Yes, in most practices. Spend two to four weeks switching on and testing the AI inside the practice management, accounting and office software you use now, then buy only for the jobs it cannot do. The exception is when the built-in AI can't reach the data a job needs, or its data terms don't meet your client obligations.
Built-in AI has one big advantage and one big limit. The advantage: it already sits next to your client records, it inherits the permissions you set up, and it is often included in the price. The limit: it only sees its own system. An assistant inside your practice management tool can summarise a client's emails but cannot read the drawings on your file server. Knowing which side of that line each job falls on is most of the decision.
Stage 1: List the AI each subscription already includes (1-2 hours)
Write down every system the practice pays for, then check each vendor's pricing page and release notes for AI features. Several have added them at no extra charge. Examples verified in September 2026:
| Software | AI it includes | Worth knowing |
|---|---|---|
| Karbon (practice management for accountants) | Email and client summaries, compose-and-refine drafting, quick replies, emails drafted from tasks, suggested assignees | Included at no extra cost on paid plans for now; Karbon says it will give notice if that changes, and states firm data is never used to train AI models |
| Xero | JAX, its AI assistant: bank reconciliation suggestions, document capture, insights | Shown across Xero's plans; Xero notes that future features may carry extra fees. Its built-in Smart Document Capture for receipts and bills was announced in July 2026 as free on all plans |
| QuickBooks Online | Intuit's AI agents for accounting, payments, finance, project management and more | Which agents you get depends on your plan, and Intuit's help page says AI features can't currently be switched off individually |
| Deltek Ajera or Vantagepoint (architecture and engineering) | Ask Dela, a plain-language assistant for project, client and staff data | Check whether it is enabled for your edition |
| Microsoft 365 business plans | Copilot Chat, web-grounded drafting and research | Grounded in the web, and it can now work on the Outlook email or file you have open; reasoning across your mail, meetings and SharePoint files together needs the paid Microsoft 365 Copilot |
| Google Workspace | Gemini in Gmail and the Gemini app on every business plan; Docs, Sheets, Meet and Drive from Business Standard up | Confirm on your Workspace admin console what your plan includes |
| SketchUp or Revit | SketchUp AI Render and assistant (monthly credits in Go, Pro and Studio); Autodesk Assistant in Revit 2027 as a tech preview | Previews change; don't build a process on them yet |
The inventory usually turns up at least one overlap. At an illustrative six-person accountancy practice on Karbon and Xero, four staff were each expensing a $20 personal assistant plan to draft client emails, $80 a month in total, while Karbon's included drafting sat unused because nobody had opened it. That is $960 a year spent on a job the practice software already did, with client emails passing through personal accounts on top.
Legal practice systems more often sell AI as a paid add-on, so check your contract and the vendor's current price list rather than assuming. AI features already in your software covers general business tools such as CRMs and helpdesks in the same way.
Stage 2: Check settings and data terms first (an hour per system)
Built-in does not automatically mean approved. For each feature you plan to use, answer five questions and write the answers down:
- Is it on, and who switched it on? Some features arrive enabled for everyone. Find the admin setting and decide who should have it.
- What data does it read? One client's record, or everything the user can see? Summaries of a client that pull in another client's emails from a shared thread are a confidentiality problem.
- Which AI provider sits underneath, and does it train on your data? Karbon, for instance, names Azure OpenAI Service and says firm data isn't used for training. If a vendor's page says nothing, ask them.
- Can you switch it off? If not, as with QuickBooks Online's AI features at present, your policy has to cover how staff review what the AI suggests.
- Is it covered by your existing contract? An AI feature added mid-contract may come with new terms. Read them; what to check in an AI tool's privacy policy and terms lists the clauses that matter.
Here are the five answers written down for one system at the six-person accountancy practice (illustrative):
SYSTEM: Karbon, client/email summaries and drafting
1. On? Yes, for all six users. Admin: practice manager.
Decision: keep on; junior staff's drafts checked by a manager.
2. Reads: the client's timeline, including shared email threads.
Risk: joint threads for couples with separate engagements.
3. Provider: Azure OpenAI Service (Karbon's AI page).
Training on firm data: no, per the same page.
4. Can we switch it off? Yes, admin setting found. Not needed.
5. Contract: included under current terms; Karbon says it will give
notice before charging. Recheck at renewal: [month].
Signed off by: [partner], [date]Question 2 matters most with linked clients. Couples, family companies and a director's own return often sit in overlapping threads. In this practice's test, the summary for one spouse's return mentioned the other spouse's dividend figure, because both were copied on the same chain. Nothing left the firm, but a draft built on that summary could have. The fix was procedural: separate threads for separately engaged clients, and a human read of any linked client's summary before a reply goes out.
When a vendor's page is silent on question 3, send support something like this and keep the reply on file:
Subject: AI features and our client data
We use [product] for client work and want to confirm three points
before enabling [feature name]:
1. Which AI provider processes our data for this feature?
2. Is any of our data, or our clients' data, used to train or
improve AI models, yours or the provider's?
3. How long are prompts and outputs kept, and where is this set
out in our contract or your terms?
A link to the relevant documentation is fine.A reply that says "your data is secure and we take privacy seriously" answers none of the three. Write back once asking for each point by number. If the second reply still names no provider and no training position, keep the feature off until it does.
Stage 3: Match the features to your three slowest jobs (30 minutes)
Features are only useful against real work. Ask the team which three recurring jobs eat the most time, then see which built-in feature, if any, touches each. Here is that table filled in for an illustrative nine-person engineering consultancy that runs projects and fees in Deltek Ajera and works in Microsoft 365 Business Standard:
| Slow job | Hours a week (whole firm) | Built-in feature that touches it | Gap? |
|---|---|---|---|
| Chasing which projects are over fee budget or have unbilled time | 4 | Ask Dela in Ajera | Test it |
| Writing fee proposals and scope letters | 6 | Copilot Chat (drafting only; not grounded in past proposals) | Partly |
| Turning site-visit notes into short inspection reports | 8 | None: notes are in phone photos and a notebook app | Yes |
The firm had been about to buy ChatGPT Business for all nine staff. The table showed that one of the three jobs was covered, one was half covered and one needed something new, which is a much smaller purchase.
Stage 4: Run a two-week test with a scorecard
For each feature, pick two people who do that job every week and have them use the AI on real work for two weeks. They log each use in a shared sheet: the task, minutes before, minutes with AI, and whether the output needed correcting.
A realistic test of Ask Dela looked like this. The finance lead typed a question in plain language:
Which active projects have more than 90% of the fee used but less
than 75% of the work complete?The illustrative reply listed six projects with fee used and percentage complete. Four were right. Two were wrong because project managers hadn't updated percent-complete in Ajera for a month, so the assistant was faithfully reporting stale figures. The AI worked; the data didn't. That is a common finding in stage 4, and fixing the monthly update habit was worth more than any new tool.
Copilot Chat on proposals showed the other kind of limit. Asked to "draft a scope letter for a structural inspection of a two-storey office with suspected roof timber decay", it produced a tidy letter with sensible headings, generic exclusions and square-bracket placeholders for the fee, programme and access arrangements. Useful as a starting point, but it knew nothing about how the firm words its own exclusions, because Copilot Chat works from the web and whatever file or email is open, not from the firm's SharePoint library of past proposals. That became a clear gap.
The filled-in scorecard after two weeks:
| Feature | Uses logged | Minutes saved per use | Needed correcting | Verdict |
|---|---|---|---|---|
| Ask Dela, budget questions | 14 | About 12 | 3 of 14 (stale data) | Keep; fix monthly updates |
| Copilot Chat, proposal drafts | 9 | About 20 | 9 of 9 (house wording missing) | Keep for structure; gap for house style |
| Copilot Chat, site notes to report | 5 | About 5 | 5 of 5 (notes had to be retyped first) | Gap |
If this is your first structured test, running your first AI pilot project covers the set-up in more detail.
Stage 5: Keep it, ignore it, or fill the gap
| What the test showed | Do this |
|---|---|
| Saves time, output mostly right | Keep. Write a short how-to and add it to your AI rules |
| Output wrong because your data is stale or messy | Fix the data habit first, retest in a month. Don't buy another tool to work around it |
| Works, but can't see the files or context the job needs | Genuine gap. Price the smallest thing that fills it, for the people who do that job |
| Rarely used, or quality too low to trust | Switch it off, so nobody relies on it by accident |
| Data terms don't meet your client obligations | Switch it off and tell staff why |
The engineering consultancy ended up buying three Microsoft 365 Copilot Business seats (the bundle with Business Standard is $23.50 per user a month on annual billing, which it compared against its existing $14 Business Standard licences) for the two directors and the senior engineer who write proposals, so the assistant could draft from the firm's own past proposals in SharePoint. For site notes it tested dictation into a phone app followed by a Copilot draft. Nine ChatGPT seats became three Copilot upgrades, costing about $28.50 a month more than before instead of about $225.
A decision like that needs checking after a quarter, not just making. Three months on, the consultancy re-ran the stage 4 log for a fortnight on the features it kept. Ask Dela's corrections fell from 3 in 14 uses to 1 in 16, because project managers now updated percent-complete on the last Friday of each month. The Copilot proposal drafts needed only wording tweaks in 7 of 10 cases, against a full house-style rewrite every time before. If the second log had looked like the first, the three seats would have gone at renewal.
The same five stages at a six-person accountancy practice
The same five stages at a six-person accountancy practice on Karbon and Xero usually find more already covered. Before a client call, a manager used to spend ten minutes scrolling a client's timeline to remember where things stood. With Karbon's client summary, an illustrative brief reads: "Year-end accounts in review, two queries outstanding on director's loan balance, client emailed 12 Sept asking about payment plan for the tax bill, invoice for bookkeeping 30 days overdue." Thirty seconds to read.
What it missed: the client had phoned on 15 September about the payment plan and a partner had already agreed it, but the call wasn't logged in Karbon. The summary was accurate to the system and wrong about the world. The practice's fix was a rule, not a tool: log calls in two lines on the day. Built-in AI is only as good as what gets recorded where it can see it.
The practice had pencilled in six team-plan assistant seats at $25 each, $150 a month, before starting. After the five stages it bought two, for the partners' advisory letters, where house style mattered and Karbon's drafts fell short: $50 a month. Add the four personal plans it cancelled and the practice spends about $180 a month less than it would have, and client emails no longer pass through personal accounts.
Where built-in AI runs out
- It can't join systems. "Which clients with overdue invoices also have a deadline this month?" needs data from two tools. That is an automation or integration job, not a feature toggle. Whether AI can work with the tools you already use explains the options. Until then, a monthly workaround does the job: export overdue invoices from the accounting system and this month's deadlines from practice management as two files, give both to a business assistant, and ask for the clients that appear in both, with the amount and the deadline. Check the match by hand the first time. Client names often differ between systems ("Ltd" in one and "Limited" in the other, or a trading name in one and the company name in the other), and the assistant can miss a pair or merge two different clients.
- It rarely learns your house style. Some tools, Karbon among them, adapt drafts to your firm's writing; many don't. Letters and reports that must sound like the practice may need a shared Project in a business chat assistant, loaded with your approved examples.
- It can change without warning. Features move between plans, previews end, free inclusions become paid, and privacy defaults move too. A real case from therapy practice software: since 16 June 2026, new users of SimplePractice's Note Taker are opted in by default to the vendor keeping de-identified transcripts. Nothing broke; the default simply changed. Recheck stage 1 and your stage 2 answers at each renewal.
- It can make you dependent on one vendor. The more of your process lives inside one product's AI, the harder it is to switch. Keep your prompts, templates and procedures in documents you own.
One realistic mistake to avoid: accepting AI suggestions in bulk because they came from a trusted system. A practice that clicks "accept all" on suggested bank categorisations for a new client can end up with a recurring supplier coded wrongly for three months before a review spots it. Built-in suggestions deserve the same review as any other AI output, at least until your test shows a feature is reliable for that client.
Further reads
- How Much Does AI Cost a Small Accounting Firm? — What a practice spends once the gaps are filled.
- Best AI Tools for Small Accounting Firms, Sorted by Task (2026) — Tools by task, for the jobs built-in AI can't do.
- What Gemini in Google Workspace Includes on Each Business Plan — What Gemini includes if your office runs on Google Workspace.
- AI Vendor Lock-In: How to Keep Your Data and Prompts Portable — Keep prompts and data portable if built-in AI ties you to one vendor.
- How to Measure Time Saved After Rolling Out AI in a Small Firm — Measure the test properly before deciding to buy.
- Per-Seat SaaS Fees vs One Custom Tool: Five-Year Costs Compared — When stacking per-seat AI add-ons stops making sense.
- How to Chase Missing Client Records With AI Before Deadlines — Build a missing-items list, send AI-written chasers on a ladder counted back from each deadline, and let AI sort the replies for you.
- AI Receipt Capture and Bank Categorisation: A Bookkeeper's Setup — Set up receipt capture and bank coding in six stages per client, with supplier rules, ledger AI suggestions and a weekly review that catches duplicates.
- AI Time Capture for Solicitors: Recover Hours You Never Billed — How passive AI time capture works for solicitors, the three ways a small firm can start, a daily review routine and honest arithmetic on what it recovers.
- AI Scribe Costs Compared: What Small Clinics Pay in 2026 — Published AI scribe prices side by side, yearly budgets for three clinic sizes, cost per note, and the extra costs and contract terms that change the total.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Karbon AI feature page, pricing page and March 2026 release notes; Xero JAX page; Intuit help article on AI agents in QuickBooks Online; Deltek pages on Ask Dela for Ajera and Vantagepoint; Microsoft 365 business plan pages; SketchUp AI subscription help page; Autodesk Revit 2027 What's New help page. Checked September 2026.