Accounting firm: a knowledge assistant staff can trust, and an end to the document chase
An illustrative 60-person accounting and advisory firm loses senior time to routine questions, has no rules for AI, and chases client documents by email. The system starts with an AI use policy, then gives staff an assistant grounded in the firm's own procedures, with citations and permissions, and moves document collection into a client portal.
This is a worked example: a realistic scenario we've built to show how we'd approach the problem. The organisation is fictional, and every number is a stated assumption or a calculation from those assumptions. It is not a client case study and makes no claim about results.
The situation
The firm's know-how is spread across years of folders on a shared drive, a procedures manual that hasn't been fully updated in a long time, partners' email and people's memories. Staff can't tell which engagement letter, workpaper template or checklist is current, so they ask whoever is nearest, usually a manager or partner.
Client documents are collected by email: a request at the start of the job, reminder emails, then phone calls. What's outstanding is tracked in spreadsheets and job notes, and attachments are renamed and filed by hand.
Partners can see that AI could help, but they worry about client confidentiality and the quality of AI-assisted work. Meanwhile some staff already use free AI tools on their own phones, and the firm has no AI policy. Doing nothing doesn't keep AI out of the firm. It just leaves it unmanaged.
Constraints we'd work within
- Registered tax agents in the firm must not disclose client information to a third party without the client's permission or a legal duty (TPB Code of Professional Conduct, item 6). The TPB's 2026 guidance on AI notes that entering client information into an AI tool can amount to disclosure, depending on how the tool is set up and used.
- The OAIC advises organisations not to enter personal information, and particularly sensitive information, into publicly available generative AI tools.
- The practice management and document management systems stay, and the assistant must respect their existing permissions.
- Partners must be able to see what the assistant answered and which document it relied on.
- Rollout has to avoid the busiest lodgement months, when nobody has time to learn a new tool.
What the Diagnostic would find
Findings by lens, with the commercial effect of each. See how the Diagnostic works.
| Lens | Observation | Commercial effect |
|---|---|---|
| Workforce capability | Juniors and new starters ask managers and partners for templates, procedures and "how do we do this here" answers several times a day. | Senior chargeable time is lost to interruptions, juniors wait instead of working, and new hires take longer to become productive. |
| Operations & process | Templates, precedents and procedures exist, but in scattered folders with no clear current version. Several versions of the engagement letter are in use. | Work product varies between offices, and outdated terms can reach clients. |
| Risk, governance & change readiness | There is no AI policy, no approved tool and no record of where AI is used. Some staff use consumer AI tools on personal devices for firm work. | Client information may be leaving the firm's control, which conflicts with confidentiality obligations, and partners can't see or manage the exposure. |
| Data & reporting | There's no reliable view of which jobs are waiting on client documents. Outstanding items live in spreadsheets, job notes and individual inboxes. | Jobs stall unnoticed, work bunches up before deadlines, and clients experience the firm as disorganised. |
| Systems & technology | The practice management system includes a basic file exchange, but few clients use it because it has no per-job checklists, reminders or status. | The firm already pays for a capability that doesn't remove the email chase. |
| AI readiness & opportunity | Much of the firm's knowledge is already written down, staff work in Microsoft 365 with existing permissions, and the questions juniors ask are repetitive. | A grounded assistant is feasible without a large data project, provided the content is cleaned up first. |
The moves
Every recommendation resolves to one of nine moves: build, buy, improve, integrate, automate, remove, change, train or leave alone. Building is only one of them.
- Change
Adopt an AI use policy before any tool goes live: approved tools, what may and may not be entered, human review of every client-facing output, a register of AI use, and client permission wording in engagement letters.
- Remove
Personal and consumer AI tools for firm work, replaced by an approved option. Duplicate and superseded templates are archived.
- Improve
The procedures manual and template library: one current version of each, with an owner, a review date and an access level, before anything is connected to AI.
- Buy
An enterprise AI service under terms that exclude training on the firm's data, with Australian data hosting where available. Building or hosting a model would add cost and risk for no benefit.
- Build
An internal knowledge assistant that answers from the firm's approved procedures, templates and precedents, cites the document and section, says when it doesn't know, and only shows what each person is allowed to see.
- Automate
Document collection through a client portal: checklists generated from the job type in the practice management system, secure upload, reminders at set intervals, and an alert to the manager when a job stalls.
- Train
Short role-based sessions on asking good questions, checking citations, what never goes into any AI tool, and how managers review AI-assisted work, with a champion in each office.
- Leave alone
The practice management system, the document management system and the tax and accounting software. They work, staff know them, and the new tools connect to them rather than replace them.
The system, end to end
The job in the practice management system drives the document request, and whatever the client uploads is filed against that job automatically. While staff work, the assistant answers from the firm's own approved procedures and templates, cites its source and shows only what that person is allowed to see, and partners still review the output.
- Job created in the practice management system
- Portal sends the client a document checklist
- Uploads filed against the job
- Staff prepare the work
- Knowledge assistant answers from approved sources, with citations
- Manager and partner review
- Client sees status in the portal
AI use policy and register
Approved tools, permitted and prohibited inputs, review rules for client-facing work, client permission wording, an incident process and a register of AI use, structured around the six essential practices in the National AI Centre's Guidance for AI Adoption.
Curated knowledge base
One current version of every procedure, template and precedent, each with an owner, review date and access level. Superseded material is archived, not deleted.
Knowledge assistant
Answers from the curated base only, cites the document and section, says when it doesn't know, and inherits folder permissions. Questions and answers are logged so content owners can fix gaps.
Client portal
Per-job document checklists generated from the job type, secure upload from a phone or computer, automatic reminders, engagement letter signing and a plain status for every job.
Integrations
The practice management system for clients and jobs, the document management system for filing, and Microsoft 365 for sign-in and permissions.
Training and champions
Role-based sessions, a one-page guide to the AI policy, and a champion in each office who gathers questions and feedback.
Delivery stages
Indicative timings.
AI policy and Diagnostic
Interview staff in each office, sample the questions juniors ask, review current tools and AI use, then write the AI policy and register and agree what the assistant may and may not do.
3 to 4 weeks (indicative)
Knowledge clean-up
Content owners confirm current versions while we restructure folders, tag access levels and archive duplicates.
4 to 8 weeks (indicative)
Assistant pilot
One team in one office uses the assistant. We test it against a set of known questions, measure accuracy and fix gaps in the source content before going wider.
4 to 6 weeks (indicative)
Client portal
Launched for one service line first, then extended, and timed to land outside the busiest lodgement months.
6 to 8 weeks (indicative)
Firm-wide rollout
All offices, champions in place, and a monthly review of logged questions, unanswered queries and policy exceptions.
4 to 6 weeks, then monthly (indicative)
The arithmetic
Assumptions first, then the calculations they produce. Change an assumption and the result changes with it, which is the point.
| Assumption | Value |
|---|---|
| Staff, excluding managers and partners, who regularly look for templates, procedures or policy answers | 44 (assumption) |
| Time each spends searching, or waiting for someone to answer | 2 hours a week (assumption) |
| Share of that time removed by an assistant that answers with citations | 40% (assumption) |
| Managers and partners answering routine questions | 16, each spending 2 hours a week, 30% of it avoidable with the assistant (assumption) |
| Jobs a year that need client documents collected | 1,800 (assumption) |
| Staff time chasing, renaming and filing documents per job | 45 minutes today, 15 minutes with the portal (assumption) |
| Working weeks a year | 46 (assumption) |
| Standard full-time week | 38 hours (assumption) |
Staff search time released
44 staff × 2 hours × 40% = 35.2 hours a week × 46 weeks = 1,619 hours
About 1,620 hours a year (modelled)
Manager and partner interruption time released
16 × 2 hours × 30% = 9.6 hours a week × 46 weeks = 442 hours
About 440 hours a year of senior time (modelled)
Document chasing time released
1,800 jobs × (45 − 15) min = 54,000 min ÷ 60
900 hours a year (modelled)
Capacity in full-time equivalents
1,619 + 442 + 900 = 2,961 hours ÷ (38 hours × 46 weeks = 1,748 hours)
About 1.7 full-time equivalents of capacity for client work and review. Capacity created, not roles removed (modelled)
What we'd measure
- Assistant answers given with a cited source, and the share staff rate as useful
- Questions the assistant couldn't answer, used to fix content gaps
- Days from document request to complete file, by job type
- Share of clients uploading through the portal rather than by email
- Staff-reported time spent searching, and interruptions to managers, surveyed quarterly
- AI policy acknowledgement and training completion, by office
Risks and how we'd manage them
Confidentiality and permissions
The assistant runs under enterprise AI terms that exclude training on firm data, inherits document permissions, and keeps client files out of the general knowledge base unless a partner approves. Usage is logged, and engagement letters cover the client's permission for AI-assisted work.
Wrong or outdated answers
The assistant answers only from curated sources, cites them and says when it doesn't know. Content owners keep the sources current, and staff remain responsible for checking anything that reaches a client, as the TPB's guidance expects of tax practitioners.
Low adoption
Tools that partners ignore don't stick. We pilot with a willing team, ask partners and managers to model use, appoint office champions, measure use monthly and keep the rollout clear of peak season.
Client uptake of the portal
Some clients will keep emailing. The portal works on a phone without an app, reminders link straight to the upload, and staff can upload on a client's behalf so every job still has one status.
The takeaway
The firm doesn't need cleverer AI first. It needs one current version of its own knowledge, clear rules about what AI may touch and a portal that ends the document chase. The assistant then makes that knowledge available to everyone, with a source a partner can check.
Services in this example
- AI assistants & knowledge systemsRole-specific AI assistants that answer from your own documents, cite the source and respect permissions. Built on Microsoft 365, Google or a custom stack.
- AI governance & implementationAI governance for Australian businesses: a use policy, approved tools, data rules and human oversight, informed by OAIC and National AI Centre guidance.
- Client & customer portalsSecure portals where customers check status, share documents, approve work, book and pay. Fewer 'where's my…' emails, with permissions and audit built in.
- AI-augmented workforceHelp your existing team do more valuable work with AI: role-based training, approved tools, prompt libraries and adoption tracking. Capacity, not cuts.
- AI strategy & roadmapsAn AI roadmap built on evidence: nine lenses, a do-nothing baseline, confidence-labelled estimates and a Now/Next/Later plan your board can act on.
Recognise your business in this?
A 30-minute call is enough to tell whether the same moves would pay off for you.
Sources
Third-party facts on this page, checked 14 October 2026.
- Confidentiality of client information, Tax Practitioners Board. www.tpb.gov.au/confidentiality-client-information
- TPB(GS) 55/2026 The use of Artificial Intelligence and the Code of Professional Conduct, Tax Practitioners Board. www.tpb.gov.au/tpbgs-552026-use-artificial-intelligence-and-code-professional-conduct
- Guidance on privacy and the use of commercially available AI products, Office of the Australian Information Commissioner. www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/guidance-on-privacy-and-the-use-of-commercially-available-ai-products
- Guidance for AI adoption: foundations, Department of Industry, Science and Resources (National AI Centre). www.industry.gov.au/publications/guidance-for-ai-adoption/guidance-ai-adoption-foundations