On the 19th of May Nicola Willis and Paul Goldsmith announced a plan to cut core public service headcount by approximately 9,000 jobs, with the headline-grabbing figure of keeping the core public service to about 1% of New Zealand's total population. Now this is not a particularly radical policy for those who are captaining the ship at the moment. It's a story old as time that feeds into the hiring / firing cycle of most western public service organisations. What is radical is the thesis that the Government will increase productivity of the remaining headcount by massive adoption of AI. Whether you buy into the AI productivity story is another discussion for a different day (not that it's unimportant), but let's hold that it is true for the argument's sake. This brings up the question of "what is the additional cost of wide scale AI adoption within the public service?" Currently the argument is the public service is not using [it enough](https://www.stuff.co.nz/politics/360980612/thousands-public-servants-lose-jobs-potentially-saving-government-over-billion-dollars). > Willis says the combined policies will amount to $2.4b over four years, with increased use of AI and fewer government departments contributing. >“Businesses and households are using AI every day and, while parts of the public sector have seized the opportunity to innovate, others are still locked into outdated ways of doing things that prioritise box-ticking over outcomes,” she said. We will then have some increased spend compared to what is currently being done in the public service. So how much? Well this is a very interesting question. Let's take a very basic stab at working out how much this will cost. Willis is targeting [55,000](https://thespinoff.co.nz/politics/20-05-2026/nicola-willis-cuts-jobs-loves-ai-hates-nz-first) core public service employees and they are likely to fall into three categories of usage. ## Bulk / Basic Usage The government is likely to go down the Copilot path given that it is primarily a Microsoft shop. Currently this sits at about $30 per month per seat, which makes things even cheaper. How has this worked for other countries? Taking a look across the ditch, the Australian Public Service undertook a Copilot study and managed to achieve a [perceived productivity boost of about 1hr per day per FTE for some users](https://www.digital.gov.au/initiatives/copilot-trial/microsoft-365-copilot-evaluation-report-full/productivity). Compared to the expected $2.4B in savings over 4 years, this makes basic numerical sense. Job done. However, not everyone will get by on just a chat style workflow with Copilot and not everyone in that pilot saved 1hr a day. ## Agentic Knowledge work Not all public servants have a workflow that can be summarised as uploading a single document and asking questions or drafting a single letter. We need to look at other knowledge work domains to see how AI usage has been adopted and the costs associated with it. For this I've taken two fields, law and finance. AI platforms for legal firms like Harvey have some fairly high costs associated with them: - Mid-market firms (50-200 attorneys): USD $1,200-$1,500/seat/month ([AI Vortex pricing analysis](https://www.aivortex.io/legal/ai-tools/harvey-ai-pricing-2026/)) - Am Law 100 firms: USD $1,500-$2,000+/seat/month - 20-seat minimum, $288K/year floor ([CostBench](https://costbench.com/software/ai-legal-tools/harvey-ai/)) - Allen & Overy deployed it to 3,500 lawyers across 43 offices ([A&O Shearman](https://www.aoshearman.com/en/news/ao-shearman-and-harvey-to-roll-out-agentic-ai-agents-targeting-complex-legal-workflows)); 40,000 queries in the pilot period - System-wide: 400K+ agentic queries/day, 445K Deep Analysis reports ([Harvey blog](https://www.harvey.ai/blog/how-agentic-search-unlocks-legal-research-intelligence)) Hebbia, which is an AI platform targeted at financial analysts: - Professional (power users designing agents): USD $10,000/seat/year ≈ $830/month ([Sacra](https://sacra.com/c/hebbia/)) - Lite (people consuming outputs and running pre-built agents): USD $3,000-$3,500/seat/year ≈ $250-$290/month - Claims ~33% of the top global asset managers by AUM as customers, deployed across 1,000+ production workflows Why finance and legal? They are great analogs to the work that some core public service officials do: working within, interpreting and drafting legislation, and of course forecasting, budgeting and modelling government spending as well as the impact government programs have on the country. The government has assumed AI is the Copilot kind, where a flat $30/seat replaces all office work. The work it actually needs done to justify cutting 8,700 FTEs, the kind currently done by senior policy analysts, legal advisors, and Treasury modellers, is the Harvey/Hebbia kind, where seats run $1,000+/month and the productivity story still depends on having a competent human to run the agent. ## Agentic coding / data analysis Of course the government also employs a lot of people that also write code, whether they are data analysts or ICT professionals. ICT professionals account for around 4% of core public service: not huge, but also likely to be the heaviest users. If we buy each of them either a ChatGPT Pro plan ([USD $100](https://openai.com/business/chatgpt-pricing/) per month) or an Anthropic Max subscription at the same [price](https://claude.com/pricing), the problem with this reasoning is that an organisation cannot use the drastically [subsidized](https://simonwillison.net/2026/May/27/product-market-fit/) Pro / Max plans that OpenAI and Anthropic use as loss leaders. As soon as you get above 150 users (the max seat count for Teams) or you want Enterprise grade controls (data privacy and governance features, SOC2 compliance, observability and security reporting) you have to switch to their enterprise plans. These start at about $20 USD per seat and then all model usage is billed at API pricing. The problem with this is it causes spend to spike anywhere between 5x to 20x from Pro and Max style plans, as the API cost is significantly more expensive than the limits provided in the prosumer / small business plans. Uber recently got stung with this with monthly per engineer costs spiking from [USD $500 to USD $2000 depending on usage](https://www.briefs.co/news/uber-torches-entire-2026-ai-budget-on-claude-code-in-four-months/). Anecdotal evidence from individual users trying to quantify their own token costs while using Max and Pro plans shows that enterprise pricing for the same usage would show a similar [spike in costs of around 10x](https://simonwillison.net/2026/May/27/product-market-fit/). This is not surprising, as it's these frontier model companies' most plausible near-term [path to profitability](https://finance.yahoo.com/news/anthropic-quietly-raises-ceiling-121501064.html). ## The outcome Three real tiers exist: | Tier | Workload | Price band (USD/seat/month) | Analog | | ---------------------- | --------------------------------------------------- | --------------------------- | --------------------- | | Bulk | Email, drafting, summarisation, simple Q&A | $30-$80 | M365 Copilot | | Agentic knowledge work | Multi-doc reasoning, research chains, briefing prep | $500-$3,000 | Hebbia Lite to Harvey | | Agentic coding / data | Tool-heavy iterative workflows | $1,000-$3,000+ | Uber on Claude Code | For 55,000 public servants the realistic mix probably looks something like 80% tier 1, 15% tier 2, [5% tier 3](https://www.digital.govt.nz/standards-and-guidance/strategy-and-planning/digital-capability-public-service-workforce/digital-workforce-data-and-evidence). Run those numbers at a middling USD $40 / $1,500 / $2,500 and you land at roughly: - 44,000 × $40 = $1.76M/month - 8,250 × $1,500 = $12.4M/month - 2,750 × $2,500 = $6.9M/month - Total: ~$21M/month, or **USD $250M/year** (NZD $425M/year) Not counting second order spending in order to support this AI infrastructure (storage, governance, etc). This quickly puts the prediction of $2.4B over four years in savings into question. ## The future of AI usage The other question to ask is whether these software engineering workflow burn rates will equate to a similar level of usage from public service style work. Possibly. Currently Anthropic's Claude Max 5x gives a token budget of approximately [88k to 110k per 5 hour window](https://milvus.io/ai-quick-reference/what-are-the-token-limits-for-claude-code). A legal / compliance document RAG search costs about [10k tokens per call](https://iternal.ai/token-usage-guide). In a given work day, a couple of these per user, combined with chat and other agentic tasks, can see a similar level of usage. Agentic coding can use millions of tokens in a single long running goal focused session, but it's much more cache friendly than free text analysis and RAG search, with cache hit rates of 70%-90% per call (check your own usage page in Claude Code to see it). A policy analyst running RAG over fresh submissions, Cabinet papers, and OIA requests has nothing to cache against. Per token, the public service workload is less efficient than the engineering one, not more. Will public servants start to leverage more agentic workflows and burn more tokens? Who knows, but that is where the software productivity coding gains have largely come from, and it will be interesting to see how this flows into other areas. With OpenAI's GPT-5.5 Instant now the [default model powering Copilot](https://www.iq-it.co.uk/blog/ai-models-powering-microsoft-365-copilot), we are likely to see Microsoft's enterprise pricing start to head in the same direction as OpenAI's if and only if we start to see other areas of the workforce adopt agentic style workflows. What's a worst case scenario? One where the public service actually does start to adopt broader and broader agentic workflows across the whole public service population. This would look something like this if we cherry picked just the Uber numbers for, say, an additional 50% of the public service: > 28,000 x $2,500 = USD $70M per month (or NZD $119M per month). Well and truly blowing out any savings, of course. That is just wild speculation that is definitely not based in any numbers about the rate of adoption of [agentic AI in other organisations](https://www.linkedin.com/posts/pneppalli_agentic-software-engineering-adoption-is-activity-7439402236541157376-6PwV/) once it hits a certain threshold of usefulness. Maybe at that level, further job cuts will pay for it? Whether you are an AI true believer or skeptic, or fall on the left or the right of the economic spectrum, the AI replacement argument made by the current NZ government is likely to come with some invoice shock. ![[Pasted image 20260528085947.png]] ### Sources - [Nearly 9000 public sector jobs to go — RNZ](https://www.rnz.co.nz/news/political/595655/nearly-9000-public-sector-jobs-to-go-government-agencies-to-merge-nicola-willis-announces) - [Budget 2026: Willis' cuts to save $2.4b — NZ Herald](https://www.nzherald.co.nz/nz/politics/budget-2026-pm-christopher-luxon-promising-job-cuts-as-nicola-willis-to-unveil-public-service-shrink/ERIJHW5CHJDRTPZHNA7O6CAYVA/) - [NZ to cut public service jobs — Bloomberg](https://www.bloomberg.com/news/articles/2026-05-18/new-zealand-to-cut-public-service-jobs-merge-some-departments-in-overhaul) - [ChatGPT Pricing — OpenAI](https://openai.com/business/chatgpt-pricing/) - [Claude Max plan — Anthropic](https://claude.com/pricing/max) - [Claude Enterprise plan — Anthropic Help Center](https://support.claude.com/en/articles/9797531-what-is-the-enterprise-plan) - [Uber torches 2026 AI budget on Claude Code — Briefs](https://www.briefs.co/news/uber-torches-entire-2026-ai-budget-on-claude-code-in-four-months/) - [Uber COO questions AI spend — Fortune](https://fortune.com/2026/05/26/uber-coo-ai-spending-tokens-claude-code/) - [Simon Willison on AI product-market fit & enterprise pricing](https://simonwillison.net/2026/May/27/product-market-fit/) - [Harvey AI Pricing 2026 — AI Vortex](https://www.aivortex.io/legal/ai-tools/harvey-ai-pricing-2026/) - [Harvey AI Pricing — CostBench](https://costbench.com/software/ai-legal-tools/harvey-ai/) - [A&O Shearman + Harvey agentic rollout](https://www.aoshearman.com/en/news/ao-shearman-and-harvey-to-roll-out-agentic-ai-agents-targeting-complex-legal-workflows) - [Harvey: How Agentic Search Unlocks Legal Research](https://www.harvey.ai/blog/how-agentic-search-unlocks-legal-research-intelligence) - [Hebbia funding and pricing — Sacra](https://sacra.com/c/hebbia/) - [Legal AI Pricing comparison 2026 — The Legal Prompts](https://thelegalprompts.com/blog/ai-legal-tools-pricing-comparison) - [Token Budgeting in Deep Legal Agent Chains — law.co](https://law.co/blog/token-budgeting-in-deep-legal-agent-chains) - [Jamie Dimon on AI and workforce redeployment — CNBC](https://www.cnbc.com/2026/02/24/jpm-ceo-jamie-dimon-ai-reshaping-workforce-redeployment.html) - [Goldman Sachs firmwide AI assistant — Fortune](https://fortune.com/2025/06/24/goldman-sachs-internal-ai-assistant/) - [Morgan Stanley 98% advisor adoption — CDO Magazine](https://www.cdomagazine.tech/aiml/98-of-morgan-stanley-wealth-management-advisors-use-its-ai-chatbot-with-improved-productivity-cao-explains-how)