AI has never been more accessible, or more capable. You can now build a genuinely useful AI tool for your accounting firm over a weekend.
And no, we’re not talking about Zato.
We’re talking about what a capable accountant can put together with Claude, ChatGPT and a bit of curiosity. Something that analyses a ledger, prepares reconciliations or drafts a set of financial statements. A few years ago that would have needed a development team. Today it needs an idea and an afternoon or two.
That’s a remarkable change, and firms of every size should be taking advantage of it.
But there’s a difference between building a tool that works well for you, one that works reliably across your team, and a production system the whole firm can depend on.
Each is achievable. Each comes with a different level of cost, complexity and responsibility.
Which brings us to the question behind this paper. It’s not whether an accounting firm can build its own AI tools. Increasingly, it can. It’s whether building them is the best use of its capital, people and time.
Answering that honestly means drawing a distinction that’s easy to miss in the early days of any technology: the difference between building a tool that does one thing well, and building a system that can carry an accounting job from the client’s first email all the way through to review and sign-off.
The first shows what AI can do. The second tests whether it actually changes how a firm works: how much time and cost come out of the whole process, whether the results can be trusted, and whether those savings hold up next year and the year after.
Executive summary
Our view, as a specialist AI-native accounting platform provider, is that firms should invest in their own AI capabilities, genuinely and seriously. But before they set out to recreate established compliance production infrastructure, they should look carefully at the full commercial and operational commitment involved.
Not every firm needs an enterprise production platform. Some will get excellent results from a handful of carefully built agents. Others may need connected workflows across their teams. The investment decision changes as the scope grows, but the questions around measurable savings, reliability and ongoing ownership remain.
Five considerations have shaped our thinking:
- Individual productivity isn’t the same as firm-wide productivity. An agent might finish a task in minutes. The saving that matters is the one left over after gathering the information, checking the output, fixing what’s wrong, and getting the job reviewed and signed off.
- Building an agent isn’t the same as building a production system. As automation spreads through a firm, data integrity, document traceability, security, auditability, user permissions and workflow controls stop being nice-to-haves and start carrying real weight.
- The real cost is ownership, not development. Software has to be tested, maintained and supported. Tax rules, thresholds, integrations and AI models all change, and through all of it, last year’s accounting treatments still have to come out right.
- Time to value matters. A build costs more than the build. It costs the hours of accountants, partners and IT specialists, and it costs whatever productivity gains the firm puts off while the work is underway.
- Build where it makes a real difference. For many firms, pairing established production infrastructure with targeted internal AI work will return more than trying to recreate everything in-house.
None of this is an argument for reflexively buying instead of building. It’s an argument for comparing like with like, understanding the full cost of each path, and testing the results before making a big bet.
What the research is telling us
Some of the most useful recent research doesn’t tell us whether AI works. It tells us something subtler: that activity and results are not the same thing.
of respondents reported improved individual productivity from AI, while 37% attributed at least some positive EBIT impact to AI at the organisation level. Separately, 32% said their organisations had decided against purchasing at least one software product or feature because agentic coding tools made internal development possible.
McKinsey, August 2026generative AI projects had been abandoned after proof of concept by the end of 2025, over weak risk controls, rising costs, poor data quality and unclear business value.
Gartner, January 2026AI magnifies whatever strengths and weaknesses a team already has. The biggest returns came from improving the underlying system, not just adopting the tools.
Google Cloud DORA, 2025These studies span many industries rather than accounting firms specifically, and we shouldn’t stretch them further than they go. But the pattern is hard to ignore. Getting AI to work, it turns out, is one achievement. Getting it to pay off is another.
01Are we actually saving time?
Think for a moment about everything that goes into a single annual compliance job.
It usually starts well before any numbers are crunched. Someone reviews last year’s file, prepares the client questionnaire, sends the document request and then, as every accountant knows, chases the responses that don’t come back. Information is pulled together from the accounting system, from IRD or the ATO, from the document management system and from a long tail of client emails. It has to be organised, reconciled and tied to supporting workpapers. Adjustments are prepared and posted. Queries and review points are worked through. Only then are the financial statements and tax returns finished, checked and signed.
AI can help at almost every one of those stages. But automating one step rarely makes the steps around it disappear.
Picture an agent that takes an uploaded trial balance and supporting documents and produces reconciliations and draft accounts. What used to take hours now takes minutes. That’s a real gain, and it’s worth recognising.
Then the harder questions begin. How long did it take to gather and prepare what went in? Was everything that should have been included actually there? Can the reviewer trace each material balance back to its evidence? Have the proposed journals been posted? Where do the open queries and review points live? And perhaps most telling of all: could another accountant pick up the file tomorrow, understand what was done, and take it through to sign-off with confidence?
To be fair to the technology, modern agents can connect to APIs, retrieve documents, prepare journals and coordinate a workflow. None of that is out of reach. The hard part is stitching those capabilities into a process that holds up reliably across different jobs, different people and different reporting periods.
The arithmetic here is simple, and a little sobering.
An agent that saves an hour on workpapers but adds 45 minutes of extra handling, checking and correction hasn’t saved an hour. It has saved fifteen minutes.
That’s why we’d encourage firms to measure the whole process, not the speed of any single task. The numbers that matter are total preparer and reviewer time, rework, quality, turnaround, and the cost of each job that is completed and signed off.
One thing we’ve learned from building accounting production software is that the difficult work often sits between the obvious tasks. Producing a reconciliation is one thing. Making sure the correct information was collected, the supporting documents are traceable, the accounting treatment is appropriate, and the resulting work can be reviewed and signed off is another. These are the details that tend to emerge when a successful demonstration becomes part of a firm’s everyday workflow.
02An agent is not a production system
Building a useful AI agent has never been easier. Building a complete accounting production system asks something more of a firm: more capability, and more responsibility.
Data integrity, security and auditability
An accounting system has to do more than produce answers that sound right.
Can the firm say where each material figure came from? Can it trace the supporting documents, reproduce the calculations and show which adjustments were made? Can a reviewer examine the work without having to rebuild it themselves? Are user permissions, client confidentiality, data retention, system access and audit trails properly managed?
These questions apply just as much to software a firm buys. Buying software doesn’t hand off a firm’s governance responsibilities.
The difference is that when a firm builds its own system, it also takes on the work of designing, implementing, testing and maintaining every one of those controls. Depending on the risks involved, that can mean secure hosting, access management, monitoring, backups, incident response, vulnerability management and independent security assessments.
None of this requires a server humming away in the back office. Managed cloud infrastructure can provide most of it. But someone still has to configure it, secure it, watch it and keep it current. For firms looking for a framework, NIST’s Generative AI Risk Management Profile and OWASP’s guidance on LLM application security are good places to start.
Keeping up with the rules
If there is one cost firms most often underestimate, it’s this one.
Tax law changes. Thresholds move. Reporting requirements and accounting standards evolve. Integrations get updated, and AI models are replaced by newer ones. A system has to absorb all of that while still treating past periods correctly.
Take a simple example. An income tax calculation for an earlier year has to apply the rules and thresholds of that year, even after the current-year logic has moved on. That means date-effective rules, version control, regression tests and the ability to reproduce historical results on demand. You can’t simply push a change across every existing file and hope for the best.
And the work is never quite finished. As a firm takes on new entities, industries and accounting treatments, new scenarios appear, each of which may need more development, testing or validation.
Getting an agent to work today is a genuine achievement. Keeping an accounting production system accurate, secure and dependable year after year is a commitment of a different kind.
Can it run without the person who built it?
Many firms have someone like this: the accountant who taught themselves to build, who stayed late, who made the thing work. Their solution may run beautifully, for them.
But can the rest of the team use it just as well? What happens when that person goes on leave, gets pulled onto something urgent, or moves on? Who fixes it when it breaks in the busiest week of the year? Who updates the integrations, trains the new graduate, and keeps the process running when the pressure is highest?
These questions matter even in a small practice, and especially in one that plans to grow. A well-built internal solution can answer every one of them. But that takes documented procedures, clear ownership, testing, support and a plan for succession: the unglamorous scaffolding that turns one person’s ingenuity into something a whole firm can rely on.
03What does ownership really cost?
Many firms already have a technically gifted accountant, an in-house IT person or a trusted outsourced provider. That’s a real advantage, and in some circumstances it can make building internally a sound commercial decision.
Still, having access to technical people isn’t the same as having the capacity to design, run and continually improve accounting production software. An IT team or provider may be excellent at networks, devices, cybersecurity, Microsoft 365 and everyday support. Building accounting-specific AI applications draws on a different set of skills: software engineering, data integration, accounting logic, testing and AI evaluation. Some providers bring all of that. Others would need specialist help or a broader engagement.
“We’ll build it in our spare time”
It’s a tempting idea, and an understandable one. But internal development doesn’t become free just because nobody new was hired.
Say two people each spend five hours a week building and maintaining an internal tool. Over 46 working weeks, that adds up to 460 hours a year. At an illustrative economic value of NZ$100 an hour, the firm is committing NZ$46,000 worth of time annually, before infrastructure, subscriptions, external services or support.
This is an opportunity-cost illustration, not necessarily cash out the door or profit lost. Its real impact depends on what those two people would otherwise have been doing. The firm may well decide it’s worth it. We’d only ask that the hours be counted honestly, and the savings measured just as carefully.
The same goes for AI-generated code. It can cut development time dramatically. It doesn’t remove the need to get the requirements right, validate the accounting logic, test the integrations, review security and maintain the finished system.
What would it take?
The cost depends on what you’re trying to build.
An individual accountant can create a useful agent with relatively little expenditure. For some practices, that may be all they need.
A small internal team can go considerably further, connecting systems and developing tools that automate work across the practice. That can be commercially attractive too, although testing, maintenance, security, staff time and ongoing support increasingly enter the equation.
Building a complete compliance production platform introduces another level of responsibility. Multiple users, entity types, integrations, accounting treatments, review requirements and financial years all need to be supported reliably.
AI-assisted development can reduce the resources required at every level. But someone still needs to own the accounting logic, security, testing, integrations and maintenance.
To illustrate what that can look like at enterprise scale, consider one possible resourcing model.
Enterprise-scale scenario: one possible resourcing model for a complete compliance production platform
| Function | FTE | Illustrative annual cost |
|---|---|---|
| Product lead and delivery coordination | 1.0 | NZ$170k |
| AI and software engineering | 3.0 | NZ$540k |
| Data and integration engineering | 1.0 | NZ$170k |
| Security and DevOps | 0.5 | NZ$100k |
| Quality assurance and automated testing | 1.0 | NZ$120k |
| Senior accounting subject-matter expertise | 2.0 | NZ$300k |
| Employment on-costs | NZ$350k | |
| Cloud, AI usage, tools, security testing and other operating costs | NZ$250k | |
| Total | 8.5 | NZ$2.0m |
Illustrative NZD estimates, excluding GST. This is one possible resourcing model for a substantial internally operated platform, not a prescribed team structure or minimum staffing requirement. Salary assumptions should be tested against current market benchmarks and the firm’s existing resources.
These functions need to be covered, but not necessarily by 8.5 separately employed people. Firms may use existing staff, external providers, shared responsibilities or AI-assisted development. The cost and operational accountability will depend on how those functions are delivered.
AI-assisted development may shrink the engineering effort, while the complexity of implementation and the sheer range of accounting scenarios may grow it. The team could also look quite different after launch than it does during the build.
But somebody, somewhere, has to be accountable for each of these functions.
At this level of spending, plus roughly NZ$250,000 in rollout and support costs over three years, the direct investment comes to about NZ$6.25 million over three years. Both figures assume this particular enterprise team and spending profile; a smaller scope, or the same functions delivered differently, would cost less.
We want to be careful here. This is an enterprise-scale scenario. It is not a realistic benchmark for a sole practitioner building a few agents, and it isn’t meant to suggest that every firm needs 8.5 people to use AI well. The point is narrower: building an equivalent production capability takes far more than code, and those responsibilities have to be funded or absorbed somewhere in the business.
The cost of waiting
There’s one more cost that rarely makes it into a business case, and that’s time.
Consider a firm that completes 1,000 annual compliance jobs. If it could cut total preparation and review time by two hours per job, it would release 2,000 hours a year. At an illustrative labour cost of NZ$75 an hour, that’s NZ$150,000 in gross annual capacity value.
| Annual jobs | Hours released | Gross capacity value |
|---|---|---|
| 100 | 200 | NZ$15,000 |
| 500 | 1,000 | NZ$75,000 |
| 1,000 | 2,000 | NZ$150,000 |
| 5,000 | 10,000 | NZ$750,000 |
Assumes two net hours saved per completed job at NZ$75 per hour. The amounts are illustrative capacity values, not guaranteed savings or increased profit.
Whether that capacity becomes money depends on what the firm does with it: reducing costs, absorbing more work, or moving people into revenue-generating services. And the savings still have to be measured, including the time spent checking AI output and correcting mistakes.
Internal development can deliver benefits along the way. But a longer build may also defer some of the gains available from an established solution. That opportunity cost belongs in the calculation too.
04Another path: shared development
A specialist production platform rests on a simple idea: the cost of building and maintaining common functionality is shared across many customers rather than carried by one.
That has consequences beyond price. A platform’s development can draw on many firms at once, across different industries, entity types, accounting scenarios and review practices. An improvement to an integration, a workpaper, a piece of accounting logic, a security control or a workflow can benefit many customers, without each firm paying separately to build the same thing.
This matters at every scale. Smaller practices gain access to capabilities that might otherwise be difficult or uneconomic to develop and maintain themselves. Larger firms can avoid recreating infrastructure that needs to operate consistently across different teams, offices and client portfolios.
There’s another advantage to an integrated production platform beyond sharing development costs. Information from accounting systems, tax authorities and document management systems can move through a connected process, into workpapers, adjustments, queries and review, without the firm having to continually move it between separate tools.
That doesn’t eliminate professional judgement or every manual step. But it can reduce the administrative work between tasks, which is often where a surprising amount of time disappears.
What Gilligan Rowe & Associates found
Gilligan Rowe & Associates (GRA), an Auckland firm of around 150 people, rebuilt its compliance production workflow around Zato. In its published case study, GRA reported running more than 1,400 compliance jobs through Zato and lifting firm-wide margin by roughly 10 percentage points within six months.
Just as notable is what the firm chose to keep. GRA held on to New Zealand quality control and partner sign-off, while reducing routine production work and redirecting local staff toward advisory, tax planning and client relationships.
A single example has its limits. GRA’s results reflect its own assessment and its own operating model, which includes outsourced preparation. They don’t isolate Zato’s contribution, and they can’t tell another firm what it will achieve. The broader lesson is the one worth carrying away: productivity gains matter most when they change the economics of the whole production process, not just individual tasks.
When buying isn’t the answer
Buying isn’t automatically the right decision.
A commercial platform may lack functionality a firm needs. It can create vendor dependence. It may fall short of a firm’s security, integration or operational requirements. Implementation and change management carry their own internal costs, and pricing and the vendor’s future product direction remain fair commercial concerns.
An internal solution, by contrast, can offer more ownership, flexibility and control, particularly where a firm already has capable technical people.
The fair comparison is between options that deliver equivalent outcomes, judged on quality, security, ongoing cost and the operational responsibilities each one places on the firm.
05So where should firms build?
For many firms, the strongest answer may be: both.
Annual compliance is, at its core, broadly the same across the profession. Every firm has to collect information, reconcile accounts, prepare workpapers, process adjustments and complete appropriate reviews. That work is essential. But it isn’t usually where a firm sets itself apart.
Internal AI development may carry far more strategic value elsewhere: in industry-specific financial insight, proprietary advisory methods, client forecasting, tax planning, benchmarking, and new client-facing services.
A hybrid approach lets a firm adopt proven production infrastructure while pointing its own talent at the things that make it distinctive. It also leaves room to build its own integrations and workflows wherever established products don’t fully meet its needs.
06Five questions to ask before committing
Before putting serious resources behind either an internal build or a commercial platform, it’s worth sitting down as a partnership and working honestly through five questions:
- What are we actually building?A single agent, a connected workflow, or a complete production system?
- Are we genuinely saving time?Net preparation, review, checking and rework time, all the way to final sign-off.
- Can it run reliably?Security, data integrity, auditability, support, accounting accuracy and historical version control.
- What will it really cost to own?Development, staff time, infrastructure, maintenance, updates, implementation and delayed benefits.
- Have we tested the alternatives properly?Compare the existing workflow, an internal solution and an established platform against representative jobs, with the same measures of time, quality, cost and risk.
A controlled commercial trial is a proportionate first step before any larger commitment. Choose a set of representative jobs, establish a reliable baseline, and measure outcomes through to final sign-off, including accuracy, rework, reviewer confidence, adoption and the effort of implementation.
Then set those results beside the realistic cost and timeframe of building something equivalent yourselves.
Conclusion
Accounting firms should absolutely invest in their own AI capabilities. The tools available today make it possible to build valuable things faster and more affordably than ever, and for many tasks, building internally will be exactly the right call.
But it’s worth being clear-eyed about the difference between creating a useful agent and taking on responsibility for an entire accounting production system. As automation spreads, the things that sit beneath the work only grow in importance: data integrity, security, auditability, keeping pace with accounting changes, technical maintenance, and continuity when people come and go.
The best decisions won’t be driven by what can be built. They’ll be driven by what can be delivered reliably, maintained economically, and turned into benefits that show up across the whole firm.
For many practices, that will mean adopting established infrastructure where it meets their needs, and investing their own AI expertise where they can create something genuinely their own.
In the end, the question was never really whether you can build it. It’s whether building it will leave your firm better off, and more sustainably so, than the alternatives.
Research and references
- McKinsey & Company, The State of AI in 2026: On the Road to ROI, August 2026.
- Gartner, Why Half of GenAI Projects Fail: Avoid These 5 Common Mistakes, January 2026.
- Google Cloud DORA, 2025 State of AI-Assisted Software Development, 2025.
- Menlo Ventures, 2026: The State of Enterprise AI, November 2026.
- National Institute of Standards and Technology, AI Risk Management Framework: Generative AI Profile, July 2024.
- OWASP, Top 10 for LLM Applications 2025, 2025.
- Chartered Accountants Australia and New Zealand, Submission on Use of AI and the Code of Professional Conduct, May 2026.
- Robert Walters, New Zealand Salary Guide 2026, 2026.
- Inland Revenue, Individual Income Tax Rates, historical and current rates.
- Gilligan Rowe & Associates, How GRA Added 10 Percentage Points to Firm-Wide Margin, September 2026.
Published by Zato, a provider of AI-native accounting production software. This article reflects Zato’s industry perspective. Financial scenarios and staffing assumptions are illustrative, not independently established minimum costs or forecasts of savings. Actual outcomes depend on each firm’s operating model and requirements.