Creating an AI budget.

Goal: Build a 12-month AI budget for your team in which every price is dated, every assumption is written down and people’s time is counted. Who it is for: Owners, practice managers and finance leads who need to approve or justify AI spending. Time: 40 minutes, plus a list of the AI tools and uses you are considering and access to the vendors’ pricing pages.

AI budgets tend to go wrong in a predictable direction. The subscription line is accurate to the cent, and the budget leaves out the hours staff spend checking AI output, the US-dollar usage bill that doubles in a busy month, and the promotional price that ends after a year. MBIE’s responsible AI guidance puts it plainly: AI tools “have different cost drivers to other known software models”, and buyers need to be alert to “the total cost of the tool and the impact of high usage, as well as the cost of ongoing support”.

What belongs in an AI budget

An AI budget has five kinds of cost. Only the first is usually visible on a pricing page.

As listed on 19 October 2026, ChatGPT Business standard seats were NZ$33 per user per month billed annually (NZ$41 monthly), and the Microsoft 365 Copilot Business add-on was NZ$34.00 per user per month on a yearly plan, shown discounted to NZ$29.14, with Microsoft stating the price does not include tax. On Anthropic’s API pricing page, Claude Sonnet 5.5 was US$2 per million input tokens and US$10 per million output tokens. Check them again before you rely on them; they change.

Exercise: build a 12-month AI budget

1. List uses and users. For each planned use, write who needs it and how often. Not everyone needs a seat.
2. Price each line from the vendor’s page. Record price, currency, GST treatment, billing term and the date you checked.
3. Estimate usage lines with arithmetic. Items per month times tokens per item times price, separately for input and output. Measure tokens on a sample of real documents if you can.
4. Add people’s time. Setup, training, review of AI output and ongoing maintenance, in hours, at a loaded hourly cost you choose and write down.
5. Build the sheet with AI. Use the worked example prompt in ChatGPT, Claude, Copilot or Gemini to lay out columns and formulas. You supply every price and assumption.
6. Add contingency and review dates. Put contingency on usage lines and diary when promotional prices, annual terms and exchange-rate assumptions need rechecking.
7. Sense-check the result. Compare total cost with the value from calculating time saved by AI. Ask the assistant which costs you have probably missed.

Worked example: an accounting practice’s first year

Illustrative example: an 18-person Auckland accounting practice planning 10 assistant seats and an API-based tool to extract data from 2,000 client documents a month. The practice is fictional; prices are as listed on 19 October 2026; every other figure is an assumption.

Build a 12-month AI budget table with columns: line, basis, monthly NZD,
annual NZD, assumption. Use formulas, not typed totals. Use only these
inputs and list any cost you think is missing at the end.

Seats: 10 x ChatGPT Business standard at NZ$33/user/month (annual).
API extraction: 2,000 documents/month; 6,000 input + 800 output tokens
each; US$2 per million input, US$10 per million output tokens.
Assume US$1 = NZ$1.70.
Automation platform: assume NZ$100/month.
Review: 1 minute per document at an assumed NZ$60/hour loaded cost.
Training: 18 staff x 3 hours at NZ$60/hour, one-off.
Contingency: 20% on API and automation lines.

Example output (illustrative, abridged):

Line                Basis                                   Annual NZD
Seats               10 x $33 x 12                            $3,960
API extraction      US$24 input + US$16 output = US$40/mth
                    x 1.70 = NZ$68/mth                         $816
Automation          $100 x 12                                $1,200
Review time         2,000 min = 33.3 hrs/mth x $60          $24,000
Training (one-off)  18 x 3 hrs x $60                         $3,240
Contingency         20% of (816 + 1,200)                       $403
Total                                                       $33,619
Software and usage: $6,379. People's time: $27,240.
Possibly missing: integration build and maintenance, admin time
managing seats, a pilot period before annual commitment.

The model costs NZ$68 a month; checking its output costs about NZ$2,000 a month. That ratio is typical of document work, and it is why the review assumption deserves the most scrutiny. Halving review time through better extraction or sampling saves far more than any model discount.

Where AI budgets go wrong

• No people line. Review and training time left out, so the budget understates cost several times over.
• Usage spikes. A busy month, a looping automation or a heavier model can multiply a usage bill. The OWASP GenAI LLM Top 10 for 2026 lists unbounded consumption, including cost exhaustion, as a risk. Set spending limits where vendors offer them.
• Exchange rates. US-dollar usage prices move with the exchange rate. State your assumption and revisit it.
• GST confusion. Mixing GST-inclusive and exclusive prices. Record which is which on every line.
• Expiring promotions. Discounted seat prices and introductory offers end. Budget the list price from the date they expire.
• Unused annual seats. Committing a year ahead for people who do not use the tool. Trial monthly first where you can.

The AI cost calculator and AI tools pricing for New Zealand businesses help with the inputs, and managing LLM costs covers technical ways to cut usage bills. Whether seats or API usage suit a task better is the subject of choosing between an API and a subscription, and choosing business AI subscriptions compares the products themselves. For custom builds, our guide to what custom AI projects cost explains the cost drivers.

Self-check

• Every vendor price has a date, currency, GST note and billing term.
• Usage lines show volume, tokens per item and price as a formula.
• Exchange-rate and hourly-cost assumptions are written down.
• People’s time for training and review is included.
• Contingency covers usage-based lines.
• Promotional prices and annual renewals have review dates.
• You compared total cost with the expected value.

Next lesson: Writing AI vendor requirements.

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