A task-based conversational AI system is built to help a person finish one job, such as booking a flight or moving a delivery, and the classic design does this with frames: lists of slots (date, city, order number) that the system fills by asking questions. According to the Jurafsky and Martin textbook draft of August 19, 2026, the frame-based approach was set out in the GUS travel system in 1977 and stayed the dominant industrial paradigm for dialogue systems for more than 30 years. No neutral ranking of "leading companies" in this field was found, so this article explains the shared design that task-based products build on.
Key takeaways
- A frame is a set of slots the system must fill before it can act, and the simplest systems ask pre-written questions until every slot is full.
- Each user message is analyzed for three things: the domain, the intent, and the slot values.
- A dialogue state tracker keeps the whole frame, not just the latest message, and a dialogue policy decides whether to answer, ask, confirm, or reject.
- Success is measured by task success rate and slot error rate, with the length of the dialogue as an efficiency cost.
- The textbook draft notes that around 2010 the GUS architecture reached commercial phone assistants such as Siri.
Frames and slots
The textbook's appendix on frame-based dialogue systems describes a frame as a knowledge structure holding the details of a user's task, made up of slots that each take a set of possible values. A hotel booking needs dates and a location, an alarm needs a time, and the system's goal is to fill the slots with what the user means and then perform the action. Slots are usually typed, so a slot of type date accepts dates and a slot of type city accepts city names.
In the simplest frame-based systems, which the draft says included most commercial assistants until quite recently, the questions used to fill each slot are pre-written templates. More sophisticated systems generate the questions as they go. The table shows an illustrative frame for rescheduling a parcel delivery, written for this article and not taken from any product.
| Slot | Type | Pre-written question |
|---|---|---|
| Tracking number | identifier | "What is the tracking number?" |
| New delivery date | date | "Which day should the parcel arrive?" |
| Time window | time range | "Morning or afternoon?" |
| Address change | yes or no | "Should the address stay the same?" |
Three jobs for every message
When a message arrives, the system does more than look for slot values. The textbook lists three linked tasks: domain classification (is this about deliveries or billing?), user intent determination (is the goal to reschedule or to cancel?), and slot filling itself. The first two choose which frame is being filled. For the message "Move the parcel to Thursday morning," an illustrative reading would be domain delivery, intent reschedule, date Thursday, window morning.
The simplest systems use handwritten rules, such as a pattern that matches phrases like "set an alarm." Most use supervised machine learning, with training sentences annotated for slots, domain, and intent. A common arrangement passes the words through a pretrained language model encoder and tags each word as the beginning of a slot, inside a slot, or outside any slot. Extracted strings are then normalized, so that synonyms for one airport map to a single code. The draft adds that industrial systems often combine rules and machine learning for each component.
Dialogue state, acts, and corrections
A dialogue state tracker records the current state of the frame, meaning every slot filler gathered so far, together with the user's most recent dialogue act. Acts such as inform, request, confirm, and affirm label what each turn does. If a customer first says the parcel can arrive on a weekday and later adds "afternoon," the state holds both constraints.
The textbook singles out user corrections as a hard case. When a system mishears, people repeat or reformulate, and in speech they often exaggerate pitch or volume, which makes the correction harder for speech recognition to read than ordinary sentences.
Confirm, reject, or proceed
Because recognition errors are common, the dialogue policy must decide how much to confirm. The draft describes the tradeoff: explicit confirmation ("Is that correct?") makes errors easy to fix but is slow and awkward, while implicit confirmation repeats the understanding inside the next question. One described policy uses speech recognition confidence in four tiers.
| Confidence level | System behavior |
|---|---|
| Below the lowest threshold | Reject and say the input was not understood |
| Above the lowest threshold | Confirm explicitly |
| High | Confirm implicitly |
| Very high | No confirmation |
How task systems are measured
The textbook names two main measures. Task success rate is the share of conversations in which the system did the right thing, such as booking the correct flight. Slot error rate is finer: the count of inserted, deleted, or substituted slots divided by the number of reference slots. In the draft's calendar example, three slots were extracted and one time was wrong, so the slot error rate was one third. Efficiency costs, such as dialogue length in turns or seconds, are tracked alongside.
These measures give a buyer or reader something concrete to ask a vendor for. A claim that a system "understands customers" says little, while a reported task success rate on a stated set of tasks can be checked.
What this does not cover
Large language model chat products work differently from frame-based systems, and the textbook appendix itself marks its section on the modern history of neural dialogue systems as still to be written. No source reviewed ranks conversational AI companies, and market lists shift with each release, so vendor comparison needs its own evidence from independent testing and public records.
FAQ
What is a slot in conversational AI?
A slot is one piece of information a task requires, such as a date or a destination city. A frame groups the slots for one task, and the system keeps asking until the required slots are filled.
What is the difference between an intent and a slot?
The intent is the goal the user is pursuing, such as showing flights or setting an alarm. Slots are the details needed to carry it out, such as the origin city or the time.
Why do some assistants repeat what was said back?
That is confirmation. Speech and text systems misinterpret input, so the policy may confirm explicitly or fold the confirmation into the next question, depending on how confident the system is.
Is there an official list of the leading conversational AI companies?
None was found in the sources checked in October 2026. Claims of leadership usually come from vendors or from commercial research, which this article does not use.
Related reading
- An AI chat app built around companion conversation
- How predictive analytics works with data
- Evidence on AI and company costs
- Researchers behind modern AI systems
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Speech and Language Processing, third edition draft, Appendix K, Frame-Based Dialogue Systems, Daniel Jurafsky and James H. Martin, Stanford University, draft of August 19, 2026
- Speech and Language Processing, third edition draft, contents page, Stanford University, August 2026