Predictions are dangerous. Most of them age badly. But after a year of deep enterprise AI delivery, Mak and I have enough signal to be usefully specific about where 2025 is heading. Some of this is extrapolation. Some of it is based on things we're already building. All of it could be wrong. But we're going to say it anyway.
What You Need to Know
- Platform consolidation is coming. The current fragmented landscape of AI tools, providers, and frameworks will consolidate significantly in 2025. Enterprises that bet on the right platforms will accelerate. Those on the wrong ones will face painful migrations.
- Agentic workflows will move from demo to production. AI systems that can take multi-step actions autonomously (not just answer questions) will enter enterprise production. Cautiously. With guardrails. But they'll arrive.
- Sovereign AI infrastructure will become a competitive requirement in regulated industries. Not just a nice-to-have. A procurement criterion.
- The gap between AI leaders and AI laggards will widen dramatically. The compound effect means early movers are accelerating while late movers haven't started. 2025 is the year this gap becomes visible.
Prediction 1: Platform Consolidation
The enterprise AI market in 2024 is fragmented. Dozens of LLM providers. Hundreds of AI application platforms. Thousands of point solutions for specific tasks. Every conference has a new "AI-powered" tool for every conceivable use case.
This is unsustainable. Enterprises cannot manage 15 AI vendors any more than they could manage 15 cloud providers. Consolidation is coming, and it will be swift.
What we expect:
Three to four major enterprise AI platforms will emerge as the default choices for most organisations. Not because they're the best at everything, but because the integration, governance, and operational overhead of managing multiple platforms is too high.
Microsoft's Copilot ecosystem and its Azure AI platform are the obvious frontrunner for enterprises already in the Microsoft stack. Google's Vertex AI is the alternative for Google Cloud shops. AWS Bedrock for Amazon-aligned organisations. And the open-source ecosystem (Hugging Face, vLLM, and associated tooling) for organisations that want control.
The point solutions will either integrate into these platforms or fade. An AI document extraction tool that doesn't integrate with the enterprise's chosen platform is a dead end.
What this means for NZ enterprises:
Make your platform bet early. Not necessarily now, but in the first half of 2025. Waiting for the "winner" to emerge is a strategy that leaves you on the wrong side of the consolidation. Pick the platform that aligns with your existing infrastructure and start building on it.
3-4
major enterprise AI platforms expected to dominate by end of 2025, down from dozens of competing options in 2024
Source: RIVER, strategic analysis, December 2024
Prediction 2: Agentic Workflows in Production
The buzziest term in AI right now is "agentic." AI systems that don't just answer questions but take actions. Read an email, determine the appropriate response, draft it, file the relevant documents, update the CRM, and notify the relevant people. Multi-step, autonomous workflows.
In 2024, agentic AI is demos and prototypes. Impressive demos. But demos. The gap between an agentic demo and an agentic production system is enormous: error handling, rollback mechanisms, approval workflows, audit trails, and the fundamental question of how much autonomy you give an AI system that can take real-world actions.
What we expect:
In 2025, agentic workflows will enter production in specific, constrained domains. Not "general purpose AI agents." Narrow agents with clear boundaries, explicit approval gates, and comprehensive logging.
Claims processing: an AI that reads the claim, extracts the data, checks the policy, and drafts an assessment for human review. Not just the extraction step. The full workflow, with the human at the end as a reviewer rather than an operator.
Customer service: an AI that reads the query, retrieves the relevant information, drafts a response, and either sends it (for simple queries) or queues it for review (for complex ones). Not a chatbot. A workflow agent.
Document management: an AI that receives a document, classifies it, extracts metadata, files it in the correct location, notifies the relevant people, and triggers any downstream processes. The kind of administrative work that currently takes a person 15 minutes and an AI 15 seconds.
What this means for NZ enterprises:
Identify one workflow that's high-volume, well-defined, and currently manual. Design the agentic version with clear human checkpoints. Build it with the expectation that the human-in-the-loop gates will tighten over time as trust builds. Don't try to automate everything. Automate one thing well.
Prediction 3: Sovereign AI Infrastructure
Throughout 2024, data sovereignty has been a discussion. In 2025, it becomes a procurement criterion.
Regulated industries (finance, insurance, health, government) will increasingly require that AI processing occurs on sovereign infrastructure. Not just data storage. Processing. The moment your data enters an AI model, the location and governance of that model matters.
What we expect:
NZ cloud providers and managed service providers will offer "sovereign AI" tiers. Azure's NZ region is already available. AWS's NZ region is coming. The missing piece is not infrastructure availability but governance frameworks: certified, auditable assurance that AI processing meets NZ regulatory requirements.
Organisations that have been using offshore AI APIs (OpenAI, Anthropic, Google) for sensitive workloads will face increasing pressure to migrate to sovereign alternatives. Open-source models running on local infrastructure will become the pragmatic choice for high-sensitivity use cases.
What this means for NZ enterprises:
If you're in a regulated industry and you're using offshore AI APIs for anything involving personal or sensitive data, start planning the migration. The regulatory requirement may not arrive in 2025, but the procurement requirement will. Your clients and partners will start asking where your AI processing happens.
Running your own models on your own infrastructure used to be the hard option. By mid-2025, the tooling will make it the obvious option for any enterprise that cares about sovereignty.
Mak Khan
Chief AI Officer
Prediction 4: The Compound Gap Widens
This is the prediction we feel most confident about, because we're watching it happen in real time.
Enterprises that invested in AI foundations in 2023 and 2024 are now building their third, fourth, fifth capabilities. Each one is faster and cheaper than the last. Their AI capability is compounding.
Enterprises that are still in the "pilot" phase, or that haven't started, are falling behind at an accelerating rate. The gap is not linear. It's exponential. A 12-month head start in AI foundation building translates to 18-24 months of capability advantage, because the early mover is compounding while the late mover is still building their first foundation.
What we expect:
By end of 2025, the gap between AI leaders and AI laggards within the same industry will be visible to customers, competitors, and boards. AI leaders will process faster, serve better, and operate more efficiently. AI laggards will not be able to close the gap by copying the leaders, because the leaders' advantage is in their foundation, not their features.
What this means for NZ enterprises:
If you haven't started building your AI foundation, start now. Not next quarter. Not after the strategy review. Now. The compound effect means every month of delay costs more than a month of catching up. The maths is unforgiving.
Prediction 5: The Talent Model Shifts
This one is specific to New Zealand. The AI talent shortage is acute. There aren't enough ML engineers, data scientists, or AI architects in the country to staff every enterprise's AI ambitions.
What we expect:
The dominant model for AI capability in NZ enterprises will shift from "hire a team" to "partner and build." Fractional AI teams, managed AI services, and AI-as-a-service models will become the default for all but the largest organisations.
The enterprises that succeed won't be the ones that hire the most AI talent. They'll be the ones that build the best partnerships with AI-capable teams and retain enough internal capability to govern, direct, and absorb the output.
2025 Prediction Areas: Expected Impact
Source: RIVER Group, strategic analysis, December 2024
The Scorecard
We'll revisit these predictions at the end of 2025. If we're wrong, we'll say so publicly. That's the deal with predictions: you have to be willing to be measured against them.
Platform consolidation. Agentic production. Sovereign infrastructure. Compound gap. Talent model shift. Five predictions, all grounded in what we've seen in 2024. Let's see how 2025 plays out.