Bringing AI into Council Interpellation: The Complete Before-During-After Workflow

Every council session sends agency staff into interpellation hell: digging through past records, compiling sentiment issues, guessing which topics members will pursue, and chasing commitments afterwards. AI has clear leverage at all three stages — pre-session issue assessment and debate preparation, real-time transcription and structuring during sessions, and post-session task tracking. This article covers the concrete practices at each stage, the data foundations required, and how to connect the three into a closed loop.

Before the Session: Issue Assessment and Debate Preparation

The core pre-session question is: who will ask what this session? AI cross-references three sources:

• Past interpellation records：Each member's past interpellation topics, follow-up patterns and constituency concerns, to estimate the likelihood of continued questioning.
• Recent sentiment hotspots：Media and social issue volume in the weeks before the session — controversies running hot in public sentiment almost inevitably enter interpellation.
• Policy progress and weaknesses：Delayed projects, contested policies and items with low budget execution rates are all predictable interpellation targets.

From these, preparation materials are produced for each high-risk issue: likely question phrasings, factual and data evidence, a recommended answer structure, phrasing to avoid, and traceable sources. Staff work shifts from compiling from scratch to reviewing and supplementing.

1. Standard fields for interpellation-response materials

For interpellation materials to be usable on the podium, the fields must be fixed. If an agency lets each division write in its own format, the executive on the interpellation floor first has to spend time hunting for which section holds which information — this is exactly why interpellation materials most often fail in practice. It's recommended to standardize on seven fields, regardless of the issue's size or which lead unit wrote it, so the layout and order stay consistent. AI is responsible for filling in the draft of the first six fields from external sentiment data and internal knowledge; staff are responsible for review and correction; the executive only needs to memorize the field order. Below are the seven fields common across agencies:

• Issue name and risk level：Define the scope of the issue in one sentence, and mark it high, medium, or low risk to determine how deeply to prepare and the order in which the executive should read ahead.
• Possible lines of questioning (including follow-up paths)：List not just the main question, but also project the second and third questions most likely to follow once the other side receives a given answer.
• Confirmed facts and data basis：Include only facts and statistics that the lead unit has confirmed; flag unconfirmed information separately, to avoid being caught without an answer if the source is challenged on the podium.
• Suggested response structure：State the conclusion first, then the basis, then the next steps — a three-part structure that keeps the response complete even under time pressure.
• Language to avoid：List wording that is easily taken out of context, conflicts with existing positions, or exceeds the agency's authority — this field is often more critical than the affirmative talking points.
• Related past commitments and their status：List past commitments made in council sessions alongside their current progress, to avoid giving answers that contradict earlier statements.
• Link to source material：Every fact should be one click away from the original news article, official document, or report, so the executive and staff can verify it independently.

2. Preparation timeline before a session opens

Interpellation prep most often fails not because of a lack of capability, but because it starts too late. Pulling preparation earlier and breaking it into four checkpoints ensures that materials have already gone through at least two rounds of manual review by the day the session opens. Below is a suggested schedule — agencies can adjust it to the length of their session, but the order of the four checkpoints should be kept as is:

• One month before the session opens：Compile sentiment hotspots for the session period and historical interpellation records, cross-reference them to produce a high-risk issue list, and assign a lead unit for each issue. The focus at this stage is coverage — it's better for the list to run long than to discover a gap after the session has already opened.
• Two weeks before the session opens：Each lead unit completes a draft of its interpellation materials in the seven-field format and finishes the first round of manual review. Facts and figures generated by AI must be checked item by item against the original source material, and only proceed to the next stage once confirmed accurate.
• One week before the session opens：Update the latest sentiment changes, add issues that newly emerged during this period, and adjust the risk level of existing issues. Sentiment tends to be most active in the week before a session opens, so this round of updates cannot be skipped.
• The day before the session opens：Produce a same-day reminder covering today's scheduled interpellating units, their corresponding high-risk issues, and sentiment changes from yesterday to now, kept within one page so the executive and accompanying staff can grasp it quickly.

During the Session: Real-Time Transcription and Structuring

The pain point during sessions is record-keeping: oral exchanges move fast, manual notes miss things, and reviewing recordings afterwards is slow. Speech recognition (ASR) with meeting intelligence tools transcribes interpellation verbatim in real time and auto-structures it into a member–question–answer–commitment list. For the team on duty, real-time structuring has a tactical value too: back-row staff can push relevant materials to the official at the podium before the member's follow-up lands.

1. Technical essentials of real-time transcription

Legislative and council settings are among the most difficult scenarios for Chinese speech recognition, and three challenges need to be acknowledged before adoption. First, proper nouns: personal names, place names, bill titles, agency abbreviations, and budget line items appear infrequently in general-purpose training data and are easily misrecognized as homophonous everyday words. Second, mixed Taiwanese and Mandarin: language switches frequently during interpellation, and the same sentence often mixes both languages. Third, overlapping speech from multiple people: interpellation and responses interrupt each other, with council staff explanations layered on top, making speaker separation harder than in an ordinary meeting. The practical solution is to build an agency-specific vocabulary — importing the agency's organization names, domain terminology, names of ongoing programs, and frequently cited regulation names — so the recognition model prioritizes these during decoding, while keeping a manual proofreading process for key passages. To understand how recognition accuracy is measured, seeASR Speech Recognition Model Architecture and CER Metric Explanationshould be referenced. It should be emphasized that real-time transcripts are positioned as working drafts for in-session tracking and post-session organization — the official record should still be the manually confirmed version.

2. Real-time staff support workflow

What actually creates the difference during a session is how fast the staff in the back row can react. The traditional approach has staff judge the issue from memory and then flip through paper materials, and by the time they've found what they need, the moment for a follow-up has often already passed. After adopting real-time transcription, the workflow changes: the system identifies the keyword of the issue currently under discussion, automatically matches it against the pre-prepared interpellation materials, and pushes the corresponding page to the screen in front of the back-row staff, who confirm it before handing it to the executive answering questions. At the same time, the system organizes the content during the session into a structured list of legislator, question, response, and commitment; staff only need to correct and annotate alongside it, so the list is nearly finished by the time the session adjourns, without having to be re-transcribed from a recording afterward.

After the Session: Commitment Tracking and Outcome Comparison

"We'll look into it," "we'll provide the data after the session" — commitments made at the podium, if not systematically managed, become next session's follow-up interpellation: "you said you'd handle it — did you?" Post-session, AI does two things: extracts commitments and assignments from the transcript into a tracking list with lead units and deadlines, and keeps comparing sentiment on related issues so the agency knows whether its response actually resolved the concern.

1. Fields for the commitment-tracking list

Commitment-tracking lists often fail because they record only a one-sentence summary. A summary can't reconstruct the context from the next session, nor can it answer "exactly what level was this commitment made at." The list should have six fixed fields, so that every commitment can still be fully reconstructed six months later:

• Original transcript passage of the commitment：Keep the original transcript text rather than a summary, and attach a timestamp so it can be played back for comparison if necessary.
• Lead unit：Specify down to the division level; for cross-unit matters, designate one lead unit and list the rest as supporting units.
• Case-handling contact：Record the job title and extension rather than just a name, so the list remains usable even after personnel changes.
• Handling deadline：If no deadline was explicitly given on the podium, the lead unit sets one on its own and fills it in within three days after the session.
• Current status：Track status as one of three states — under review, in progress, or replied by letter — with a timestamp recorded whenever the status changes.
• Corresponding sentiment change：Track the discussion volume and sentiment trend for the issue after the commitment was made, to judge whether the response actually resolved public concern.

This list is effectively next session's guaranteed question bank. When producing the high-risk issue list one month before a session opens, any item still under review or in progress, with sentiment volume not yet declining, should go straight onto the high-risk list. Items already replied to by letter but still contentious in public sentiment also need follow-up explanations prepared. In other words, the quality of post-session tracking directly determines the starting point for the next round of pre-session preparation — a well-maintained list means next session's prep is effectively half done already.

The full picture of AI support across one session

Laying the three stages out on a single timeline shows that the division of labor between AI and staff stays consistent throughout: AI handles collecting, cross-referencing, and drafting; staff handle confirming, judging, and external communication. This boundary needs to be made clear to every division early in adoption, to prevent staff from mistakenly believing the system's output can be submitted directly, and to prevent the executive from assuming everything has already been verified. The table below takes one complete session as an example, listing the AI output and corresponding human work at each point in time — agencies can use it directly as a division-of-labor explanation during adoption.

Local Councils vs. Legislative Yuan: Differences in Implementation Design

The closed-loop framework is universal, but the data design must be adjusted by level. Local council issues center on electoral districts — the construction progress of a road, how quickly a complaint case is handled, the maintenance condition of a public facility. These issues spread extremely fast through local community groups and local media, and often form public sentiment before a formal complaint is even filed. Legislative Yuan issues center on institutions — a bill's review progress, the conditions for freezing and releasing a budget, division of responsibility across ministries. Discussion concentrates in national media and policy communities, with a longer issue cycle but broader reach. The two also need different tracking fields: for local councils, the tracking focus is whether individual cases are actually resolved and whether promised completion dates are met; for the Legislative Yuan, the tracking focus is the progress of article amendments, the timeline for submitting reports, and the state of cross-ministry coordination. When setting issue categories, sentiment source weighting, and tracking fields, agencies should adjust for the level they operate at. The table below summarizes four major dimensions of difference:

Connecting the Three Stages into a Closed Loop

Each stage delivers value on its own, but the real payoff is the connection: the post-session commitment list is the "guaranteed exam questions" for next session's preparation; session transcripts accumulate into historical data on members' attention patterns; pre-session sentiment assessment decides which issues need live support during the session. This loop is the five-stage AI copilot model — sense, understand, assess, recommend, track — unfolded in the parliamentary setting.

On data foundations: pre-session assessment needs external sentiment monitoring (such as InfoMiner) and historical parliamentary data; in-session transcription needs Chinese speech recognition; post-session tracking needs to connect with the agency's internal case-management mechanisms. For adoption, pilot with a single session and a single committee, validate the workflow, then scale.