Skip to main content
Tax Academy

Legal

AI Use and Governance Policy

Where Tax Academy of Singapore uses artificial intelligence, exactly what information goes to it, and how personal data and course materials are kept away from it.

Last updated 19 November 2026

Summary

Tax Academy of Singapore is a private training provider. We are not a government body. This page explains exactly where we use artificial intelligence (AI), what information goes to it, and how we keep your personal data and our course materials away from it.

  • We use one AI provider: Anthropic, through its paid business service (the Claude API).
  • AI does not decide anything about you. It plays no part in admissions, marks, certificates, payments, refunds or access to courses.
  • Our systems never send an AI model your name, email, phone number, address, payment details, registrations, attendance, marks, certificates, medical certificates, support requests or sign-in records. Personal data can reach it in two ways only: names spoken in a video recording, which are part of the caption text used for video descriptions (see 2 below), and anything a staff member types into the staff chat (see 3 below). Our rules forbid typing personal data into the chat, the chat screen warns against it, and our code removes contact details and identity numbers before sending, but it cannot reliably remove names.
  • The AI models cannot reach our database or our files. They see only the text our own code sends them, built from a fixed list of fields.
  • When staff ask our chat assistant to email a group of people, our code finds the people. The AI is never sent their names, email addresses or ids, only how many people are in each group, and a person always confirms before anything is sent.
  • Under Anthropic's business terms, what our portal sends is not used to train its models and is deleted within 30 days. Anthropic keeps content longer only when it is flagged under its usage policy or the law requires it. See Anthropic's privacy centre.

Where we use AI

We use AI in four places. Nothing else in the portal sends data to an AI model, and automated checks in our build block any new connection to an AI service unless it is added on purpose and reviewed.

1. Tax News translations

What it does: writes an English version of news headlines and summaries that a publisher released in another language.

What is sent: the headline and summary text the publisher put in its public news feed. Nothing about you or about what you read.

What comes back: an English translation, shown automatically next to the original and labelled as an AI translation.

2. TaxTOC video descriptions

What it does: drafts the short description shown under a TaxTOC video.

What is sent: the video's caption text (the words spoken in the video), its title, its category, its length, and the titles of the published TaxTOC courses it belongs to. If the first draft breaks our writing rules, Claude's own draft and the list of problems are sent back once for a second try. The caption text is part of our course material, so we list it here. It is sent as it is, and spoken words can include the names of speakers, of audience members who speak, or of people a speaker mentions. This applies to every video with uploaded captions, including videos not yet released and videos from our partners. No audio or video file is sent to any AI service.

What comes back: a short description. It is written automatically for a video that has captions and no description yet, or when staff press Write with AI. It appears on the site without a person checking it first, and staff can edit it at any time. A description written by our staff is never replaced.

3. AI agents for our staff

What it does: a staff-only tool that researches the tax education market and reviews our own offerings, and a chat assistant (the chief of staff) that answers staff questions about these reports and their cost. Only administrators, and staff they give access to, can use it. Learners never see or talk to it.

What is sent:

  • Public information about our offerings, for items published or announced (events that have not ended: upcoming and ongoing events only): programme codes, titles, levels, formats, durations, prices, the public summaries and descriptions our staff wrote, module titles, delivery modes and the number of intakes; event titles, dates, formats, prices, categories and summaries; TaxTOC course titles, summaries, prices, and the number and minutes of their videos. Text our staff wrote for the public can include the names of trainers or speakers, as published on our site.
  • Counts with no names: registrations per programme and per event in the last 30 and 90 days, TaxTOC plays per course, news reads per tax area, and how many people list each interest. A count below 5 is shown as "fewer than 5". The daily brief, the weekly reports and the monthly review all receive these counts.
  • The names of competitors, professional bodies and universities our administrators ask it to follow, and the names of our tax areas.
  • Pages and search results from the public web. The search queries the AI writes go to Anthropic's search service.
  • A focus our staff write for a report, which our code checks for contact details and identity numbers before sending. It only adds to the standard report; the standard sections, safety rules and data limits always stay.
  • For the chat: what our staff type, the earlier messages of that conversation, and the agents' own reports, run records and spending figures.

Before anything is sent, our code removes email addresses, phone numbers, identity numbers, postal codes, unit numbers, invoice and registration numbers and long codes from the text. It cannot reliably remove names.

What these agents are never sent by our code: lesson content, e-learning, quizzes, exam papers, files, video recordings or captions, marks, payments, support requests, or any record about a learner.

  • To help staff draft text: what they ask for and the text the assistant writes (for example a job description, a LinkedIn post, a web page or a testimonial). For a job description, our code first sends the job titles of our staff, how many staff hold each title (a count below 5 is shown as "fewer than 5") and the portal roles held for each title. It sends no names, email addresses, ids, biographies or pay. Job titles are sent as staff entered them (after our code removes contact details and identity numbers), and a title held by one person can identify that person. Role codes are sent only for titles held by 5 or more people.
  • To prepare an email to a group: the list of groups staff can choose from (public event, intake, programme and tax area names, with how many people are in each; a count below 5 is shown as "fewer than 5"), and the subject and message the assistant writes. Email recipients are never sent to the AI, only group counts.

What comes back: one-page reports for our staff, kept in the portal. Text the assistant drafts is saved as a private draft for the staff member who asked, who may share it with colleagues. A testimonial is drafted with a placeholder instead of a name and stays marked "Draft, not approved by the person named" until staff record that the person approved the wording. A report made with a staff member's own focus is private to that person and the colleagues they choose to share it with (only people who can use the AI agents page), and it is never emailed. A daily summary is emailed to staff who ask for it. The agents cannot publish content or change settings, and they cannot send an email. When staff ask the assistant to email a group, it only prepares the email: the staff member sees the message, the number of people and, for a small group, their names, and presses Confirm before anything is sent. A reminder is about one event or one intake and goes to the people signed up to it or to its trainers, or to staff as an internal message; the staff member ticks that it is a service message, not marketing, and a reminder that reads like marketing is blocked. Promotional email goes only to people who opted in to marketing, through the same consent, daily limit, suppression and unsubscribe rules as our other marketing email. A change to settings that an agent suggests takes effect only when an administrator presses Confirm.

4. Building and maintaining the portal

Our developers use AI coding assistants to write and check code. Every change is reviewed by a person and approved before it goes live.

When investigating a problem, an assistant may read from the live database. Our rule is that it does so only through our lookup tool, which:

  • can only read (the database refuses any change made through it);
  • replaces, with "[masked]", columns that hold names, email addresses, phone numbers, identity numbers, addresses, sign-in details and secrets, and any email address, phone number or identity number found inside other text, before the result reaches the assistant.

The tool cannot catch every case, for example a name written inside free text, so assistants are also instructed not to ask for personal data. The developer's computer holds the database connection, so using the tool is a rule we follow and check, not a barrier that makes other access impossible.

Payment card details are never in our database (Stripe holds them).

How we protect your data

  • Least data: each AI use above is built from a fixed list of fields in our code. Adding a field is a code change that is reviewed by a person, and automated checks that run in every release build stop a release that adds a new AI connection or widens what our AI agents may read.
  • No access to our systems: an AI model receives text and returns text. It cannot query our database, open our files or sign in. Our agents' web browsing is blocked from our own website and file storage.
  • Web pages are treated as information, not instructions, so a web page cannot make an agent do something.
  • People stay in charge: the AI agents cannot change settings, publish content or send email. An email they prepare is sent only when a staff member who has the right permission confirms it, and a change to settings they suggest takes effect only when an administrator confirms it. Video descriptions and news translations are published automatically and staff can correct them. Daily briefs are emailed to the staff who asked for them. Drafts are saved for the staff member who asked and are not published anywhere.
  • Records: every AI agent run is logged with the model used, the amount of text processed and its cost. Every confirmed or cancelled email, every shared draft and every recorded testimonial approval is written to our audit log. Administrators can preview what our offering review sends.
  • Spending limits: monthly and daily caps stop the agents when they are reached.
  • Access control: the AI agents page needs a permission that only administrators hold or grant.

How long information is kept

  • At Anthropic: deleted within 30 days, as described in the summary above.
  • With us: AI agent chats are deleted after 12 months without use. Reports are kept as internal records. Drafts are kept until the person deletes them or their account is erased. Emails TAddy prepared, with their subject and text, are deleted after 12 months, or sooner if the person deletes them or their account is erased. Translations and video descriptions stay with the news story or video.

Where the data is processed

Anthropic is based in the United States and may process what we send outside Singapore. Our Privacy Statement explains how we protect data sent overseas.

Responsibility and review

Tax Academy's administrators own this policy. Every change to how we use AI is reviewed by a person and by an independent check before it goes live, and this page is updated when our use of AI changes.

Questions

For questions about this policy or your personal data, contact our Data Protection Officer.

Write to us through a support request and we will pass it to our Data Protection Officer.