Rising component costs may adjust prices on selected items without prior notice. Supply constraints may add shipping time — we’ll notify you if affected. OEM costs and supply tightening — contact us to secure stock and pricing. For projects or bulk orders, choose “Request for Quotation” on any item.
Call our specialists(65) 6978 3502 Email our sales teamsales@sourceit.com.sg

AI & Workstations

Your AI Meeting Summary Might Be Wrong. Who Takes Responsibility?

When an AI meeting summary is wrong, the responsibility still sits with people, not the software. The vendors say plainly that AI output can be inaccurate. Under Singapore's PDPA the organisation has to make a reasonable effort to keep personal data accurate. In practice that means the meeting owner has to review a summary before anyone acts on it.

This post looks at how AI summaries go wrong, what Microsoft and Zoom actually say about their own tools, where accountability should sit, and a simple review workflow your team can adopt. If you are still deciding whether AI assistants belong in your business at all, start with our analysis Your AI Assistant Might Be Your Biggest Security Risk.

How do AI meeting summaries actually get things wrong?

An AI recap is built in two steps. First, speech is turned into text. Then a large language model condenses that text into notes, decisions and tasks. An error can creep in at either step, and an error in the first step gets carried into the second.

Microsoft says its intelligent recap is based on the meeting transcript and is powered by GPT and other AI models. Zoom says its meeting summary relies on speech-to-text data and may use third-party models. Neither approach can recover words that were never captured clearly in the first place.

Common failure points include:

  • Mishearing. Product codes, names, figures and local terms are easy to get wrong when audio is poor or people talk over each other.
  • Wrong language setting. Microsoft advises that the meeting's spoken language should match what people are actually speaking if you want an accurate transcript. Mixed-language meetings, common in Singapore offices, make this harder.
  • Wrong attribution. A remark gets credited to the wrong speaker, especially when several people share one room microphone.
  • Over-confident condensing. A suggestion someone floated becomes "agreed". A "maybe next quarter" becomes a firm action item with a name next to it.
  • Invented content. Speech-to-text models can produce words nobody said. In October 2024 the Associated Press reported that researchers found OpenAI's Whisper model sometimes inserted whole fabricated sentences into transcripts. That report was about one specific model, not the Teams or Zoom engines, but it shows that this class of error is real.

What do Microsoft and Zoom say about accuracy?

Microsoft's support page for intelligent recap in Teams is direct: "Some AI-generated content might be inaccurate, incomplete, or inappropriate." The same page tells users to edit an AI summary draft for accuracy before emailing it to others. In other words, Microsoft expects a human check.

Zoom's support article on Meeting Summary with AI Companion has no equivalent accuracy warning that we could find. It does say that only the host and co-host can start or stop the summary from the meeting controls. The host can edit or delete a summary once it is ready, and the host chooses whether participants receive it automatically. That puts the host in an editorial position, whether or not they think of it that way.

Neither vendor publishes an accuracy rate for its meeting summaries, so any figure you see quoted elsewhere should be treated with caution.

So who is responsible when the summary is wrong?

The software licence won't take the blame for you. In practice, responsibility splits across three roles:

Role What they own Practical duty
Meeting owner (organiser or host) The official record of the meeting Review and correct the summary before it is shared or filed; mark it as reviewed
Each action-item owner Their own tasks and deadlines Confirm or reject tasks assigned to them within an agreed window
IT and data protection officer Settings, retention and policy Decide where summaries may be used, how long they are kept and who can see them
Anyone forwarding a summary What they send outside the meeting Forward only reviewed versions; never send raw AI output to customers or suppliers

There is also a legal angle in Singapore. The PDPA's Accuracy Obligation requires organisations to make a reasonable effort to ensure personal data they collect is accurate and complete. This applies where the data is likely to be used for a decision about the individual, or disclosed to another organisation. PDPC's guidance says the effort expected depends on the circumstances, including the harm an individual could suffer if the data is wrong.

Not every meeting summary contains personal data that feeds a decision. A summary of a performance discussion, a disciplinary meeting or a customer complaint call clearly can. If a wrong AI summary misattributes a comment to an employee and that summary later shapes an HR outcome, saying "the AI wrote it" is unlikely to help the organisation. This is our reading of the PDPC guidance, not legal advice. Check specific cases with your data protection officer.

Which meetings need a human-reviewed record?

Not every stand-up needs formal review. A risk-based approach works better than a blanket rule:

  • Low stakes: internal brainstorms and team check-ins. The AI summary is a convenience. A quick skim by the organiser is enough.
  • Medium stakes: project decisions, supplier calls and anything with budgets or deadlines. The organiser must review it, and action owners must confirm their tasks.
  • High stakes: HR, legal, board, customer disputes and anything regulated. Treat the AI output as a draft only. A named person writes or approves the final minutes. Consider whether AI should be in the meeting at all; our companion piece Should AI Be Allowed to Attend a Confidential Board Meeting? covers this.

Can better audio really reduce summary errors?

The transcript can only be as good as the audio it hears. This is the one part of the problem that hardware can help with. A laptop's built-in microphone at the far end of a table picks up room echo and air-conditioning noise along with speech. A purpose-built speakerphone or headset puts the microphone closer to the person speaking.

For small rooms, the Jabra Speak2 75 MS Wireless Conference Speakerphone (S$415.00) is one example. Jabra says it uses four beamforming microphones, with pickup up to 2.5 m and a recommended room of up to 4.5 m by 4.5 m. It also has a light-ring indicator that tells people when they need to move closer. Jabra makes these claims; we have not lab-tested it. For room acoustics more broadly, see Your S$50,000 Meeting Room Still Sounds Terrible. Here's Why.

On the software side, AI recap in Teams comes with licensing such as Microsoft 365 Copilot Annual Subscription (S$460.00) or bundles like Microsoft 365 Business Premium and Copilot Business (S$490.00). Microsoft says intelligent recap is available with Teams Premium or a Microsoft Copilot licence. Buying the licence is the easy part. The review process is what makes the output safe to rely on. Prices as listed on 25 Oct 2026.

What should a simple review workflow look like?

A workflow only works if it's light enough that people actually follow it. Here is a version that fits most teams:

  1. Announce it. Say at the start that an AI summary will be produced. Our sibling post Should Your Meeting Room Be Allowed to Record Everything You Say? covers consent and notice.
  2. Close with a verbal recap. Spend the last two minutes saying the decisions and owners out loud. This gives the AI a clear, quotable record of the decisions to summarise.
  3. Review within one working day. The organiser checks names, numbers, dates and every action item against what was said, and corrects or deletes anything wrong.
  4. Label the status. Mark the summary "AI draft" or "Reviewed by [name], [date]" so readers know which version they are looking at.
  5. Let task owners push back. Anyone assigned an action can reject or reword it before it becomes a commitment.
  6. Keep the source. Where policy allows, keep the transcript for a set period so a disputed point can be checked, then delete it on schedule.

What to do next

  • List which of your meeting types are low, medium and high stakes, and write the review rule for each.
  • Check who in your tenant can start transcription and AI summaries, and who receives them by default.
  • Add a "reviewed by" line to your summary template.
  • Fix the audio in the rooms that produce your most important meetings.
  • Brief staff that an AI summary is a draft, not minutes, until a person signs it off. If unapproved note-taking bots are turning up in your calls, read The Meeting Was Confidential. So Why Is There an AI Transcript?

Frequently asked questions

Are AI meeting summaries accurate enough to use as official minutes? Not without review. Microsoft states that some AI-generated content might be inaccurate or incomplete, and tells users to edit summaries for accuracy before sharing them. Treat AI output as a draft until a named person has checked it.

Who is responsible if an AI summary assigns the wrong action item? The meeting owner is responsible for the record, and each person assigned a task should confirm it. The AI vendor does not take on responsibility for decisions your team makes based on its output.

Does the PDPA apply to AI meeting summaries? It can. Where a summary contains personal data that may be used to make a decision about someone or disclosed to another organisation, the PDPA's Accuracy Obligation expects a reasonable effort to keep that data accurate and complete. Check specific cases with your data protection officer.

Can a Zoom host edit an AI Companion meeting summary? Yes. Zoom's support article says the host can edit and delete meeting summaries once they are available, and the host decides whether participants receive the summary automatically.

Will a better microphone improve AI meeting notes? It helps with the first step. Summaries are built from speech-to-text, so clearer audio gives the AI a better transcript to work from. It cannot stop the AI from misreading what a discussion meant, so human review is still needed.

SourceIT Approved

SourceIT Approved

Researched and verified by the SourceIT team for Singapore buyers.

Share This Article
Facebook

More from our blog


← Older Post Newer Post →