Most AI tools marketed to accountants present a chat box and hope the user will trust the output. Accupe's docs-only AI mode is built on the opposite premise - that for a regulated professional services firm, the tool's job is to answer strictly from the documents the firm has uploaded, to cite each statement back to its source passage, and to refuse cleanly when the answer is not there. Hallucination is not mitigated; it is designed out.
This piece is a walkthrough of how the docs-only mode is engineered, what it does and does not answer, how it fits with the three operating modes (Fast, Planning, Ultra-Detailed), and what a partner needs to know to deploy it confidently inside the working paper workflow.
What docs-only mode actually does
In docs-only mode, the AI has access to one thing - the documents the firm has uploaded for the question at hand. It does not have access to the open web, it does not draw on the model's general training knowledge for substantive claims, and it does not invent context to fill gaps. When the user asks a question, the system retrieves the most relevant passages from the uploaded documents, passes them to the language model with the user's question, and instructs the model to answer only from those passages, with a citation for every statement.
When the documents do not support an answer, the model says so. It does not approximate. It does not extrapolate. It does not produce a plausible-sounding response that might be right. The refusal is a feature, not a failure, and the workflow is designed so that a refusal is the prompt for the user to either upload more material or to take the question to a different channel.
The three modes and where docs-only fits
Accupe's AI runs in three modes - Fast, Planning, and Ultra-Detailed - each suited to a different kind of work:
- Fast - quick first-pass questions, extraction of standard data points, short summaries; minimises latency and works well for the high-volume routine questions
- Planning - structured drafting work, research note skeletons, file note structures; produces output in the shape the working paper file expects
- Ultra-Detailed - deep document interrogation, complex cross-document analysis, the kind of question that needs the model to do real work; takes longer but produces the most thorough output
Source citation as a first-class output
Every statement in a docs-only response is accompanied by a citation to the specific passage in the specific document that supports it. The citation is not an afterthought at the end of the answer; it is attached to each claim so the reviewer can verify the chain at the sentence level.
In practice this changes how the reviewer works with the output. The question is no longer "do I believe the AI?" - it is "does the cited passage actually say what the AI says it says?". A reviewer can verify a multi-paragraph answer in under a minute by spot-checking three or four citations, which is the workflow that makes AI productive rather than risky.
Client-context awareness
Docs-only mode is also client-context aware. When the user is working inside a client record, the AI has access to the documents associated with that client - the engagement letter, the prior year accounts, the AML file, the correspondence, the working papers - and answers questions in the context of that specific engagement. When the user switches to a different client, the context switches with them.
This matters because the most useful AI questions in a practice are client-specific. "What did we agree about the year-end timetable for this client?" is a question that depends on the engagement letter and the recent correspondence; "what did we conclude about the deferred tax position last year?" depends on the prior year working papers. The AI answers these because it has the client's actual documents in scope.
How the refusal pathway is engineered
The refusal pathway is the single most important engineering decision in a docs-only system. It is built in three layers. First, the retrieval layer returns a confidence score on the passages it finds; below a threshold, the system treats the question as unanswerable from the corpus. Second, the model is instructed in its system prompt to refuse rather than guess when the retrieved passages do not support an answer. Third, the user interface treats refusals as a normal output category, not as an error, so users get comfortable with them rather than rephrasing the question until something appears.
This three-layer refusal is what distinguishes a tool that is genuinely safe for regulated work from a tool that merely claims to be safe. Without it, users learn to extract answers from a system that should be refusing, and the protection is lost.
What docs-only mode does not do
It is worth being explicit about what docs-only mode does not do, so the team's expectations are calibrated. It does not give a tax opinion based on legislation not in the uploaded corpus. It does not give a regulatory answer based on rules that the user has not provided. It does not give a general accounting standards interpretation unless the standard text has been uploaded. It does not browse the web, query Companies House live, or pull HMRC manual pages on demand.
These are deliberate boundaries. A tool that answers from documents the firm has placed in scope is auditable in a way a tool that draws from anywhere is not. The firm decides what corpus to expose, and the AI works within that corpus.
Integration with the working paper workflow
A docs-only answer is not a free-floating chatbot response; it sits inside the client record and feeds the working paper file directly. The prompt, the documents in scope, the AI output, and the citations are captured as artefacts against the engagement. When the reviewer signs off, the sign-off is captured as well. The working paper file is therefore reconstructible after the fact, which is the standard ICAEW, ACCA, and UAE auditor inspections expect.
This is the operational benefit of the AI being part of the practice management platform rather than a separate tool. The team does not have to remember to document the AI use; the workflow does it.
What this means for the firm's AI policy
A firm that has Accupe's docs-only mode on its approved tool list can give a clean answer to the questions an insurer or regulator is most likely to ask. The data stays inside the firm's processing boundary. The output is source-cited. The tool refuses cleanly when the answer is not in the documents. The audit trail is captured in the working paper file. The contractual basis for processing is documented.
These are the questions that consumer chatbots cannot answer well. They are the questions the firm wants its AI policy to be able to answer in writing at the next professional indemnity renewal.
How it sits alongside the rest of the platform
The docs-only AI mode is one component of the wider practice management platform. It sits alongside Smart Boards for visual job management, AML/KYC screening via OpenSanctions, Companies House integration, an encrypted client portal with structured document request lists, built-in e-signatures, time tracking, and Compliance Radar. The AI works on documents that arrived through the portal, supports the work that is moving across the Smart Board, and produces output that lands in the working paper file the partner reviews.
The integration is what turns AI from a productivity novelty into an operational asset. The firm does not run AI as a separate tool; it runs AI as part of how the practice operates.
How to evaluate it in your firm
The right evaluation is the small acceptance test that works for any AI tool. Pick five real client documents from different engagement types, prepare five questions you know the answer to, and ask the tool in docs-only mode. Look at whether the answer is right, whether the citations are right, whether the tool refuses cleanly when it should, and whether the output is reviewable in a sensible amount of time. A tool that passes this test in the office will behave well in production.
Accupe's docs-only mode is designed to pass this test. The per-firm pricing from £20/month means the cost of evaluation is essentially zero, and the cost of running the tool across the whole firm is small relative to the productivity it returns.
How Accupe helps
Accupe is the practice management platform built around exactly this AI design - docs-only mode with three operating gears (Fast, Planning, Ultra-Detailed), client-context awareness, per-statement source citation, a clean refusal pathway, and full integration with the working paper workflow, Smart Boards, AML/KYC via OpenSanctions, the encrypted client portal, and built-in e-signatures. Per-firm pricing from £20/month makes it accessible from the smallest practice upwards.