Slot Filling

What is slot filling?

Slot filling is the process an AI assistant uses to gather the specific pieces of information it needs before it can complete a task. Each required detail is a "slot", a labelled placeholder such as order number, delivery date, or account email. When a customer's request is missing one of those details, the assistant asks a follow-up question to fill the empty slot, and only acts once it has everything the task requires.

The idea comes from task-oriented dialogue systems. Booking a table needs a date, a time, and a party size; tracking an order needs an order number. Until every mandatory slot holds a valid value, the assistant cannot reliably finish the job, so slot filling is what turns a vague request into a structured, actionable one.

How slot filling works

A slot-filling flow has a few moving parts:

• A frame of slots. Each task defines the slots it needs and which are mandatory. Others are optional and only refine the result.
• Extraction. As the customer types, the assistant pulls known values straight out of their message. This is where entity extraction does the work, lifting an order number or a date from free text and dropping it into the matching slot.
• Prompting for the gaps. Any mandatory slot still empty triggers a targeted follow-up: "What's the order number?" The assistant asks only for what is missing, rather than restarting the conversation.
• Validation. A value is checked before it is accepted: an order number in the right format, a date that exists. If it fails, the assistant re-prompts.
• Confirmation and action. Once every mandatory slot is filled, the assistant confirms and carries out the task.

Because slots persist across the exchange, a customer can supply details in any order, or several at once, and the assistant keeps track of what it still needs. That state-tracking is handled by dialogue management, which decides the next question based on which slots remain empty.

Slot filling and multi-turn conversations

Slot filling is one of the clearest reasons a support assistant needs to remember context. Collecting three details usually takes more than one exchange, so it depends on a multi-turn conversation in which earlier answers are retained. A single-turn system that forgot the order number the moment it asked for the delivery date could never complete the task.

It also pairs closely with intent recognition. Intent recognition works out what the customer wants to do (track an order) and slot filling gathers the details needed to do it. One identifies the task; the other collects its inputs. You will sometimes see the technique described as "slots and intents", which captures exactly this pairing.

How to apply slot filling well

Keep the number of mandatory slots to the minimum the task genuinely needs: every extra question adds friction. Ask for one missing detail at a time, phrase prompts in plain language, and let customers give several values in a single message when they want to. Validate each value as it arrives, so mistakes are caught early rather than at the end. And design a graceful exit: if a slot cannot be filled, or the customer stalls, hand over to a human with everything gathered so far, so nobody has to repeat themselves.