The stack learned to think. Now it has to be trusted.
Something happened across the Canadian accounting stack in 2026 that is easy to miss one product announcement at a time, and obvious once they are lined up. Every major platform a firm relies on shipped an artificial intelligence agent that works inside the engagement rather than beside it. Caseware launched Verity. Countable relaunched Luka. Xero built Just Ask Xero. Sage rebuilt Sage Intacct around named agents. Dext shipped AI Assist. Wolters Kluwer pushed its Expert AI across the CCH suite. Different companies, different price points, one pattern: an agent that operates within the workflow, learns the firm’s own methodology, and keeps a person accountable for the result.
The word for this is agentic, and it marks a real break from the assistant era that preceded it. An assistant answers when it is asked. An agent takes a sequence of actions across a process, and that difference is the whole story, because an engagement is a sequence of actions, and software that can carry several of them changes what a firm is buying and what it remains responsible for.
What agentic actually means
Caseware, the largest name in Canadian assurance, was explicit about the distinction when it introduced Verity at its 2026 conference. Verity is not framed as a chatbot bolted to the side of the software. It is described as an orchestration layer with agents that run across the engagement lifecycle, drawing on the context of the engagement and the firm’s own methodology rather than answering generic questions in isolation. The competitors are making the same move in their own language. The common claim is that the agent understands the work, not merely the question.
That claim is precisely what raises the stakes. A tool that drafts an email is judged on convenience. A tool that drafts a workpaper is judged on whether a partner can sign the file behind it. The moment software moves from assisting to acting, the question stops being how much time it saves and becomes whether its work can be trusted, reviewed, and defended. Which is the question Caseware chose to build a framework around.
The Trust Stack, read from the outside
Over a series of essays this spring, Caseware’s chief executive set out what the company calls the AI Trust Stack, a five-layer way of describing how artificial intelligence should operate inside a profession built on trust. The layers move from the raw intelligence of the model, up through agents that can act, the workflows they act within, the professional context that tells them what a given firm actually does, and finally governance: the requirement that every output be reviewable, explainable, documented, and accountable.
Set aside that it comes from a vendor, and the framework is genuinely useful to an independent firm, because it supplies a vocabulary for a conversation most firms are having without one. A partner does not need to adopt Caseware’s products to adopt Caseware’s question. When any agent proposes to touch a file, the governance layer is the part that matters: can it show its work, and can the firm stand behind the result.
“An agent that cannot show its work is not a time-saver. It is an unreviewed opinion delivered in a confident tone.”
Why the incumbent and the challenger argue the same case
The most telling signal is not that the market leader is talking about governance. It is that its challenger is using the same words. Countable, which describes itself as the artificial-intelligence-native engagement platform for Canadian CPA firms, relaunched Luka as, in its framing, the first AI preparer for year-end engagements, bringing compilation, review, and tax into a single agentic flow. The pitch leans on speed, a reported four hours and more removed from a file. But the reassurance underneath it is identical to Caseware’s: every action logged, attributed, and defensible.
When the largest platform in the market and the insurgent trying to unseat it independently arrive at the same thesis, the direction is settled. Agents will do more of the preparation. The value a firm sells moves up, toward the judgment that reviews the agent’s work and signs the file. And the currency both vendors are competing on is not only capability. It is defensibility.
What this asks of an independent firm
The temptation is to read all of this as a reason to buy the biggest platform, and that is the wrong lesson. An independent firm does not need to purchase the largest suite. It needs a standard, applied evenly to every agent it lets near client work, whichever vendor supplies it. The standard is the governance layer stated plainly: the output must be reviewable, the reasoning explainable, the trail documented, and a named person accountable for the file. None of that changes because the first draft came from a machine.
Read that way, the agentic turn is not a threat to the independent firm. It is leverage, and unusually even-handed leverage, available to a two-partner practice on the same terms as a national. But it is leverage only for firms that adopt it deliberately, on their own terms, with the governance intact. The Practice Intel section of this issue turns the standard into five questions to ask before an agent touches a single file.