Most firms that have tried an AI assistant have tried the generic kind: a chat window bolted onto the side of a product, answering from whatever the underlying model happened to learn during training, with no connection to the firm's actual clients, jobs or deadlines. It is entertaining for five minutes and then a burden, because every answer has to be checked against the real record before anyone can act on it. The question a partner actually wants answered is rarely "what does a VAT return generally involve". It is "which of my clients has a VAT return due in the next ten days and hasn't sent us their sales ledger yet", and a model with no access to the firm's data cannot answer that at all, however fluently it writes.
Accupe's AI assistant is built the other way around. It lives in two places: the assistant panel that opens from the sparkles icon on every page, and the fuller AI Chat page reached from the sidebar. In both places it works against the firm's own jobs, clients and compliance records through a defined set of tools, rather than guessing from general knowledge. What makes it worth a closer look is not the chat interface, which looks much like every other chat interface, but the machinery underneath: what the assistant is actually allowed to do on its own, what it has to ask permission for first, and where the two surfaces genuinely differ.
Two Surfaces, Not One Assistant
The panel and the page look similar and are not the same tool, and firms that treat them as interchangeable miss half of what each one is for. The assistant panel, opened from the sparkles icon in the top bar, is available on every page in the product and always answers the same way: one model, no mode selector, no document upload. It is built for the question a staff member has mid-task, answered quickly, from the firm's live data, without leaving the page they are working on. The AI Chat page, reached from the sidebar and available on Growth and Elite plans, is the deeper surface. It adds document upload, a response mode selector (Fast, Planning or Ultra Detailed), a Docs Only toggle, and a scope selector for what the assistant is allowed to draw on. Choosing "Ask AI assistant" from the command palette actually opens the AI Chat page rather than the panel, a small distinction worth knowing before assuming the two behave the same way.
The plan gating behind the modes is real and enforced in two places. Fast and Planning are usable by anyone who can open the AI Chat page at all, which means anyone on Growth or Elite. Ultra Detailed is reserved for Elite. If a firm's plan does not include a mode, its pill in the selector shows a small lock icon, and clicking a locked mode never sends anything to the model; it opens an upgrade card instead, naming the mode and the plan that unlocks it. The same check runs again on the server as a backstop, so a stale or scripted request for a mode outside the firm's plan is declined server-side too, not merely hidden in the interface. Rate limits apply across both surfaces too, deliberately, so that one enthusiastic user asking the assistant fifty questions in ten minutes does not slow down everyone else's answers; the interface tells a user when they are close to a limit rather than letting them hit it without warning.
What the Assistant Can Actually Do, Not Just Say
The panel's real value is that it can act on the firm's data, not just describe it in the abstract. Ask it what's overdue this week and it does not summarise VAT deadlines in general terms; it calls a tool that queries the firm's own jobs and tasks and returns the actual list, with client names attached. The assistant has a defined set of tools it can call, each scoped narrowly enough that the model cannot wander outside what the firm has actually authorised it to touch. A read-only question never risks changing anything, which is precisely why the firm can let the assistant answer this category of question without a second thought.
It is tempting to focus on the tools that create or send something, because they are the more dramatic capability, but in practice the read-only tools above do most of the daily work. A manager who wants to know whether a particular client's KYC is current does not have to open the client record, scroll to the compliance tab and cross-reference the expiry date; get_client_details returns the answer in one exchange. A partner planning Monday's priorities can ask list_tasks scoped to the whole team rather than opening everyone's individual task lists in turn. None of this changes any record, so none of it needs a confirmation step, and that is deliberate: the assistant should never make a person stop and approve something that was never going to alter the firm's data in the first place. Reserving the confirmation gate for the tools that genuinely warrant it is what keeps it meaningful rather than becoming a click a user learns to approve without reading.
- search_clients: look up a client by business name or reference before acting on anything else
- get_client_details: pull a full view of one client, including open jobs, KYC verification status and Compliance Radar score
- list_jobs and list_tasks: filter active work firm-wide, by team member, or by client and due date
- lookup_companies_house: resolve a UK company number to its officers, filing deadlines and SIC codes
- get_firm_pulse: a live snapshot of active jobs, overdue items, this week's revenue and urgent inbox items
- calculate_tax: compute UK or UAE VAT, Corporation Tax or income tax from a given amount, bound to the firm's own region
Where the Assistant Has to Ask First
A second category of tool does not just read, it changes something the client or the firm will see, and those tools carry an explicit confirmation gate. create_task, create_job and apply_workflow_template create real records on the Smart Board. send_client_message and draft_signature_request go further still, because they put something in front of a client. Every one of these tools is built to the same two-step pattern: the first call, made without confirm set to true, returns a confirmation request describing exactly what is about to happen, which the interface shows to the person using the assistant. Only a second call, with confirm explicitly set to true after the person has approved it, actually creates the record or sends the message. There is no shortcut version of the tool that skips the confirmation and acts directly, whatever the wording of the request that triggered it.
This matters more than it sounds. An assistant that can search, summarise and calculate is a useful productivity tool on its own. An assistant that can also draft a client message, a signature request or a new job is materially more useful, and materially riskier, if it can act without a person in the loop. The confirm gate is the design decision that lets Accupe ship the more useful version without inheriting the risk of the more useful version. A partner can ask the assistant to draft a message to a client about a missing document, read exactly what it proposes to send, edit a line if the tone is wrong, and only then approve it, all before anything leaves the firm.
Docs Only Mode: When the Answer Has to Come From One Document
The AI Chat page's Docs Only toggle is a different kind of guardrail, built for a different kind of risk: not an unauthorised action, but an ungrounded answer. When Docs Only is switched on, the assistant is instructed to answer strictly from the documents uploaded to that conversation, an engagement letter, a scanned lease, a set of client-prepared accounts, and to end every factual sentence with a citation back to the specific document it came from. If no documents have been uploaded, it refuses outright rather than answering from general knowledge, and says so plainly rather than quietly filling the gap.
Before the answer ever reaches the screen, Accupe runs two automatic checks against it. The first counts how many of the answer's factual sentences carry a citation; if more than 40 percent arrive uncited, the reply is withheld and replaced with a message stating that the draft could not be verified. The second extracts the key names and figures the answer relies on and checks that they genuinely appear in the uploaded documents; if too many do not, the reply is withheld again, this time naming the specific terms it could not confirm, so the user knows exactly what to check rather than guessing. Either failure produces a blocked message, never a plausible-sounding but unverified answer. Docs Only is the right setting when accuracy against one specific source matters more than breadth, a contract clause, a scanned engagement letter, a figure buried in client-prepared accounts, and the wrong setting for the broader questions a firm wants answered from its live data as well as any attached documents.
The document handling underneath is honestly scoped rather than oversold. The AI Chat page reads PDF, Word, Excel, CSV and image files, and runs OCR automatically over the first three pages of a scanned PDF, English only. PowerPoint files and older binary .doc and .xls files are accepted on upload but never actually read, and are marked accordingly in the document tray rather than silently ignored, which matters because no amount of rephrasing a question fixes a document that produced no extractable text in the first place. A single message can draw on at most ten uploaded documents, and any one file above 500,000 extracted characters is rejected outright rather than partially processed. None of these are hidden limits a firm discovers by accident; they are stated so a team knows what the tool can and cannot read before relying on it for something time-sensitive.
Why the Guardrails Are the Point, Not an Afterthought
2026 has been the year professional bodies stopped treating AI use in practice as a side issue. The CCAB ethics group has been working through a statement to the profession on the ethical use of AI, and the Professional Conduct in Relation to Taxation guidance, the shared standard ICAEW and the other PCRT bodies publish, was updated this year to address AI use in tax work directly. The direction of travel across that guidance is consistent: firms are expected to be able to say what an AI tool actually did, on what basis, and with what oversight, rather than simply asserting that "the AI checked it".
The confirm gate and the Docs Only verification checks exist for exactly that expectation. A firm using Accupe's assistant can show that no client message or signature request left the building without a person explicitly approving the exact text, and that no Docs Only answer reached a user without first passing a citation check and a fact-coverage check. That is a materially easier conversation to have with a professional indemnity insurer, or with a practice assurance reviewer asking how the firm governs its use of AI, than simply pointing at a chat log and asserting that the output was reviewed.
A Realistic Morning
A manager opens a client file before a call and asks the panel to pull the client's current status. get_client_details returns the open jobs, the KYC verification state and the Compliance Radar score in one pass, so the manager walks into the call already knowing what is outstanding rather than tabbing between three pages to assemble it themselves. After the call, they ask the assistant to draft a message summarising the two documents still needed and send it to the client. The assistant returns the draft text first; the manager tweaks a sentence, approves, and only then does send_client_message actually fire. Later that day, a partner opens the panel and asks how the firm is tracking this week; get_firm_pulse returns active jobs, overdue items, and this week's revenue in one answer, the kind of question that used to mean opening four separate dashboards before the first coffee.
A different scenario the same afternoon: a new client has just signed, and a staff member asks the assistant to set them up on the firm's standard annual accounts workflow. apply_workflow_template is a confirm-gated tool for good reason, because it does not just create one job, it spins up the job and every templated task inside it with sequenced due dates in a single action. The assistant matches the template by name, shows what it is about to create, the client, the template, the starting date, before touching anything, and only proceeds once the staff member confirms. What would otherwise be several minutes of manually replicating a template happens in one exchange, with the same review step a person would want even if they were building the job by hand.
What It Does Not Do
The assistant does not file anything with HMRC, the FTA or Companies House; it drafts, summarises and, once approved, actions items inside Accupe. It does not replace the technical review a qualified team member gives to a tax position or an accounts figure, calculate_tax returns a structured breakdown for reference, not a filed computation the firm can submit unchecked. It has no access to anything outside the tools listed above, meaning it cannot read a client's live bank feed, browse the open internet, or see another firm's data on the platform. And every tool call, confirmed or not, is bound to the calling firm's own data by the caller's identity; there is no version of the assistant that reaches across firms, however it is asked. The underlying model itself is also deliberately tiered: the fast, everyday path behind the panel and Fast mode favours speed for the volume of small questions a firm asks in a day, while Planning and Ultra Detailed switch to a slower, deeper reasoning pass, which is exactly why those two are gated to paid plans rather than given away for every query regardless of how much thinking it actually needs.
Closing
An AI assistant earns its place in a regulated practice by being predictable under scrutiny, not by being clever in a demo. Accupe's version reads the firm's actual data, proposes actions rather than taking them unilaterally wherever a client is on the other end, and can show afterwards exactly what it was allowed to do and when a person signed off on it. The assistant does the retrieving and the drafting; the firm still decides what actually goes out the door.
