Building conversational interfaces the right way

Conversational interfaces are powerful, but they are not always the best choice. At Airgentic we start every design with a simple question: what is the easiest, fastest and most reliable way for a user to achieve their goal? In some cases the answer is not a chat-style agent.

Here are three situations where a conversational UI is likely to be the wrong tool:

• The task requires a long, rapid-fire sequence of questions. Forms or guided multi-step pages are faster to complete and easier to validate.

• The user needs to consume a large amount of information. A well-designed web page is easier to scan and reference than a sequence of dialogue turns.

• The interaction depends on autocomplete to select from a large set of options. A conversational interface can struggle to find the precise input that an autocomplete field would present.
  

Hutt City Council made this trade-off very concrete. The council’s Too Good To Waste website enables citizens to enter an address and immediately see their next bin collection details. The existing address autocomplete field is simple, familiar and effective. The challenge was to re-imagine this experience as a conversational interface without making it worse. Our initial question was straightforward: would removing the autocomplete degrade the user experience?

After analysis we realised the key was to let the AI replicate what a human does at the keyboard. A typical human will type, watch suggestions, adjust for errors, backtrack when necessary, accept a suggestion when it looks correct and confirm when uncertain. The agent needed to behave the same way.

We therefore developed an “autocomplete brain” for the agent. The component implements the keyboard behaviour in a conversational context and includes the following capabilities:

• Progressive refinement: the search moves from broad to narrow as more of the address becomes clear, so matches converge quickly.

• Intelligent backtracking: the agent recognises when the search trail goes cold because of typos, missing characters or unusual spacing and it backtracks to recover.

• Confidence-driven selection: the agent builds confidence as it refines matches and only auto-selects when the match is strong.

• Confirmation gating: low-confidence outcomes are not guessed; the agent asks for a brief confirmation before proceeding.

• Human-friendly disambiguation: when multiple plausible matches remain, the agent presents a short, easy to read shortlist for the user to choose from.

• Input normalisation: messy inputs are normalised behind the scenes for casing, spacing and common variants so lookups remain reliable.

• Dataset hints: lookup behaviour is guided by dataset-driven signals that nudge the agent towards the most likely completion, similar to a good autocomplete field.

• Operational robustness: query budgeting and caching keep the agent responsive under load and avoid unnecessary lookups.
  

The result is that citizens can simply message their address to the agent. The agent handles the equivalent of keystrokes, corrects rough edges, and resolves ambiguity so the correct lookup is returned naturally, quickly and reliably.

Designing conversational interfaces means choosing the right interaction model for the task. A chat agent can deliver a great experience when it is the right fit, and with components such as the autocomplete brain it can also replace more traditional UI elements without compromising usability.