From 23 vendor demos to one working workflow in 9 weeks.

The firm is a 40-staff accounting practice in the Bay of Plenty, with three partners, twelve senior accountants, and the usual mix of compliance, business advisory, and tax work. Their clients are Tauranga and Waikato SMBs, mostly trades, primary industry, and professional services.

By early 2026 the managing partner had been on the receiving end of 23 vendor pitches in six months. AI bookkeeping co-pilots. AI tax research tools. AI document review platforms. AI client-meeting note-takers. AI working paper assistants. Each pitch was 30-60 minutes, often followed by a free trial that the team would try for a week and then quietly abandon. Three pilots had progressed to paid subscriptions. None of those subscriptions were being used by more than two people.

The MD's brief to us was disarmingly honest. "I think we should be doing something with AI. The Big 4 firms are clearly doing something. I don't know what we should do. Every vendor I talk to tells me their product is the answer. I don't believe any of them anymore. Help."

The firm had three things going for it. First, partner-level buy-in, with a stated willingness to spend on the right tool. Second, a clean technology stack (Xero, Karbon, MyBriefcase, Microsoft 365), with no major legacy data problems. Third, a team that wanted to learn but had been burned by half a dozen failed trials and was tired of being asked to evaluate yet another vendor on top of their billable hours.

Three things were working against them. The vendor noise was making it impossible to think clearly about what they actually needed. The team's tolerance for trials was depleted. And the partners had no shared mental model of where AI moves the needle for a firm their size, which meant every conversation about which tool to buy turned into a strategy debate.

Phase one: Readiness Audit (weeks 1-2)

Two weeks of structured diagnostic work. Eight stakeholder interviews across the three partners, the operations manager, and four senior accountants chosen to span business advisory, tax compliance, and audit work. A current-state stack assessment mapped against their real workflows. A data hygiene check on their working paper templates and Xero ledger consistency. Workflow-by-workflow ROI mapping, scoring each candidate on impact and effort.

The audit identified seven high-ROI candidate workflows, ranked by partner-perceived value and team feasibility. The top three were tax-season working paper preparation, client meeting note generation, and IRD-letter and advisory-memo drafting. The audit also surfaced two compliance flags that needed sorting before any client-data tool went live (data residency on one shortlisted vendor, and a thin audit trail on another).

The audit deliverable was a 46-page document and a 90-minute partner walkthrough. The output the partners cared about most was the one-page ranked candidate list, with a clear "do this first, then this, then revisit" sequence and a budget envelope on each step.

Phase two: Tool Stack Selection (weeks 3-4)

From the 23 vendors the firm had been pitched, we built a long-list of 17 candidates that genuinely played in the top-three workflow areas. We pruned aggressively on three filters: data residency and Privacy Act 2020 fit (eliminated 6), pricing and total cost of ownership over three years (eliminated 4), and concrete fit with their existing stack (eliminated 3). That left a shortlist of four.

Three of those four went into a live pilot. Twelve senior accountants used each tool for two weeks on real client work (with appropriate anonymisation where the pilot vendor's data handling wasn't yet locked down). We built a structured scoring rubric in advance: output quality, time saved, integration friction, team enthusiasm, vendor responsiveness, and ongoing cost. Each user filled it in weekly. We collated and compared after four weeks of pilot time.

One tool emerged clearly ahead on output quality and team adoption. A second was close on time saved but lost on integration friction. The third was eliminated halfway through when a Privacy Act issue surfaced in the data-handling small print. The winner was the firm's third choice going in, not their first. The partner who had been most skeptical of the pilot process said afterward that the live testing was the first time he'd seen any of these tools used the way a real accountant would actually use them.

Phase three: Workflow Integration (weeks 5-9)

Five weeks of embedded work. We started with the highest-ranked workflow, tax-season working paper preparation, and built it out with three senior accountants who had been pilot leads. Prompt engineering, system prompts, integration with their working paper template library, and a verification step so every AI-generated draft was reviewed against the underlying source documents before signoff.

Weeks 6 and 7 were team training. Each of the 12 senior accountants ran the workflow on live client work, with one of us shadowing for the first day each, then weekly check-ins thereafter. We tracked adoption metrics from day one: who was using it, on what, how often, and whether the output was accepted, edited, or rejected. By week 7, eight of the 12 were using it independently. By week 9, all 12 were live with a documented handover, a runbook, and a named owner inside the firm.

The 22 cancelled vendor pitches stayed cancelled. The partners now had a clear yes for one tool, a clear no for 22 others, and a defensible answer when the next vendor called.

Three months after go-live, the firm reported the following outcomes (audited against pre-engagement baselines we agreed in week one, not vendor-supplied numbers).

Time reclaimed. The 12 senior accountants reported an average of 7.2 hours per week per person back, primarily on working paper preparation and advisory note drafting. Aggregate: roughly 87 hours per week across the senior team. Some of that has been redirected into higher-value advisory work for existing clients. Some has been recovered as capacity for new clients during tax season.

Cost avoided. Of the 22 vendors cancelled, 19 had been priced. Cumulative year-one subscription cost for those 22, had the firm continued to evaluate and procure them piecemeal, was approximately $48k+GST. None of that money was spent. The actual deployed tool came in at $14k+GST per year for the 12 seats, well within the original budget envelope.

Adoption. Adoption rates settled at 92% weekly active across the 12 senior accountants by month three. The remaining 8% (effectively one senior accountant who took longer to come around) was at 70% by month four and 100% by month five. Adoption inside the partner group was uneven (the MD was an early heavy user, one of the other two partners took longer to engage), but neither partner blocked progress on the rollout.

Compliance posture. A board-ready acceptable use policy was drafted as part of the engagement, signed off, and circulated. Data residency was confirmed New Zealand or Australia for all retained vendor relationships. The firm passed an internal data-handling audit four months after go-live with no AI-related findings.

What didn't happen. Worth naming. We didn't try to extend AI into bookkeeping or transaction-level work, which was lower on the ROI list and higher on the risk register. We didn't replace any staff (this came up explicitly at the partner kickoff; the engagement was scoped around reclaiming time for higher-value work, not headcount reduction). And we didn't promise the firm a moving target. The deployed tool is reviewed quarterly under a Fractional AI Director arrangement, with a planned re-evaluation at month 12 in case the broader market has shifted enough to justify a switch.