A plain reading of what Dext publishes about AI Assist, and the five things a Canadian practice should settle before turning it on across a book of clients.
Key takeaways
Dext AI Assist is an agent inside Dext that learns your firm's bookkeeping decisions and turns them into automation applied across clients. On Dext's own product page for the feature, the described sequence is: it learns from every correction, category change and tax rule your team applies; it generates guidance reflecting your firm's standards; it explains each suggestion; and your team decides whether to review, refine, reject or auto-apply.
It is, in other words, a way of propagating one firm's judgment across a book of clients — not a way of replacing that judgment. Everything worth arguing about is in that distinction.
The same page lists the categories it targets: applying tax treatment consistently across clients, categorising transactions using document context and the firm's historical decisions rather than static rules, automating everyday steps such as reviewing transactions and processing documents, remembering client-specific rules so they are applied to future documents, and detecting missing or incorrect document fields to correct or flag them.
Note what those have in common. Each is a repetition of a decision the firm has already made. None is a new decision. That is the honest reading of the feature, and it is also why the deployment question for a practice is about the quality and consistency of the decisions being propagated, not about the model.
Firms conflate these three and then evaluate the wrong one.
Capture is the extraction layer. Dext's capture page describes submission by mobile photo, upload, email or invoice fetch, extraction of supplier, date, tax and total, and line item extraction, alongside a vendor accuracy claim of 99.9% — Dext's own figure for its own document mix.
Data Health is a monitoring layer. Dext's Data Health page describes reviewing client accounting data from Xero or QuickBooks Online daily, flagging issues such as unreconciled bank transactions, duplicate or missing contacts and overdue invoices and bills, and giving each client a percentage health score that moves as issues are resolved.
AI Assist is the decision layer between them. If capture is reading and Data Health is checking, AI Assist is the part that proposes treatment. A practice that has not standardised its treatment will find AI Assist propagating inconsistency at speed, which is worth knowing before rollout rather than after.
A practice with sixty owner-managed clients wants consistency in how a familiar set of suppliers is coded. Today, three staff code the same supplier three ways depending on who has the file.
The wrong rollout enables the feature for all sixty and lets it learn from all three habits at once. The result is faster inconsistency and a harder problem to unpick, because the pattern is now embedded rather than merely habitual.
The rollout that works starts with one staff member's decisions on one client segment, agreed in advance with a partner: this supplier, this account, this tax treatment, these exceptions. Dext's page states that a firm can apply automation practice-wide or manage it at the client level and can review, refine, reject or auto-apply — so use that granularity deliberately. Expand only once a month's output has been sampled and the exceptions are understood.
1. Sales tax treatment. Consistency is only a virtue if the treatment is right. Provincial variation in GST, HST, QST and PST means a rule learned from one client segment is not automatically correct for another. Partner sign-off on the tax rules being propagated is not optional.
2. Where records live. Subsection 286(1.2) of the Excise Tax Act requires a record to be kept in Canada, in English or French, unless the Minister authorises otherwise, and subsection 286(3.1) requires records kept electronically to be retained in an electronically readable format. Where the source documents are stored is therefore a compliance question, and the answer belongs in writing from the vendor.
3. Whose obligation it is. Client records remain the client's obligation. Section 230 of the Income Tax Act sets the general six-year retention period, and our sister firm's note on CRA books and records requirements is a useful thing to send a client who assumes the software is the record.
4. Privacy. Client documents contain personal information, and the OPC's principles for generative AI apply to organisations using such tools, not only to those building them.
5. Who signs. The engagement output is signed by a person. Whether the file is a compilation, review or audit changes what that person is attesting to; the Auditing and Assurance Standards Board sets out how those engagement types differ, and no automation setting alters the responsibility. If it goes wrong, the analysis is professional negligence against the firm, not against the vendor.
Does AI Assist post entries without us?
Dext's page states that the firm decides how automation is applied — review, refine, reject, or auto-apply — and that automation can be managed practice-wide or per client. Auto-apply is a setting a firm chooses, so treat enabling it as a policy decision with a named owner.
Is it useful for a very small practice?
The benefit scales with repetition across clients. A practice with a handful of clients and one person coding gets less from propagating decisions, because there is nothing to propagate to. Capture and Data Health may be the better starting point.
Does it replace a bookkeeper?
No. It repeats decisions a bookkeeper has already made, and it needs someone to make and maintain those decisions and to handle everything it flags. If you are looking at the function inside a single business rather than a practice serving clients, see AI in bookkeeping.
A 30-minute call is enough to tell you whether AI pays for itself here.
