In March 2026 the Tax Practitioners Board released an exposure draft on the use of artificial intelligence and the Code of Professional Conduct. We read it closely.

The draft gets the principle right: existing Code obligations apply to AI-assisted tax agent services. But we thought it left one practical gap. It describes AI as something a practitioner deliberately chooses to use — selecting a tool, entering information, reviewing output. In most firms today, AI arrives through platform settings, connectors, browsers and agents that nobody experienced as a decision at all. You cannot comply with a disclosure obligation you don't know you've triggered.

So we made a submission. It's published here in full, including the proposed drafting we offered the Board.

Kognitive Submission in Response to Exposure Draft TPB(I) D62/2026

The Use of Artificial Intelligence and the Code of Professional Conduct

About Kognitive

Kognitive is an AI strategy and implementation consultancy that works with regulated professional services firms — including tax practitioner associations and accounting practices — to adopt AI tools in a manner consistent with their obligations under the Tax Agent Services Act 2009 (TASA), the Privacy Act 1988 (Cth), and related frameworks. Kognitive's recent work includes designing and implementing AI governance programmes for organisations subject to TASA — covering acceptable use policy, verification procedures, incident response, and structured vendor assessment.

General Comments

We welcome the TPB's decision to issue guidance on AI and the Code of Professional Conduct. The draft correctly establishes that existing Code and Determination obligations apply to AI-assisted tax agent services.

Our submission raises four points we believe the finalised guidance should address, and a note on the scope of "AI" as defined in paragraph 4. Our perspective is that of an AI implementation consultancy working with SMEs and microfirms; our recommendations are framed accordingly, focused on making the compliance path as navigable as possible for the smaller end of the profession, which represents the majority of the profession by headcount.

1. Helping Practitioners Recognise Their Firm's Full AI Footprint

Paragraphs 6, 9, and 22 describe AI use predominantly at the level of deliberate practitioner action — selecting a tool, entering information, reviewing output. That framing captures part of the AI landscape, but not the deployment patterns that have become mainstream over the past two years, and which a practitioner often does not recognise as a moment of client disclosure.

Four such patterns are widespread in practices today:

• Embedded AI in productivity and accounting platforms. Microsoft 365 Copilot, Google Workspace Gemini, Xero, MYOB, and QuickBooks run AI features — including tenant-level indexing of client content — once enabled. Enablement is typically a firm-level administrative decision, often made by an external IT provider, and does not surface to the engagement-signing practitioner as an AI-adoption event.
• Connector-based AI assistants. ChatGPT, Claude, and comparable assistants can access a firm's email, calendar, documents, CRM, and practice management systems through connectors. One configuration step authorises ongoing access; every subsequent query draws client content into the vendor's processing pipeline.
• AI-native browsers. ChatGPT Atlas, Perplexity Comet, and Dia route page content through an AI model as part of ordinary operation. Enablement is often a single click within ChatGPT or Perplexity — tools already widely used across the profession — after which viewing any client portal, webmail page, or practice management dashboard exposes that content to the AI vendor.
• Agentic AI and computer-use capabilities. Recent products — Anthropic's Claude Cowork, OpenAI's Operator, Perplexity Computer, and other agentic tools — allow an AI model to take actions on the practitioner's behalf: navigating applications, reading content across windows, interacting with systems the practitioner is logged into. The vendor's processing reach becomes effectively the reach of the practitioner's session.

Worked example (de-identified). In a compliance review of an SME accounting firm with fewer than 20 employees, leadership had explicitly instructed staff not to enter client information into AI chat interfaces. A review of actual staff practices found that several team members had independently enabled AI connectors in email, calendar and document storage tools, and installed AI-native browsers on their work machines. Client information was routinely passing through third-party AI vendors — outside the chat interface leadership had focused on, and without the leadership's or the affected staff's awareness that it was occurring. This pattern — leadership addressing the visible part of AI use while the larger operational footprint goes undetected — is consistent with what we see in advisory conversations across the profession.

We are not suggesting the draft excludes these patterns from Code item 6's reach. The compliance risk is that the draft does not make them visible to the practitioner who needs to comply. The same recognition gap also bears on Code items 9 and 10, section 30, and section 40 — each presupposes the practitioner knows what tools are operating in their practice.

What we recommend

Three clarifications — each sized to sit within the draft's existing structure — would materially improve practitioners' ability to comply:

1. Expand paragraph 6 to name the common deployment patterns through which AI reaches tax practices today — embedded platform AI, connector-based assistants, AI-native browsers, and agentic/computer-use capabilities — alongside the existing list of AI capabilities. This helps practitioners recognise AI as they encounter it in their own environments. (Proposed drafting: Appendix A.1.1.)
2. Add a consideration to paragraph 9 directing practitioners to identify how AI is invoked in their practice — whether by deliberate user action, platform-level configuration, connectors, AI-native browsers, or agentic permissions — as a prerequisite to addressing the other considerations already listed. (Proposed drafting: Appendix A.1.2.)
3. Clarify in paragraph 22 that the consent obligation applies wherever client information is disclosed to an AI vendor, however the disclosure is technically occasioned; and confirm — consistent with the mechanism TPB has already accepted in structurally similar situations (e.g. TPB(PN) 1/2017, Cloud computing and the Code of Professional Conduct) — that this mechanism extends to disclosures to AI vendors. (Proposed drafting: Appendix A.1.3.)

These additions are small in drafting terms but substantial in compliance terms. They preserve the architecture of the draft while addressing the single biggest barrier — practitioner recognition — to the mechanisms the Code already provides operating as intended.

2. Vendor Assessment Needs a Scalable Compliance Path

Once practitioners have recognised the AI footprint within their firm [per Point 1], the next question is how they assess whether each tool is suitable for use with client information.

Paragraph 24 of the draft states that practitioners should "complete appropriate review of commercial AI tools." Two practical problems arise for the smaller end of the profession, which represents the majority of the profession by headcount.

A tax-specific overlay is missing. The draft signposts OAIC's Guidance on privacy and the use of commercially available AI products, which provides detailed general AI governance guidance. What OAIC cannot provide is the tax-specific overlay: which vendor attributes are load-bearing for the statutory obligations tax practitioners actually hold under the Code and Determination. Only the TPB is positioned to provide this overlay, and absent it, practitioners are left to translate general privacy guidance into tax-specific compliance on their own.

Individual vendor assessment does not scale. Appropriate review, done properly, involves considerations across multiple dimensions — typically including jurisdiction, legal and contractual position, security, AI-specific configuration, and a documented local decision about which categories of information each tool is approved to handle. That is a level of due diligence most micro firms cannot produce in-house or afford to commission externally. If every sole practitioner must independently assess every AI vendor they use, compliance becomes economically irrational: the assessment cost exceeds the tool cost. The risk is that a risk-averse practitioner abandons AI altogether, foregoing productivity gains the profession otherwise stands to benefit from.

Worked example (de-identified). In conducting a vendor assessment for a membership organisation serving the tax profession, two tiers of the same chat assistant produced different outcomes. The enterprise tier — providing an explicit APP commitment, Australian-governing-law contract, contractual exclusion of customer data from foundation model training, and SOC 2 Type II certification — was assessed as suitable for use with client communications subject to documented consent. The consumer tier of the same product, lacking equivalent contractual commitments and administrative controls, was assessed as suitable only for non-client information.

What we recommend

1. The finalised TPB(I) provide a short tax-specific overlay to paragraph 24 — identifying, at a practical level, which vendor attributes practitioners should verify in order to satisfy their obligations under Code item 6, section 30 of the Determination (record-keeping), and section 40 (quality management). (Proposed drafting: Appendix A.2.1.)
2. The finalised TPB(I) explicitly recognise that a sector-level vendor assessment (produced by a professional association or other qualified source) combined with a local suitability determination by the practitioner is one acceptable form of evidence of "appropriate review" under paragraph 24. This does not foreclose other approaches; it gives smaller practices a viable path that is consistent with how TPB already treats engagement letter templates under TPB(PN) 3/2019. (Proposed drafting: Appendix A.2.2.)

Appropriate review is not a pass/fail threshold. Different answers to the relevant questions can lead to different compliant outcomes — use with client data, use for non-client tasks only, or do not use — provided the questions are asked, the answers documented, and the resulting use scope respected.

The practical value of such an overlay is illustrated by a question practitioners are already asking: whether an AI tool can be used to process a document that contains a client's tax file number. The answer engages at least three obligations simultaneously — Code item 6, the Privacy (Tax File Number) Rule 2015, and the vendor's contractual and security posture — and cannot be resolved by reference to any one of them alone. This is the kind of question the finalised guidance should help practitioners navigate.

3. Signpost the Interacting Regulatory Frameworks

Tax practitioners adopting AI must navigate obligations across multiple frameworks simultaneously — the Code and Determination, the Privacy Act 1988 and the Australian Privacy Principles, the Privacy (Tax File Number) Rule 2015, the OAIC's guidance on commercially available AI products, the APESB's technology-related revisions to APES 110, the Voluntary AI Safety Standard, and — for those providing designated services from 1 July 2026 — additional obligations under the Tranche 2 AML/CTF reforms.

We recommend the finalised TPB(I) include a short signposting section that maps the key decision points a practitioner faces when adopting an AI tool to the relevant obligations — not to reconcile other regulators' frameworks, but to help practitioners identify which obligations are triggered at each step. Even a simplified reference would materially reduce the burden on practitioners who are otherwise left to identify these connections independently — for example:

"Before adopting an AI tool for use with client information, practitioners should consider their obligations under Code item 6, APP 6, APP 8, APP 11, and where applicable, the Privacy (Tax File Number) Rule 2015."

This is a modest addition to the guidance. It places no obligation on the TPB to reconcile other regulators' frameworks, but it directs the reader to the right questions. The TPB is uniquely positioned to do this for the tax practitioner population, and we encourage the Board to do so.

4. Transition Provision for Existing AI Deployments

On the day the finalised TPB(I) takes effect, a substantial proportion of the profession will already have AI tools deployed under engagement letters and general authorities that were signed before this guidance existed. Without an explicit transition provision, the finalised guidance creates instant technical non-compliance for early adopters — a perverse outcome given that early adoption of AI in the profession is something the TPB, the ATO, and the Australian Government's National AI Plan 2025 otherwise actively encourage.

What we recommend

The finalised TPB(I) provide a reasonable transition period — consistent with the TPB's treatment of analogous transitions in prior information products — during which practitioners may update engagement letters and general authorities for existing client relationships to reflect the AI footprint already in operation. The transition should apply only to AI deployments in place at the date the finalised guidance takes effect; new deployments from that date forward should be expected to comply at the time of adoption. (Proposed drafting: Appendix A.4.)

A Note on the Scope of "Artificial Intelligence"

Paragraph 4 of the draft adopts the OECD definition of AI (via the National AI Plan 2025):

"a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment."

On a plain reading, this definition captures functionality that tax practitioners have used for years without regulatory attention — including email spam filters, predictive text in email clients, transaction auto-categorisation in accounting software, and pattern-matching features such as Excel's Flash Fill.

We do not suggest the TPB should adopt a narrower definition — the OECD definition is the accepted standard. We do suggest that the finalised TPB(I) clarify the intended scope, so that a literal reading does not require vendor assessment and client consent for a spam filter.

Conclusion

The exposure draft is a constructive step towards regulatory clarity on AI use in tax agent services. Our central observation is that AI in tax practices today is often embedded in platforms, authorised through connectors, routed through AI-native browsers, or granted agentic control of the practitioner's session — patterns that produce disclosures practitioners do not always recognise as disclosures. The finalised guidance can address this most effectively by making these deployment patterns visible in paragraphs 6, 9, and 22, supported by a scalable vendor-assessment path, practical signposting, and a reasonable transition for deployments already in place.

We would welcome the opportunity to participate in a TPB roundtable on AI governance or to brief the relevant working group, and are happy to discuss contributing further de-identified worked examples should that be useful for the finalised guidance.

Ethan Glessich
Managing Director
Kognitive Pty Ltd
www.kognitive.com.au

Appendix A — Proposed Drafting for the Board's Consideration

For each recommendation in the body of this submission, we have included proposed paragraph wording that the Board may wish to consider, adapt, or replace. The wording is offered to reduce the drafting burden on TPB staff and has been framed to slot into the existing structure of the Information Sheet. The TPB's existing tone and style conventions have been followed to the extent practicable.

A.1.1 — Expansion of paragraph 6

Context: Addition to the existing bullet list in paragraph 6, following the capabilities already listed.

In addition to the AI capabilities described above, tax practitioners should be aware that AI reaches a practice through a range of deployment patterns, including:

• direct use of AI tools, such as entering information into an AI chat interface;
• embedded AI features in productivity and accounting platforms, including email clients, office productivity suites, and practice management or accounting software, which typically activate at the firm or tenant level once enabled;
• connector-based AI assistants, which authorise an AI tool to access client information in email, calendar, documents, CRM, or practice management systems through a one-time configuration step;
• AI-native browsers, which route web page content — including content from client portals, webmail, and practice management dashboards — through an AI model as part of ordinary operation; and
• agentic and computer-use AI, which allows an AI tool to take actions on the practitioner's behalf by interacting with applications the practitioner is logged into.

Each pattern can result in the disclosure of client information to a third-party AI vendor, whether or not the practitioner perceives the use as a discrete action.

A.1.2 — Additional consideration in paragraph 9

Context: New first bullet inserted into the existing paragraph 9 bullet list.

If A.1.1 is adopted:

• how AI is invoked within the practice, having regard to the deployment patterns described at paragraph 6, and the categories of client information each pattern may reach;

If A.1.1 is not adopted:

- how AI is invoked within the practice — whether by deliberate user action, platform-level configuration, connectors, AI-native browser features, or agentic permissions — and the categories of client information each pattern may reach;

A.1.3 — Clarification of paragraphs 22 and 23

Context: Two-part change. The first replaces the parenthetical in paragraph 22; the second appends a sentence to paragraph 23.

Replacement for the parenthetical in paragraph 22, replacing "(which can include entering client information into AI chatbots/copilots, depending on how these tools are configured and used)":

A disclosure occurs wherever client information is transmitted to an AI vendor, regardless of the technical means — including, but not limited to, the deployment patterns described at paragraph 6. When obtaining client permission for such disclosure, tax practitioners should clearly inform the client about the proposed disclosure, including to whom the disclosure will be made and where data will be stored.

Sentence to append to paragraph 23:

This general authority extends to disclosures to AI vendors, consistent with TPB(PN) 1/2017 Cloud computing and the Code of Professional Conduct*.*

A.2.1 — Tax-specific overlay to paragraph 24

Context: Indicative content for a tax-specific overlay, which could be added to paragraph 24 as a continuation of the existing paragraph or as a new paragraph immediately after.

In conducting this review, tax practitioners should consider vendor attributes that bear directly on their obligations under the Code and Determination, including:

• jurisdiction and legal framework — where data is stored and processed, and which privacy regime applies;
• contractual position — whether the vendor is bound by the Australian Privacy Principles, whether the contract is governed by Australian law, and whether client data is contractually excluded from use in training AI models;
• security and certifications — independent attestations such as SOC 2 or ISO 27001, and the vendor's data handling and retention policies;
• AI-specific configuration — tenant-level controls that prevent client data being used to train foundation models, limit data retention, and restrict sharing with other users or third parties; and
• scope of approved use — a documented decision, made by the practitioner, as to which categories of client information each approved tool may be used to process.

These considerations relate directly to Code item 6, section 30 of the Determination (record-keeping), and section 40 (quality management).

A.2.2 — Sector-level assessment recognition

Context: New paragraph following A.2.1.

The TPB recognises that the resources required to conduct an appropriate review of each commercial AI tool can be disproportionate for sole practitioners and smaller firms. Consistent with the approach taken in TPB(PN) 3/2019 Letters of engagement*, a vendor assessment produced by a recognised professional association may be relied upon by a tax practitioner, provided the practitioner:*

• satisfies themselves as to the competence and independence of the source;
• makes a documented local determination as to which categories of client information each approved tool may be used to handle within their practice; and
• retains the records required under section 30 of the Determination.

This approach provides a scalable, proportionate compliance path. It does not foreclose tax practitioners from conducting their own vendor assessments where they prefer to do so.

A.4 — Transition provision for existing AI deployments

Context: New paragraph at the end of the Information Sheet, before the "Further information" section.

The TPB recognises that, at the date this Information Sheet takes effect, a proportion of the profession will already have adopted AI tools under engagement letters and general authorities established before the issuance of this guidance. In respect of AI deployments in place at that date, tax practitioners will have a reasonable period of twelve months within which to update engagement letters, general authorities, and supporting records to reflect the AI footprint in operation within the practice. AI deployments adopted from the date this Information Sheet takes effect will be expected to comply with this guidance at the time of adoption.