Custom AI vs off-the-shelf tools for New Zealand businesses.

What is the difference between custom AI and off-the-shelf tools?

Off-the-shelf AI is a product you subscribe to, such as ChatGPT Business, Claude Team or Microsoft Copilot. Custom AI is software built for your workflow on model APIs. Most of the useful decisions sit in a third option between them: a product configured for your business.

How New Zealand firms actually split: in the AI Forum’s March 2025 survey, 72% used pre-existing AI solutions and 13% custom-built ones. It was a self-selected sample from an industry body, not a representative one, but the direction is clear.

How do you test buying before building?

Run the product and, if needed, a narrow prototype on the same 30 real cases. The product goes first; the build has to beat it by a margin worth paying for.

1. Invoice coding. Give 30 recent supplier bills to Claude with the Xero connector, or to ChatGPT with an export, and ask for the account code for each. Count correct codes. Note that the connector is read-only, so a person still keys every bill.
2. Tenancy questions. Load 40 tenancy agreements into a Project and ask 30 questions property managers really ask. Score answers, and check whether anyone could see agreements for properties they do not manage.
3. Inbox triage. Classify 30 emails from a shared inbox into your categories and urgency. Score against your own calls.

For each task, record accuracy, minutes of correction, and what the product could not do at all, such as writing back with approval. That last list is your case for a build. Our invoice coding workflow and tenancy document search pages show the custom versions.

What does buying or building cost over three years?

Off-the-shelf: 12 people on ChatGPT Business Standard at NZ$33 a seat a month cost 12 × 33 × 12 = NZ$4,752 a year, or NZ$14,256 over three years, with GST treatment unconfirmed. Copilot’s Business add-on at NZ$34.00 excluding tax is similar. Our lesson on choosing business AI subscriptions compares the seat options.

Custom, model fees only: coding 2,000 invoices a month, assuming 3,000 input and 300 output tokens each.

• On Claude Haiku 5.5 (US$0.10 in, US$0.50 out per million tokens): 2,000 × (3,000 × 0.10 + 300 × 0.50) ÷ 1,000,000 = US$0.90 a month.
• On Claude Sonnet 5.5 (US$2 and US$10): 2,000 × (3,000 × 2 + 300 × 10) ÷ 1,000,000 = US$18 a month.

Measure tokens with the model you will use; Anthropic says its newer tokenizer produces about 30% more tokens for the same text than older models. Either way, model fees are not the cost that matters.

The payback rule, in staff hours: payback months = B ÷ (hours saved a month − hours of review and maintenance a month), with B expressed in staff-hour equivalents. Example: a build costing the equivalent of 600 staff hours that saves 50 hours a month and needs 10 hours of review and upkeep pays back in 600 ÷ 40 = 15 months. Our AI business case template sets this out for approval.

Neither side is fixed-cost. Products add usage charges: monday.com requires AI credits for new customers from 6 May 2026, and Atlassian’s Rovo extra-usage billing starts on 3 December 2026, on by default. Custom builds meet platform charges too: Xero has charged developers since 2 March 2026, and its remote MCP server needs the Core plan, A$35 a month excluding tax, for real organisations.

Which should you choose?

Choose off-the-shelf if

• The work is generic: drafting, summarising, research, meeting notes.
• A read-only connector or a file upload gives the tool all the data it needs.
• Volumes are low enough that per-seat pricing is cheaper than building.
• Your 30-case test shows the product is good enough with light editing.

Choose custom if

• The product fails your test on something that matters, and you can measure the gap.
• The task must write to your systems, with approvals and an audit trail.
• Volumes are high, or the work is part of a product or service you sell.
• You need a specific model or processing region, such as Claude inference kept in Australia and New Zealand.

If the product is close but not quite there, configure it before building. Our guide to Copilot Studio versus custom agents covers the configured route for Microsoft 365 firms.

Where do both fail?

• Off-the-shelf without rules. Seats bought, nobody trained, client data pasted into personal accounts. A product does not enforce your workflow.
• Custom without an owner. A system nobody inside the business owns degrades quietly as data and models change. OpenAI shut down its Assistants API on 26 August 2026; builds on it had to migrate.
• Write access without checks. Xero’s remote MCP tools do not ask for confirmation before writing data. Any custom build that writes to accounting systems needs its own approval step.
• No evaluation. Whether bought or built, keep the 30 cases and rerun them after every change.

When the test says build, our custom AI development team starts from your 30 cases.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.