Intent
An intent is the classification primitive used by Salesforce Einstein Bots and the legacy Einstein Intent Service to recognize what a user is trying to do in a conversation.
Definition
An intent is the classification primitive used by Salesforce Einstein Bots and the legacy Einstein Intent Service to recognize what a user is trying to do in a conversation. Each intent has a name, a description, and a set of training phrases (typically 20 to 100 per intent) that an NLP model learns from. When a user message arrives at a bot, the model computes a confidence score for every defined intent and returns the highest as the picked intent. The dialog or action wired to that intent runs next.
Intent classification is the foundation of every traditional conversational bot. It tells the bot "the user wants to check their order status" or "the user is asking about a refund" so the bot can route to the correct dialog. In Agentforce, the equivalent concept is the agent topic, which uses natural-language classification descriptions instead of training phrase lists. The intent model is deterministic and trainable on closed examples; the topic classification model is LLM-driven and more flexible at the cost of less direct control.
Why intent is the right abstraction for bots and the wrong abstraction for agents
The training-phrase model and how generalization works
An intent learns from training phrases. Each phrase is a literal example of a user message that should match this intent. The NLP model converts each phrase into a vector representation and learns the cluster of vectors that map to this intent. At runtime, a user message is converted to a vector and matched against every intent's cluster. The intent with the closest cluster wins. The quality of the cluster depends on the variety of training phrases. Twenty phrases that all say "where is my order" produce a tight cluster that only matches near-identical messages. Twenty phrases that vary in length, vocabulary, and grammar produce a wider cluster that generalizes to real user variation.
Confidence scores and the threshold decision
Every intent classification returns a confidence score (typically 0 to 1). The picked intent is the highest scorer. A confidence threshold (commonly 0.7) decides whether the picked intent is trusted or treated as "no match." Below threshold, the bot takes a fallback path: ask the user to clarify, route to a default intent, or escalate to a human. Tuning the threshold is one of the most consequential bot decisions. Too low and the bot picks wrong intents with low confidence and frustrates users. Too high and the bot rejects valid messages and escalates too often. Most production bots settle on a threshold between 0.6 and 0.8 after observing real conversation logs.
Negative training and anti-phrases
Negative training tells the model what an intent does not match. Einstein Bots supports negative training phrases on each intent. A common pattern: the Order Status intent has positive training phrases like "where is my package" and negative training phrases like "I want to return this order" (which belongs to the Returns intent). Negative training improves intent boundaries when two intents naturally overlap in vocabulary. Without negative training, two intents with similar phrasing produce confusing classifications that flip between turns. With negative training, the boundary is explicit and the bot picks consistently.
Intent vs entity: classifying versus extracting
Intents classify; entities extract. The intent of "I want to return order ORD-1234" is Returns (classification). The entity in that message is ORD-1234 (an OrderNumber extraction). Bots use both together: the intent picks the dialog, and the entity provides the input the dialog needs to act. Confusing the two is a common error. A bot author who adds order numbers as training phrases on the Order Status intent will produce a model that matches based on the presence of an order number rather than the user's actual goal. Entities go in entity definitions; phrasing patterns go in intent training phrases.
Tuning, evaluation, and the confusion matrix
Intent quality is measured with a confusion matrix that shows for every intent, how often the model picks it correctly versus picking a different intent. The Einstein Bots Intent Insights report surfaces this matrix from production logs. The off-diagonal cells are the targets for improvement: an intent that gets confused with another intent needs more distinct positive training phrases or negative training phrases that exclude the other. Tuning is iterative. Most production bots reach acceptable quality after three or four tuning passes, each driven by the confusion matrix from the prior round of production conversations.
The relationship to Agentforce topics
Agentforce topics serve the same role as intents in the bot stack: classify the user message into a category that drives downstream behavior. The mechanism differs. Topics use natural-language classification descriptions that an LLM evaluates against the message. Intents use training phrase clusters that a smaller NLP model evaluates. Topics are easier to author for new categories (no need to invent 50 training phrases) and harder to constrain precisely (the LLM may match against vocabulary the author did not consider). Intents are harder to author (the training phrase set is real work) and more predictable in production. Most teams use the right tool for the conversation: intents for highly scripted bots, topics for adaptive agents.
Lifecycle: where intents are still relevant
The intent concept is alive but the surface is shifting. Einstein Bots continues to use intents and continues to be supported. Agentforce uses topics. The Einstein Intent Service (a standalone API for intent classification outside of bots) is in maintenance mode. The Einstein Bots intent model continues to receive updates per release. New conversational AI work in Salesforce is going to topics; existing intent-based bots continue to run and tune as expected. The right framing in 2026: intents are a stable, well-understood tool for the bot use cases that fit them, and topics are the active investment area for new capability.
How to design and tune intents that hold up in production
The intents that work in production are designed against real user phrasing, not against the author's imagination. Pull conversation logs (even from a competing channel like email-to-case) and write training phrases that match how customers actually word their questions, then tune against the confusion matrix from real bot runs.
- List the intents the bot must recognize
Start with five to ten intents covering the most common user goals. Add an Out of Scope intent that the bot uses to escalate when no other intent fits. Resist starting with twenty intents; you cannot tune that many in parallel.
- Pull real user phrasing from existing channels
Look at email-to-case subject lines, web form submissions, and existing chat logs if any. Extract one to two hundred real messages per intent. Real phrasing trumps invented phrasing every time.
- Author 30 to 50 varied training phrases per intent
Vary length (short, medium, long), vocabulary (formal, casual, slang), and grammar (questions, statements, fragments). Twenty near-identical phrases produce a brittle model. Forty varied phrases produce a model that generalizes.
- Add negative training for intents that overlap
For each pair of intents that share vocabulary (Order Status and Returns both mention "order"), add three to five negative training phrases that explicitly belong to the other intent. The boundary gets crisp.
- Set the confidence threshold based on initial logs
Start at 0.7. Look at the first two weeks of production conversations. If too many low-confidence messages route to the picked intent and fail, raise the threshold. If too many valid messages escalate, lower it.
- Use the Intent Insights confusion matrix to drive tuning
After two to four weeks of production, open the Intent Insights report. The off-diagonal cells show intent confusions. Add training phrases or negative training to crisp up the boundaries.
- Re-train monthly for the first three months, then quarterly
Intent models drift as user phrasing evolves. Monthly retraining for the first quarter catches the early variation. Quarterly thereafter is sufficient for most bots.
Literal examples of user messages the intent should match. Aim for 30 to 50 varied phrases per intent.
Messages that should not match this intent. Critical for intents that share vocabulary with another intent.
Minimum confidence required to trust the picked intent. Tuned to balance over-matching and over-escalation.
Catch-all intent for messages that fit no defined intent. Drives escalation behavior.
How often the intent model retrains against new conversation data. Monthly for new bots, quarterly for stable ones.
- Twenty near-identical training phrases produce a brittle model that only matches near-identical messages. Vary length, vocabulary, and grammar across the training set.
- Confusing intents (which classify) with entities (which extract) is a common authoring error. Order numbers belong in entity definitions; phrasing patterns belong in training phrases.
- Without negative training, intents that share vocabulary flip between turns. Add explicit negative phrases for every overlapping pair.
- The confidence threshold cannot fix bad training data. Tuning the threshold helps at the margins; tuning the phrases fixes the underlying issue.
- Intent models drift as user phrasing evolves. A model that worked in Q1 may need retraining in Q3. Schedule the retraining cadence rather than waiting for complaints.
Trust & references
Cross-checked against the following references.
- Einstein Bots intent referenceSalesforce
- Agentforce topics as intent successorSalesforce
Straight from the source - Salesforce's reference material on Intent.
- Intents in Einstein BotsSalesforce Help
- Intent Insights ReportSalesforce Help
- Negative Training PhrasesSalesforce Help
About the Author
Dipojjal Chakrabarti is a B2C Solution Architect with 29 Salesforce certifications and over 13 years in the Salesforce ecosystem. He runs salesforcedictionary.com to help admins, developers, architects, and cert/interview candidates sharpen their fundamentals. More about Dipojjal.
Test your knowledge
Q1. What is an Intent in Einstein Bots?
Q2. How does the bot identify intents?
Q3. What's the key to good intent design?
Discussion
Loading discussion…