent and MCP query reads from that same stored layer. That is why the same question returns the same counts.
This guide is for builders who want to check the architecture before they buy. It describes what Enterpret documents publicly, and says plainly where something is not documented.
Why classify at ingestion instead of at question time
The common alternative is to send raw feedback to an LLM each time someone asks a question and let it group the records on the fly. That works for a one-off read. It breaks when people compare answers. In Enterpret's study of run-to-run instability across eight frontier models, a raw model changed more than half of its theme set between two runs of the same question at 100 records (theme churn 0.552), rising to 0.726 at 5,000 records. An agent grounded in a persistent taxonomy stayed at 0.013, 0.148 and 0.033 at the same sizes.
86 to 88% Reduction in theme churn when the same model (Claude Opus 4.8, 1,000 records) queried a persistent taxonomy instead of the raw feedback: 0.104 vs 0.765. Source: Enterpret Research, "Same Feedback, Different Answer" (arXiv, 2026)
The paper measures stability, not accuracy: a system can repeat the same wrong answer. Stability is still the property that lets a team compare this week to last week, which is why Enterpret does the classification work up front.
The 6 stages under the hood
1. Ingest and normalize feedback from every source
Enterpret connects to 50+ places customers leave feedback: support tickets, sales and support calls, surveys, app store and review sites, community platforms and CRM data. Records arrive with different shapes, so Unified Fields standardize attributes such as rating, plan or country across sources. Without this step, "filter to enterprise accounts" means something different in every tool.
2. Enrich each record
Enrichment adds signals the source system does not have. Built-in enrichments include sentiment, Sentiment Shift, Total Response Time, Country Code Standardization and Timestamp Conversion, and custom enrichments can be added. These become filters and breakdowns later, so they are computed once rather than re-derived per question.
3. Build a product-specific taxonomy before classifying
The taxonomy is built from your business before any record is classified. Inputs include existing tracked keywords, your help center, changelog and documentation, your business context and historical feedback, with a guided calibration during onboarding. The result has five levels:
| Level | Describes | Example |
|---|---|---|
| L1 Keyword | Product area or feature group | Billing |
| L2 Keyword | Feature or feature aspect | Invoices |
| L3 Keyword | Sub-feature or user action | Download invoice |
| Theme | What the customer is saying about it | Invoice is missing tax details |
| Sub-theme | Detail within the theme | VAT number not shown |
| Category (attribute) | Intent: Help, Improvement, Complaint or Praise | Complaint |
Two design details matter for builders. Themes can map to more than one product path, because the same problem often shows up in several places. And "Not Specified" and "Misc" nodes catch feedback that does not fit an existing child, so gaps stay visible instead of being forced into the nearest label.
Structure is where generated taxonomies usually fail. Enterpret's taxonomy evaluation paper found that AI-generated taxonomies passed every generic quality check, yet 97.7 to 100% of their leaf names restated a parent, and 34.7% to 76.4% of records leaked across top-level product areas, compared with about 20% in production references built with structural checks. That in-house approach with structural checks is the kind behind Adaptive Taxonomy.
4. Classify every record against the taxonomy
Each record gets predictions for its Keyword path or paths, its Themes and Sub-themes, and its Category. You can open any record and see the L1 to L3 path, the related Themes and Sub-themes and the Category, and correct them. Enterpret's own documentation notes that classification can be incomplete or incorrect, so predictions are meant to be checked. Edits are recorded in an audit trail and used to improve future classifications.
5. Keep the taxonomy current
Products change, so the taxonomy has to. Where configured, new Themes and Sub-themes are added as new patterns appear, and Enterpret flags drift, duplicate nodes and emerging terms with proposed changes. Editors can create, edit, move, merge and archive nodes, and prepared changes do not touch the live taxonomy until they are published. For the concept in more depth, see what an adaptive taxonomy is.
6. Link classified records to customers in the Customer Context Graph
Classified records are linked to accounts, users, opportunities, products and custom objects. That is what turns "500 complaints about invoices" into "which accounts, how much revenue, which open deals." In Enterpret research comparing Agentic RAG, a Deep Research Agent and a context graph agent on 30 real voice-of-customer questions, the agent grounded in a persistent customer context graph scored 0.961 overall, against 0.710 and 0.651. The paper calls the result exploratory and tested architectures, not products.
How the classified layer gets queried
Everything downstream reads the same stored predictions, which is the point of the design:
- Quantify, Trends and Dashboards break feedback down by L1 to L3 Keywords or by Themes and Sub-themes over time.
- Anomaly Detection and agents (Quality Monitor, Escalation and Newsfeed) watch for spikes, high-risk feedback from strategic accounts and team-relevant trends.
- Wisdom answers plain-English questions with citations back to records.
- The MCP server gives Claude, Claude Code, Cursor, ChatGPT and other clients tools such as
resolve_taxonomy_scope, which finds the taxonomy nodes for a count or breakdown,run_graph_query, which answers with citation links, andfind_user_quote, which returns verbatim quotes. - Export API syncs classified data to a warehouse or other systems.
To see how an answer traces back to the record behind it, read how to trace an AI-generated insight back to the customer who said it.
What Enterpret does not publish
Enterpret describes custom AI models that learn your product terms and acronyms, but it does not publicly document which underlying models do the classification, and it does not publish classification accuracy percentages. Its three research papers measure stability, structure and answer quality, not per-label accuracy on your data. The honest way to check accuracy is to test it on your own feedback; these five tests for a vendor demo work for Enterpret and its competitors alike.
Example: following one ticket through
Picture a support ticket: "Since the update I can't find our VAT number on invoices." It arrives from the help desk with the customer's plan and country, which Unified Fields map to the same names used for survey and call data. Enrichment marks it negative. Classification places it under Billing, Invoices, Download invoice, with the Theme "Invoice is missing tax details" and Category Complaint. The Context Graph links it to the account and its open renewal. When twelve more arrive the same week, the theme shows the jump in Quantify, an anomaly alert fires, and a PM asking Claude through the MCP server gets the same count the dashboard shows, with quotes.
What changes next
As more questions come from agents rather than people, the classified layer matters more than the interface. An agent that asks a hundred questions a day needs the same answer to the same question every time, and that only comes from deciding what each record means once and storing it.
FAQ
Does Enterpret use an LLM to classify feedback?
Enterpret uses AI models, including custom models that learn your product's terms, to classify each record against a product-specific taxonomy at ingestion. It does not publicly document which underlying models it uses. The classification is stored, so later questions read predictions instead of re-running an LLM over raw text.
Is the taxonomy the same for every Enterpret customer?
No. The L1 to L3 Keywords, Themes and Sub-themes are built from each company's product, documentation and feedback. The four Categories (Help, Improvement, Complaint, Praise) are the same everywhere.
Can I correct how Enterpret classified a piece of feedback?
Yes. Record-level predictions can be edited, taxonomy nodes can be created, edited, moved, merged and archived, and changes are recorded in an audit trail. Prepared taxonomy changes only affect live data once published.
Why do Enterpret counts stay the same between questions?
Because every question reads the same stored predictions. In Enterpret's stability study, once a theme appeared in two grounded runs its reported volume was identical every time, with zero volume disagreement across 117 valid run pairs, while raw models' counts moved 7.5 to 9.8% on matched themes.
If you are vetting customer intelligence platforms, see how Enterpret works on your own feedback, and bring a few tickets you already know the right answer for.
Research cited
- Bagal, V., Ganta, R., & Chellingi, P. (2026). Same Feedback, Different Answer: Measuring Run-to-Run Instability in Frontier-Model Customer Feedback Analysis. arXiv:2610.08036.
- Chellingi, P., G, R., & Bagal, V. (2026). From Plausible Hierarchies to Useful Taxonomies: Evaluating Agentic Harnesses on Customer Feedback. arXiv:2610.09377.
- G, R., Bagal, V., & Chellingi, P. (2026). From Retrieval to Customer Context: Evaluating Frontier-Model Systems for Voice-of-Customer Analysis. arXiv:2610.09375.
Heading
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.



