If your business has no rules for AI yet, it is worth checking whether staff are already using personal AI accounts for work. The Office of the Australian Information Commissioner (OAIC) built one of its own worked examples around exactly that: a fictional company whose employee uploaded a customer's hardship application to a publicly available AI tool.
The useful question is no longer whether your team will use AI assistants. It is whether they use them on the right work, with the right information and the right checks. This article covers where role-specific assistants earn their place, where they don't, how they find answers in your own documents, and a 30-60-90 day way to start without creating a privacy problem.
What is a role-specific AI assistant?
A role-specific AI assistant is a generative AI tool set up for one job, such as a proposal writer, a service coordinator or an accounts officer. It works from approved instructions, templates and a defined set of your documents, and it produces drafts, answers and summaries that a person checks before they are used. It supports the person in the role. It does not make their decisions.
The "role-specific" part matters. A general chat tool knows a lot about the world and nothing about your pricing rules, your scope templates or how your team writes to customers. An assistant built for a role carries that context, so staff spend less time explaining and more time checking.
Where assistants genuinely help
The pattern is consistent. Assistants are strong where the work has a known shape, the source material exists, and a competent person can check the result in less time than it would take to produce it.
Drafting from a known pattern
Proposals, scope documents, client update emails, job notes, position descriptions and policy drafts all follow patterns. Give an assistant your templates, two or three strong past examples and the facts of the case, and it produces a first draft in minutes. The value is in a faster first draft, not in skipping review.
Finding answers in your own documents
Staff lose time hunting through shared drives, procedure manuals and old email threads, or interrupting the one person who knows. An assistant connected to an approved document library can answer "what is our process for a warranty claim?" and show which document the answer came from.
Summarising long material
Meeting transcripts, tender documents, contracts for a first read, customer feedback and long email chains can be condensed into a page of key points, open questions and actions. The reader still decides what matters.
First-pass analysis
Sorting survey responses, finding themes in support tickets or comparing three supplier quotes against a checklist are good first-pass jobs. Treat the output as a starting hypothesis to test, not a finding.
| Task | Good fit when | What a person checks |
|---|---|---|
| Proposal or scope draft | You have templates and strong past examples | Pricing, commitments and client facts |
| Policy or procedure question | The answer exists in a current, approved document | The cited source, whenever the answer matters |
| Meeting or document summary | The source is complete and readable | Actions, owners and anything contentious |
| First-pass analysis | Categories or criteria are defined in advance | A sample of items, checked by hand |
We're building our own inbox triage on an AI system, so these are patterns we work with ourselves, not theory.
Where they don't help, and why
Judgement calls
Whether to extend terms to a struggling customer, how to handle a complaint from a long-standing client, which candidate to hire, whether a job is safe to start. These depend on context, accountability and values that sit with people. An assistant can gather the facts and draft options. It should not make the call, and staff should not feel they can hide behind it.
Facts the assistant cannot verify
Generative AI produces fluent text whether or not it is correct. The OAIC's guidance on commercially available AI products warns about inaccurate outputs and suggests your records make clear which information is AI output, a probabilistic assessment rather than fact. In practice: if an answer can't be traced to a source document, a system record or a person who knows, it doesn't go to a customer.
Sensitive data without controls
The OAIC recommends, as best practice, that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools. Health details, hardship applications, HR matters and client files need an enterprise tool where you have restricted the provider's use of your data for training, limited retention where possible and controlled who has access. If you can't show those settings, the data stays out.
How retrieval over your own documents works, in plain English
Most useful business assistants rely on a method called retrieval-augmented generation, usually shortened to RAG. The model is not retrained on your documents. It looks things up before it answers.
- Prepare the library. You choose which documents the assistant may use: procedures, templates, product information and approved past work. Drafts and out-of-date material come out.
- Index it. The documents are split into passages and converted into numerical representations, so they can be searched by meaning rather than exact keywords.
- Retrieve. When someone asks a question, the question is converted the same way and matched against the library to find the most relevant passages.
- Answer with sources. Those passages are added to the question, and the model writes its answer from them, citing which document each point came from.
- Keep it current. When a procedure changes, the library is updated, so the answers change with it.
Two practical consequences follow. First, permissions matter. Microsoft, for example, says Microsoft 365 Copilot only surfaces organisational data that each user already has at least view permission for. That is sensible, but it means any folder that is overshared today becomes easy to find through the assistant tomorrow. Fix permissions before you switch it on.
Document quality beats model choice
Second, the library sets the ceiling. If three versions of your leave policy exist, the assistant may quote any of them. In our view, poor answers more often trace back to the library than to the model: duplicates, outdated procedures, scanned files it can't read, or knowledge that was never written down. Budget time for cleaning source material. It is the least glamorous part of the work and usually the most valuable.
A 30-60-90 day way to start safely
Days 1 to 30: one role, one library, written rules
- Pick one role with high-volume, pattern-based work and a manager who wants it to succeed.
- List the five to ten tasks the assistant will help with, and the tasks it will not touch.
- Choose an enterprise-grade tool, restrict the provider's use of your data for training, and record where data is stored.
- Assemble and clean the document library. Remove drafts, duplicates and personal information the role doesn't need.
- Write a one-page use policy. Our practical guide to AI governance in Australia has an outline.
- Measure the baseline: how long first drafts take now, and how often they are reworked.
Days 31 to 60: pilot with a small group
- Run the assistant with three to eight people in that role.
- Log every answer that was wrong, unsupported or unhelpful, and fix the document or instruction behind it.
- Hold a short weekly review: what saved time, what created rework, what nobody used.
- Keep a human approval step on anything that leaves the business.
Days 61 to 90: decide, then expand or stop
- Compare the pilot measures with the baseline.
- Choose one of three outcomes: expand across the role, add a second role, or stop. Stopping is a legitimate result if the work was not a good fit.
- Name an owner for the library, the instructions and the measures, so the assistant doesn't decay once the project team moves on.
How to measure value
Usage counts tell you people opened the tool. They don't tell you whether the work got better. Track a small set of measures before and after the pilot.
| Measure | How to capture it | What it tells you |
|---|---|---|
| Time to first draft | Time from brief received to draft ready, sampled weekly | Whether drafting time is actually falling |
| Rework rate | Share of drafts needing substantial changes at review | Whether quality holds, or effort has just moved to reviewers |
| Answer accuracy | A monthly sample of 20 answers checked against their sources | Whether the library and instructions are working |
| Adoption | Share of the role using it weekly for the agreed tasks | Whether it fits real work or sits unused |
| Unanswered questions | Questions the assistant could not answer from the library | Gaps in your documentation, worth fixing anyway |
A worked example of the arithmetic
Every number below is an assumption for illustration. Replace them with your own baseline.
- Assumption: 8 people work in the role.
- Assumption: each produces 6 first drafts a week.
- Assumption: time to first draft falls from 45 minutes to 25 minutes, with review time unchanged.
- Assumption: 46 working weeks a year.
Capacity created: 8 people × 6 drafts × 20 minutes saved = 960 minutes, or 16 hours a week. Over 46 weeks, that is 736 hours a year. In our Altitude Diagnostic this figure would carry the label "modelled" until the pilot measures it. We report it as hours released for higher-value work, such as client contact or business development, rather than jobs removed. If the rework rate rises at the same time, part of the saving has moved to reviewers and the real figure is lower.
Where to go from here
For a quick read on where your business stands, start with the AI readiness assessment. If you already know the role and the documents, our AI assistants & knowledge systems work covers library preparation, the assistant itself and the measures. AI-augmented workforce is the broader program for lifting capability across several roles, and AI governance & implementation puts the policy, tools register and oversight rules in place.
To see how this fits together in an accounting and advisory firm, read the professional firm knowledge assistant worked example. It is an illustration with labelled assumptions, not a client case study.