The Resurgent AI Journey: from bookkeeping to an AI-native finance firm

Start with the problem, not the technology

Resurgent runs a follow-the-sun model. A file opened in Montreal is progressed overnight in Asia and returned before the next business day. That model has an obvious benefit and a less obvious cost. The benefit is that the working day never really ends. The cost is the handover — because every handover is a re-explanation.

Somebody in Dhaka has to know what the reviewer in Montreal was worried about. Somebody in Mumbai has to know that the warrant tranche was repriced, and why. Multiply that by every client, every entity and every close, and the tax on a 24-hour firm is context transfer. You can pay that tax in meetings and long emails, or you can build systems where the context lives in the work itself.

That is the actual reason we went down this road. Not cost. Not novelty. Context.

We did not set out to build AI. We set out to stop re-explaining the same file twice a day.

Four phases, in strict order

The sequence mattered more than the ambition. Each phase was only possible because the one before it had actually landed — and firms that skip a step tend to end up with an impressive demonstration sitting on top of a filing cabinet.

1. 01

   Digitize

   Cloud accounting on QuickBooks and Xero, paperless, real time, across every entity we touch. This is the unglamorous foundation, and it is not optional. Nothing intelligent can be built over documents that only exist as paper in somebody's drawer.

2. 02

   Automate

   Workflow before intelligence. Our AP Payment Portal moves an invoice from capture through approval to payment and reconciles it back to the ledger, on deterministic rules that a controller can read and audit. No model required — and that was the point.

3. 03

   Agentic

   The AI CFO Suite: one agent per recurring finance job, each with a defined input, a defined output and a professional standard it has to meet. Cap tables, related-party disclosure, marketable securities, exploration schedules, buyback registers, statement compilation.

4. 04

   Insight

   The deliverable stops being a file emailed at month-end and becomes a live surface. Financial statements, MD&A inputs, AP ageing and market data published to a dashboard a director can open on a Sunday night before a board meeting.

Phase 02 taught us the most

The AP portal is the least sophisticated thing we have built and it changed the most. It taught us that most of what looks like judgment in a finance function is actually routing: who approves this, against which budget, in which currency, before which cut-off. Route it properly with plain rules and the volume of genuine judgment left over is surprisingly small — and the people doing it are no longer exhausted by the routing.

The firms that struggle with AI in finance are usually the ones that tried to put a model on top of a broken workflow. The model then has to guess at things the workflow should simply have known.

Phase 03: what an agent actually owns

The insight that unlocked the agentic phase was structural rather than technical. Public-company finance work is not an undifferentiated mass. It is a set of recurring, well-specified jobs that come round every quarter and are done to a written standard. That is precisely the shape a well-scoped agent can hold — and precisely the shape a generic chatbot cannot.

Note what these have in common. Each has an unambiguous source of truth, an output whose format is prescribed, and a professional standard against which the output can be checked. Where those three conditions hold, an agent is genuinely useful. Where they do not, we have not deployed one.

What we refuse to automate

This is the part of the journey that matters most to the boards and auditors we work with, so we state it plainly.

• A CPA signs, not a model. Every filed number, note and statement is reviewed and signed by a qualified accountant. The agent drafts; the professional is accountable.
• Agents are read-only by default. They read the books. Postings, payments and filings stay behind human approval.
• Every figure traces back. Outputs are source-linked to the ledger entry, broker confirmation or contract clause they came from, so a reviewer can check rather than trust.
• Client books are used to serve that client. They are not pooled across clients or used to train third-party models.
• Judgment stays human. Materiality, going concern, impairment triggers, related-party characterisation and disclosure calls are professional judgments and are treated as such.

What actually changed

We are cautious about publishing our own productivity numbers, because every firm defines them flatteringly. What we will say is qualitative and verifiable by any client: books stay current rather than being caught up before a deadline; the mechanical assembly of a reporting pack happens overnight, so the working day is spent on review; and adding a geography no longer means adding proportional headcount.

The wider profession is reporting the same direction of travel.

The same research is equally clear about the brake: trust in the agent's data is the single largest barrier to adoption, ahead of systems integration and skills. That matches our experience exactly. Capability has not been our constraint for some time. Evidence is.

The regulator is already in the room

In January 2026 the PCAOB set out that the profession's core values — independence, professional skepticism, responsibility — must be preserved in an AI-augmented audit environment, and named agentic auditing as a near-term prospect rather than a speculative one. The open question it raised is a good one: where does professional skepticism sit when an agent identifies a risk, designs the test, runs it and documents the result?

Our answer is that skepticism sits with the reviewer, and the reviewer needs evidence to be skeptical about. So our working papers record what the agent did, on which inputs, and who reviewed the output. Firms that automate first and document later will be rebuilding that trail under deadline pressure. We would rather not.

Where this goes next

The next phase is deeper ledger integration — permissioned, read-only, live — so that a consolidation or a covenant test is a query rather than a project. After that, continuous monitoring: the interesting version of a month-end close is one that never needs to happen because nothing was allowed to drift in the first place.

None of that changes the shape of the firm. Governance stays CPA-led and stays in Montreal. The agents make the follow-the-sun model cheaper and quieter. They do not sign anything.