Short answer
Intent drift is the gradual change in what customers contact a business about, so categories, routing and automation built for the old mix start missing. In AI agents, it also means an agent drifting from the goal it was given.- Also called
- Topic drift; concept drift in machine learning; goal drift in agents
- Caused by
- Launches, policy changes, seasons, carrier problems, new channels
- First sign
- A growing pile of messages that fit no category
- Hard to see because
- Systems keep assigning familiar labels to unfamiliar requests
- Mitigated by
- Allowing no match, reviewing unmatched messages, updating categories
Every set of support categories is a snapshot of what customers asked when it was written. Customers keep changing what they ask. Intent drift is the gap that opens between the two, and it rarely announces itself: the old labels keep getting applied, just to the wrong messages.
What causes intent drift
Drift follows changes in the business and around it. In a Shopify store, the usual causes are:
- A new product line. A store that adds footwear starts getting sizing questions it has no category for.
- A policy change. Shortening the return window turns some return requests into disputes about the window.
- Carrier problems. A delayed carrier turns "where is my order" into "my tracking hasn't moved in a week", which needs a different answer.
- Seasons. Gift messages, holiday cutoffs and Black Friday discount questions arrive every year and then disappear.
- A new channel. Adding Instagram brings pre-sale product questions that email rarely saw.
Three kinds of drift
Drift shows up in three forms, and each needs a different fix.
| Kind | What changes | Example | Fix |
|---|---|---|---|
| New intent | A request type that did not exist | "Does the new jacket run small?" | Add a category |
| Changed intent | The same request needs a different answer | "Where is my order?" after a carrier switch | Update the answer or the process |
| Mix shift | The share of each request type moves | Returns double in January | Rebalance staffing and automation |
In machine learning terms, a changed intent is close to concept drift, a mix shift is close to data drift, and a new intent is a class the model has never seen. The support version is simpler to state: the questions changed, the system did not.
With Aide, the changed kind often surfaces as a knowledge gap: every week Aide lists questions the team answers that knowledge does not cover and articles that say something the team no longer does, each checked against the whole knowledge base and proposed with the text and the article it belongs in.
Signs of intent drift
Most drift is visible in data the team already has, if someone looks.
- More messages labelled "other", or matched to no category at all.
- Lower confidence on classifications that used to be clear.
- More conversations reassigned after the first routing decision.
- Handoffs or reopens rising on one category while the rest stay flat.
- Agents correcting the same label again and again.
In Aide, Analytics can be filtered by topic and shows handoff reasons by week and channel, and every number opens the exact conversations it counts, so a rise on one topic can be read message by message.
Agent intent drift
In AI agents, intent drift has a second meaning: an agent losing track of the goal it was given during a long conversation or task.
It happens in three ways. The agent follows a tangent and never returns to the original request. It keeps working on the first request after the customer changed it. Or text in a message or document redirects it, which is how prompt injection works.
A shopper asks to change a shipping address, then asks whether the jacket can be returned if it doesn't fit. A drifting agent either ignores the second question or forgets the address change. A well-built one answers the return question and still finishes the address change, or hands off with both noted.
How to mitigate intent drift
Mitigation is routine maintenance, not a one-off project.
- Let classification return no match. A message forced into the nearest category hides the drift; a message left unmatched shows it.
- Read unmatched and corrected messages every week, grouped by what they are about.
- Add, split or retire categories as requests change, and add real recent messages as examples.
- Re-test any automation that depends on a category after you change it, on recent real conversations.
- For agents, restate the goal at each step, limit how many steps run without a check, and hand off when a request changes in a way the agent was not set up for.
The practice this page rejects is setting up categories once and trusting them indefinitely. A taxonomy that is never revised is drifting, whether or not anyone measures it.
Where Aide fits
Aide classifies every customer message against the team's own topics and picks up to three, or none when nothing fits. When no topic fits, it suggests a new one right on the message, so new requests show up as suggestions instead of hiding in an existing topic. The team can add examples to a topic and move them between topics, and Aide classifies with the latest examples. Within a conversation, a procedure can be set to switch when the customer moves to something another procedure fits better. See what your topics tell you.