From Prompt to Intent: Why AI's Real Transformation Isn't What You Think
Return to Part 0: Table of ContentsPrevious Article: Part 5 Title
Three companies now control 88 percent of all enterprise AI usage.
Not 88 percent of consumer chatbot conversations. Not 88 percent of AI research papers. Eighty-eight percent of the AI that enterprises use to draft contracts, assess credit risk, route supply chains, and make hiring recommendations runs through Anthropic, OpenAI, or Google.
Menlo Ventures surveyed more than 150 technical leaders in January 2026 and found that these three providers collectively account for 88 percent of enterprise large language model API usage. [1] More than half of the enterprises surveyed reported they do not use open-source models at all. The remaining 12 percent of usage is spread across Meta's Llama, Mistral, Cohere, and a long tail of smaller providers, none of which individually commands meaningful market share. Preference for closed-source models has been increasing steadily, driven by the pace of improvement in proprietary systems, limited internal AI talent, and a perhaps counterintuitive belief that vendor-managed security is more trustworthy than self-hosted alternatives.
That concentration would be worth monitoring on its own. But it becomes something more consequential when you combine it with a second shift: these models are no longer waiting for instructions. They are pursuing goals.
The previous chapters in this section traced the contours of the agentic economy. Part 3 introduced the 23 percent of organisations scaling agentic AI, the 62 percent experimenting, and the threshold at which human intervention in many workflows has become effectively ceremonial. [2] Part 4 laid out the productivity economics of digital assembly lines, where legal review costs collapse from dollars to cents and cycle times compress by a third. Part 5 quantified the governance gap: 81 percent of organisations in the earliest stages of AI maturity, fewer than one percent governing properly.
This chapter examines what sits beneath all three patterns. The operational pivot from prompt to intent is the mechanism through which AI stops being a tool you use and starts becoming a system that acts on your behalf. Understanding that mechanism, and its implications for sovereignty, concentration, and organisational control, is not optional. It is the prerequisite for every governance decision that follows.
The Staff Officer, Not the Co-Pilot
A year ago, the dominant metaphor for enterprise AI was the co-pilot. Microsoft named its flagship product after it. The metaphor implied human primacy: the pilot flies the plane; the co-pilot assists. The human stays in command.
That metaphor is quietly dying.
In practice, the relationship has inverted. Gartner projects that 15 percent of day-to-day work decisions will be made autonomously by AI agents by 2028, up from zero percent in 2024. [3] If that projection holds, it describes a transfer of authority, not a support function. PwC reports that 79 percent of organisations have already adopted AI agents at some level, with 96 percent planning to expand their use in 2025. [4] These figures point to a workforce transformation that, for many organisations, has already passed the point of reversibility.
The better metaphor is the staff officer. In military doctrine, the staff officer does not merely relay information or assist with decisions. The staff officer gathers intelligence, analyses options, drafts operational plans, and presents the commanding officer with a recommended course of action. The commanding officer may accept, modify, or reject the recommendation. In theory. In practice, time pressure and information asymmetry mean the recommendation is accepted far more often than it is questioned.
That is what agentic AI does in enterprise settings today.
In finance, autonomous portfolio managers rehedge in near real time, surfacing decision points to humans only where the risk profile demands intervention. In legal departments, agent frameworks parse discovery troves for precedents, draft motions, and assemble filings before junior counsel has checked the inbox. In engineering, they manage release cycles, merge pull requests, and document changes with a precision that reduces human QA workload by half. Microsoft Copilot and AWS Q draft policy briefs, update market analyses, assemble budgets, and schedule executive calendars. In every sector where data, rules, and outcomes are well specified, agentic frameworks are proving faster, cheaper, and, in many implementations, more consistent than the human workflows they replace.
In every case, a human is nominally in the loop. In most cases, the human approves what the machine has already decided. The oversight exists on the org chart. It does not exist in practice.
This is the prompt-to-intent transition. Prompting is transactional: you ask a question, you get an answer. You remain in control of the workflow, deciding what to ask next, when to stop, and what to do with the output. Intent-based systems operate differently. They receive a goal, decompose it into sub-tasks, select their own tools, execute across multiple steps, and report back. The system decides what information to gather, which sequence to follow, and when the task is complete. The first model is a calculator. The second is a contractor.
The distinction matters because it changes the risk profile of every deployment. A prompt-driven system fails small: a bad answer to a single question. An intent-driven system fails at scale: a misaligned goal propagated across dozens of automated steps before any human notices.
Gartner projects that by 2028, one third of enterprise software applications will include agentic AI, enabling 15 percent of day-to-day work decisions to be made autonomously. [3] By 2027, one third of agentic deployments are expected to use multiple agents with different specialisations working together. These are not isolated tools. They are digital teams, coordinating across functions, negotiating between competing objectives, and executing with a consistency that human teams cannot match at scale.
The speed of this transition is part of the problem. Blue Prism's 2025 Global Enterprise AI Survey found that 29 percent of organisations are already using agentic AI and 44 percent plan to implement it within the year. Yet 78 percent said they do not always trust these systems, and 69 percent of AI projects never make it into production. [5] The technology is deploying faster than the trust infrastructure required to support it.
The staff officer metaphor raises a governance question. But the concentration data raises a sovereignty one.
When Menlo Ventures published its mid-2025 survey, Anthropic commanded 32 percent of enterprise LLM market share by usage, OpenAI held 25 percent, and Google held 20 percent. [6] By January 2026, Anthropic's share had grown to 40 percent of enterprise LLM spend, with OpenAI at 27 percent and Google at 21 percent. [1] The trend is towards consolidation, not diversification.
This concentration carries specific consequences. First, provider dependency. When three vendors supply the intelligence layer for nine out of ten enterprise deployments, a pricing change, a policy shift, or a service disruption from any one of them becomes a systemic risk for the organisations that depend on it. The OECD's January 2026 guidance on responsible AI due diligence explicitly addressed this dynamic, noting that disengagement from AI business relationships may be impossible when only a handful of enterprises control essential supply. [7]
Second, sovereignty exposure. For a country like New Zealand, with five million people and no domestic foundation model capability, the dependency is more acute. When New Zealand public servants use AI tools, they are typically accessing models hosted in US data centres, governed by US law, and optimised for US English and US regulatory contexts. The Government Chief Digital Officer's cross-agency survey recorded 272 AI use cases across 70 agencies in 2025, with 55 now operational. [8] The survey does not record how many of those use cases run on infrastructure that New Zealand has no ability to influence if terms change.
Third, capability gatekeeping. Menlo Ventures reported that preference for closed-source models has been increasing steadily, driven by the rate of change in model quality, limited internal AI talent, and, surprisingly, perceived security advantages. [1] When enterprises conclude that proprietary models offer better performance and better security than open alternatives, the economic moat around the three dominant providers deepens. What looks like a market preference today becomes a structural dependency tomorrow.
This echoes a pattern Gutenberg would recognise. Before the printing press, the Catholic Church controlled the production and distribution of knowledge through monastic scriptoria. Gutenberg's press broke that monopoly, scattering printed pages across Europe and fuelling revolutions in religion, trade, and politics. But the press itself required capital, technical skill, and distribution networks. Those who controlled the presses shaped what was printed. The parallel to 2026 is uncomfortable: AI promises to democratise decision-making, but the infrastructure of that democratisation is controlled by fewer hands than most users realise.
New Zealand's position in this landscape is instructive for any small, open economy navigating the agentic transition.
The Reserve Bank of New Zealand's February 2026 analytical note found that approximately 30 percent of New Zealand workers face high joint exposure to AI and robotics automation, scoring above 0.8 on a 0-to-1 scale. [9] Software programmers scored 1.000 for AI exposure, the maximum possible. Workers with bachelor's or master's degrees showed no statistically significant difference in AI exposure compared to those with no qualifications. The credential shield, the assumption that higher education protects against automation, has collapsed for cognitive work.
This finding intersects directly with the prompt-to-intent transition. The tasks that knowledge workers have traditionally been paid to perform (structured analysis, document preparation, research synthesis, regulatory interpretation) are precisely the capabilities that agentic systems now execute autonomously. The analytical note maps a transformation already underway, one significant enough to feature in the Reserve Bank's labour market analysis.
The GCDO's cross-agency survey noted agentic AI as an emerging adoption category for the first time in 2025. [8] The barriers reported by agencies are revealing: skills shortages, funding constraints, security concerns, and privacy requirements. Technological barriers, which ranked fifth in 2024, dropped to eighth in 2025. The tools are no longer the bottleneck. The people, the governance, and the institutional readiness are.
New Zealand's Responsible AI Guidance for the Public Service, published in February 2025, provides a structured framework covering human oversight, transparency, and accountability requirements. [10] What it cannot address, because no domestic guidance can, is the concentration question. When 88 percent of enterprise AI runs through three providers headquartered in the United States, a New Zealand governance framework operates within constraints it did not set.
The workforce data makes this concrete. Employment Hero's January 2026 analysis showed New Zealand job numbers rising 4.9 percent year-on-year while total hours worked fell 4.7 percent. More workers, fewer hours. That pattern is consistent with AI restructuring work into smaller, task-level units, precisely what the prompt-to-intent shift enables. The agentic system does not eliminate the role. It absorbs the routine cognitive tasks within the role, compressing the hours required and concentrating the remaining human contribution on judgment, relationship, and exception handling.
For Māori workers, the RBNZ data introduces a complex picture. Lower average AI exposure, largely due to concentration in manual, community-focused, and agricultural roles, intersects with higher average robotics exposure and lower average pay. What looks like protection from cognitive automation may simply be a different form of economic vulnerability. The prompt-to-intent transition does not affect all communities equally, and governance frameworks that ignore this will entrench rather than address existing inequity.
The prompt-to-intent shift is not a technical curiosity. It is the mechanism through which AI moves from assisting human decisions to making them, and through which control over that decision-making capacity concentrates in a small number of providers.
Understanding this mechanism creates three practical obligations for enterprise leaders.
First, map your intent surface. Know where in your organisation AI systems are receiving goals rather than prompts. These are the points where decision-making authority has transferred, whether or not anyone authorised the transfer. The difference between a tool that drafts an email when asked and an agent that monitors your calendar, identifies scheduling conflicts, and resolves them autonomously is the difference between assistance and delegation. Both may live inside the same product.
Second, assess your provider concentration. If your organisation's AI capabilities depend substantially on one or two providers, you have a single point of failure that no internal governance framework can mitigate. The OECD's due diligence framework recommends that enterprises identify control points in their AI value chain and develop contingency plans for scenarios where key providers change terms or become unavailable. [7] For most organisations, this assessment has not been done.
Third, distinguish between nominal and effective human oversight. A human who clicks "approve" on every recommendation without the time, context, or expertise to evaluate the underlying analysis is not providing oversight. They are providing a liability shield. McKinsey's research found that high-performing organisations are distinguished not by having humans in the loop but by having defined processes for when model outputs require genuine human validation. [2] The question is not whether a human reviewed the output. It is whether the human could have changed it.
Executive Takeaway
The prompt-to-intent shift is the operational mechanism of the agentic era. It transfers decision-making from humans to machines. The machines run on infrastructure controlled by three companies that hold 88 percent of enterprise usage. For leaders, three actions matter now.
Map your intent surface. Identify every workflow where AI receives goals, not just prompts. These are your points of delegated authority.
Audit your provider concentration. If one vendor's outage would halt your AI-dependent operations, you have a governance gap that no policy document can fix.
Test your human oversight. Ask whether the humans reviewing AI outputs could realistically override them. If the answer is no, you have a process, not a safeguard.
For New Zealand leaders specifically: 30 percent of workers face high automation exposure and 88 percent of the AI they interact with runs through offshore providers. The governance challenge is not just internal. It is structural.
The Gutenberg press freed knowledge from monastic control. It also created new gatekeepers: the printers, the publishers, the distributors. Five centuries later, we face the same fork. AI promises to distribute decision-making power. Whether it actually does depends on whether leaders treat the prompt-to-intent shift as a feature to adopt or a transfer of authority to govern.
Your organisation almost certainly has workflows where AI is no longer waiting for instructions. It is acting on intent. When did you last audit which of those workflows have genuine human override capability, and what would happen if the provider behind them changed terms overnight?
• • •
The views expressed in this article are entirely my own, informed by more than 30 years of professional experience in architecture, security, and technology leadership in New Zealand. They do not represent the views of my employer, any government agency, or the New Zealand government. My commentary on legislation and policy is analytical, drawing on publicly available sources and my professional expertise in architecture, security, and AI governance. I follow the Public Service Commissioner's Code of Conduct for the Public Sector and social media guidance.
• • •
Andreas Hamberger is a New Zealand leader in Architecture & Security and Associate Member of the Institute of Directors. The Hamberger Report: Generative AI 2026 provides enterprise leaders with evidence-based analysis of the AI landscape.
This article was drafted with AI assistance (Claude by Anthropic). All analysis, frameworks, and editorial decisions are my own. I verify every claim, choose every word, and take full responsibility for the final text.
Published: 25 February 2026Copyright: © 2026 Andreas Hamberger. All rights reserved.
• • •
References
[1] Menlo Ventures. (2026, January). "2025: The State of Generative AI in the Enterprise." https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/
[2] McKinsey & Company. (2025, November). "The State of AI in 2025: Agents, Innovation, and Transformation." Global Survey. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
[3] Gartner, Inc. (2025, August). "Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026." Press Release. https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025
[4] PwC. (2025). "2025 AI Agent Survey." Survey of 1,000 US business leaders. Cited in multiple industry analyses.
[5] SS&C Blue Prism. (2025). "Global Enterprise AI Survey 2025." https://www.blueprism.com/resources/blog/ai-agentic-agents-survey-statistics/
[6] Menlo Ventures. (2025, July). "2025 Mid-Year LLM Market Update." https://menlovc.com/perspective/2025-mid-year-llm-market-update/
[7] OECD. (2026, January). "Due Diligence Guidance for Responsible Business Conduct in the AI Value Chain." OECD Publishing.
[8] Government Chief Digital Officer (NZ). (2025, August). "2025 Cross-Agency Survey of Use Cases for AI." New Zealand Digital Government. https://www.digital.govt.nz/
[9] Reserve Bank of New Zealand. (2026, February). Analytical Note AN2026-02: AI and Automation Exposure in the New Zealand Labour Market.
[10] Government Chief Digital Officer (NZ). (2025, February). "Responsible AI Guidance for the Public Service: GenAI." New Zealand Digital Government.
[11] Employment Hero. (2026, January). SmartMatch Employment Report: New Zealand.