Why intent matters
Intent drives three critical decisions: where to route the contact, which automation can handle it, and what information the agent needs ready.
When someone calls saying “I want to cancel,” that intent tells you several things immediately. They need retentions, not general support. They’re potentially frustrated or dissatisfied. They might be price-sensitive or have had a service issue. Route them wrong and you’ve wasted their time and increased the chance they follow through on cancelling.
When someone chats asking “where’s my order,” that intent is straightforward enough for automation. A bot can look up the order status and provide tracking information without human involvement. But only if you correctly identify that the intent is order tracking, not complaint about delayed delivery or request to change the shipping address.
Intent detection separates contacts that automation can handle completely from those needing human judgment. Miss the intent and you either trap customers in automation that cannot help them, or waste agent time on queries that could have been automated.
How intent gets detected
Traditional approaches use keywords and rules. If the customer says “refund,” “money back,” or “return payment,” the system identifies refund intent. If they say “cancel,” “close account,” or “stop service,” it identifies cancellation intent.
This works until it doesn’t. Someone saying “I don’t want a refund, I want this fixed” triggers refund detection despite explicitly stating the opposite. Someone asking “can I cancel my appointment?” gets routed to account cancellation instead of appointment management.
Modern intent detection uses natural language processing (NLP) and machine learning. Instead of matching keywords, the system understands context, sentence structure, and meaning. “I don’t want a refund” and “I want a refund” are distinguished properly. “Cancel appointment” and “cancel service” route differently.
The system learns from examples. You show it thousands of messages labelled with their true intent. It identifies patterns in how people express each intent. Then it applies those patterns to new messages, getting better over time as it sees more examples and receives corrections when it gets things wrong.
Intent in different channels
Phone systems traditionally rely on IVR menus or speech recognition. The customer speaks naturally and the system identifies intent from their words. “I need help with my bill” gets classified as billing intent. “My internet stopped working” becomes technical support intent.
Chatbots and virtual assistants analyse typed messages to identify intent. The AI for customers needs to understand variations – “My WiFi is down,” “Internet not working,” and “Can’t get online” all express the same technical support intent despite using different words.
Email and messaging channels have the advantage of complete messages to analyse. The system can read the entire email, not just the first sentence, to understand what the customer needs. This improves accuracy but introduces delay – the intent gets identified after the email arrives, not during the interaction like with voice or chat.
Intent versus sentiment
Intent and sentiment are different things that get confused constantly. Intent is what the customer wants. Sentiment is how they feel about it.
Someone might have “billing query” intent whilst being furious (negative sentiment) or just mildly curious (neutral sentiment). The intent tells you where to route them. The sentiment tells you how urgently they need help and which agent can best handle their emotional state.
Both matter, but they serve different purposes. Intent determines routing and automation. Sentiment influences prioritisation and agent matching. Mixing them up means routing correctly but missing that the customer is about to escalate, or identifying their anger but sending them to the wrong department.
Common intent problems
Too many intents
Some contact centres create hundreds of specific intents. “Billing query – invoice,” “Billing query – payment,” “Billing query – charges,” “Billing query – dispute.” Each slightly different from the others.
This creates problems. The system struggles to distinguish between similar intents. Agents get contacts that were almost classified correctly but landed in the wrong specific bucket. Reporting becomes meaningless because everything is fragmented across too many categories.
Better to have broader intents that work reliably. “Billing” covers all billing-related contacts. Sub-categorisation can happen after routing, once a human or more specialised automation is handling it.
Too few intents
The opposite problem is having so few intents that they’re useless. “Customer service” and “technical support” as your only two intents provides almost no routing value. Everything becomes a judgement call about which vague category fits better.
Effective intent structures typically have 15-30 primary intents covering the main reasons customers contact you. Enough to route meaningfully, not so many that accuracy collapses.
Misidentified edge cases
Most intent detection works well for common, clearly expressed intents. Problems emerge with edge cases, ambiguous messages, or customers who bury their actual intent in rambling explanations.
“I got an email saying my payment failed but I’m sure I paid and my bank says the money left my account so I’m confused about what’s happening and whether I need to pay again” contains multiple possible intents – payment query, billing dispute, account status check.
The system needs to either ask clarifying questions to narrow down intent, or default to routing to someone who can handle any of those possibilities. Getting it wrong means customers explaining themselves multiple times.
Intent drift
Customer language changes over time. New products launch with new terminology. Competitors introduce new services that change how people describe things. Intent models trained on last year’s data might struggle with this year’s language.
Regular retraining prevents drift. The system learns from recent interactions, incorporates new patterns, and stays current with how customers express their needs.
Intent and automation
Intent detection determines what automation can handle. Simple, high-volume intents like order tracking, balance enquiry, or appointment scheduling are perfect for bots and self-service.
Complex intents requiring judgment, empathy, or authority need humans. Complaints, vulnerable customers, complex troubleshooting – these should route to agents even when the intent is correctly identified, because automation cannot handle them properly.
The challenge is building systems that know their limits. A bot that correctly identifies “refund request” intent but cannot process refunds should hand off cleanly to humans rather than failing slowly whilst the customer gets frustrated.
Intent confidence scores
Most intent detection systems provide confidence scores alongside their predictions. “I’m 95% confident this is billing intent” versus “I’m 60% confident this is billing intent.”
High confidence means the system is certain about what the customer wants. Low confidence means the message is ambiguous or doesn’t match known patterns well.
Smart systems handle these differently. High confidence intents can route automatically. Low confidence intents might trigger clarifying questions (“Are you calling about billing or technical support?”) or route to agents who can work it out through conversation.
Ignoring confidence scores means routing ambiguous contacts as confidently as clear ones, which creates mistakes that damage customer experience.
Multi-intent conversations
Sometimes customers have multiple intents in the same contact. “I want to update my address and also ask about my bill and check when my appointment is.”
Traditional systems struggle with this. They identify the first intent and ignore the rest, or get confused by multiple signals and classify incorrectly.
Better systems track multiple intents within a conversation, ensuring all get addressed before the contact ends. The bot or agent handles the address update, answers the billing question, and confirms the appointment. Nothing gets missed because the system only looked at the first sentence.
Intent in omni-channel environments
When customers move between channels, intent should follow them. Someone who started a chat about technical support but then called should arrive at the phone agent with their technical support intent already identified.
Without this, customers repeat themselves. “I was just chatting with someone about my internet problem.” The agent has no context, so the customer explains from scratch. Effort increases, satisfaction drops, and the technology that should connect everything becomes another barrier.
Omni-channel intent tracking ensures the customer’s goal carries forward regardless of channel changes.
Getting intent right
Effective intent detection starts with understanding your actual contact drivers. What are the top 20 reasons customers contact you? Those become your core intents.
Then look at the language customers use. Don’t use your internal terminology. If customers say “refund” but your systems call it “credit reversal,” train intent detection on “refund.” Meet customers where they are, not where your processes are.
Test with real messages, not hypothetical ones. Pull 500 recent contacts and see if your intent model classifies them correctly. Where it fails, understand why. Is the intent definition unclear? Is the training data insufficient? Are customers expressing this intent in ways you didn’t anticipate?
Monitor accuracy over time. Intent detection degrades as language changes and new patterns emerge. Regular retraining keeps it working reliably.
Most importantly, design for failure. What happens when intent cannot be identified? Default to human routing, ask clarifying questions, or provide multiple options. Failing gracefully is better than routing confidently to the wrong place.
The difference it makes
Good intent detection means customers reach the right place without explaining themselves repeatedly. Automation handles what it can. Humans handle what they must. Nobody wastes time because the system couldn’t work out what the customer wanted.
Poor intent detection creates the opposite. Customers bounced between departments. Bots that cannot help but won’t let you speak to someone. Agents receiving contacts they cannot handle. The technology exists but makes everything worse instead of better.
Intent is the foundation of intelligent routing, effective automation, and good customer experience. Get it right and everything downstream works better. Get it wrong and even the best agents cannot compensate for starting every conversation in the wrong place.
Your Contact Centre, Your Way
This is about you. Your customers, your team, and the service you want to deliver. If you’re ready to take your contact centre from good to extraordinary, get in touch today.
