AI for Professional Services

Treat that as a no. A consumer account is personal rather than the firm’s, there is normally no processing agreement between the vendor and your practice, retention is set by the vendor’s defaults, and you cannot answer a client asking where their data went. If the capability is genuinely needed, procure a tool properly rather than writing a policy that asks people to resist a convenient habit.

It is a fluent, confident, wrong answer. You do not prevent it at the model level; you prevent it at the system level. Restrict the answer to a defined document set, require a citation on each claim, refuse rather than guess when the source does not contain the answer, and check the extracted text when the source was a scan. Most wrong answers in a firm trace back to an empty extraction rather than to invention.

Almost certainly because it never read the document. A scan is an image, so a system reading the text layer of the PDF gets nothing, and a model given nothing will still answer. Check whether the tool ran optical character recognition, how many pages it processed, and what languages it recognises, and read the extracted text before relying on the answer.

It changes the mix of work rather than removing the role. The tasks it absorbs first are the ones juniors currently learn on, which creates a real training problem: if nobody retypes a set of figures, nobody develops the instinct that tells them when a figure looks wrong. Firms handling this well are deliberate about how juniors now build that judgement, rather than assuming it still arrives by itself.

Yes, and it can be one page. Approved tools, what may and may not be uploaded, what always requires human review, who to ask, and what to do if something goes wrong. Professional bodies and insurers increasingly expect one, and more usefully it gives your team a clear answer at the moment they are deciding whether to paste something into a browser tab.

No, and any tool claiming otherwise is describing something a regulator will treat as your firm’s work regardless. AI can draft, extract and look up. The treatment, the conclusion and the signature stay with a qualified person, and your engagement letter almost certainly already says so.

Be straightforward about it. Most clients are unbothered by a firm using software to extract data or draft a first version, and considerably more bothered by discovering it later. If your engagement letter is silent on subcontracting and processing, that is worth reviewing anyway, and it is a natural place to cover this.

It answers questions grounded in your own firm data such as clients, jobs, deadlines and documents, extracts data from uploaded documents with OCR so scanned PDFs are read rather than skipped, and offers a docs-only mode that restricts answers to your uploaded documents with citations and withholds an answer it cannot ground. It respects the same roles and permissions as the person asking, and asks for confirmation before anything with a real-world effect. It does not file returns and it does not give advice.

Treat it as an unanswered question rather than a soft answer. Accupe blocks a response when too little of it is supported by the documents in scope, and that block is the feature, not a limitation to work around. The failure mode worth designing against is the plausible, fluent, uncited paragraph that a busy reviewer accepts because it reads like the others. A tool that declines is safer than one that guesses, and it is far easier to defend to a professional indemnity insurer.

On scans and on formats the extractor does not read. OCR handles the first few pages of a scanned PDF in English, so a long scanned lease or a document in another language will come back partial. Presentation and legacy Office formats are not extracted at all. The practical discipline is to check that the extraction covered what you think it covered before relying on a figure, because a partial extraction looks identical to a complete one until you reconcile it against the source.

The person who signs it, exactly as before. AI changes the drafting cost, not the professional duty, and no regulator or insurer recognises a distinction between a document a junior drafted and one a model drafted. The workable policy is narrow and written down: which tasks may use it, which client data may go near it, and what a reviewer must check before their name goes on the output.

Start with one task, one reviewer and a written note of what was checked. The usual failure is the opposite: a firm-wide policy drafted in the abstract, agreed by nobody in particular, and quietly ignored within a month because it does not match how anyone actually works. Pick something narrow and verifiable, such as extracting figures from a document set where a human already reconciles the output, and run it for a few weeks. You will learn where it breaks on your own files, which is the only evidence worth writing a policy from. Two rules are worth having on day one regardless of scale: no client data goes into a tool the firm has not assessed, and a named person signs off anything that leaves the building. Everything else can be written once you know what you are describing.

Five, and the answers should be in writing rather than in a sales call. Where is the data processed and stored, and does that location satisfy the firm’s own engagement terms. Is client content used to train the provider’s models, and can that be switched off contractually rather than in a settings panel. How long is content retained after a session ends, and what does deletion actually mean. Who inside the provider can read the content, and under what controls. What happens to the firm’s data if the contract ends or the provider is acquired. A vendor who answers these crisply has thought about regulated customers; a vendor who deflects to a generic security page has not. None of this is exotic diligence, and it is the same set a firm would apply to any subcontractor handling client papers.

Yes, where the task is how somebody learns judgement. Reading a full set of accounts, or working through a client’s bank statements line by line, is slow and largely automatable, and it is also how a trainee builds the instinct that later lets them notice something is wrong in thirty seconds. A practice that automates every entry-level task efficiently removes the ladder its own senior people climbed. That is not an argument against the tooling, it is an argument for being deliberate about which work stays manual and why, and for saying so out loud rather than letting it erode by default.