The last quarter of 2025 felt less like “AI is coming” and more like “AI just moved into your spare room, changed the Wi‑Fi password, and started shipping features.” If Q3 was about shiny models, Q4 was about infrastructure, regulation and a quiet but very real shift in what it means to build products as a PM.
1. AI is no longer a feature, it’s the environment
Remember when “let’s add an AI feature” sounded bold and visionary? In Q4 it started to sound more like “let’s add CSS.”
A few things crystallised:
- Consumer spend on AI apps grew triple‑digit year‑on‑year, with productivity and dev‑tools eating most of the pie.
- CEOs started talking less about “experimenting with AI” and more about “AI as a line item in the P&L and infrastructure strategy.”
- The interesting part: access is basically universal. The moat is no longer “we use a powerful model” but “we actually designed a product that does something useful with it.”
If you are still pitching “we’ll plug in a model and see what happens,” you are playing last season’s game.
2. Welcome to the Agent Era (for real this time)
In Q3, agents were mostly a buzzword plastered on top of chatbots. In Q4 they started to look like systems. Not “ask me a question,” but “give me a goal and I’ll go do things.”
On the supply side:
- Major players doubled down on agentic AI: orchestration, tools, memory, multi‑step workflows became first‑class citizens in the ecosystem.
- The infra world woke up: AI‑optimised data platforms, vector‑native storage and “AI data engines” became fashionable ways to say “we’d like your entire data stack, thanks.”
On the product side, this changes your job in a very specific way:
- You’re no longer designing single interactions, you’re designing behaviours over time.
- You stop asking “what’s the prompt?” and start asking “what is the loop: observe → decide → act → learn?”
If your discovery deck still frames AI as “smart autocomplete,” you’re underestimating what’s now possible – and what your competitors are testing.
3. Regulation stopped being background noise
Q4 was the quarter when AI regulation moved from conference slides to Jira epics.
A few highlights:
- The EU AI Act moved into the “this is happening, plan for it” phase, including obligations for general‑purpose models and high‑risk use cases.
- The US took a slightly different turn with a new executive order focused on keeping innovation and competitiveness alive while still nudging companies toward responsible AI.
- China and a handful of other countries pushed new standards around safety, transparency and governance for generative systems.
For PMs, this isn’t just legal trivia. It affects:
- How you log model outputs and user interactions.
- How much explainability you need to bake into the UI.
- Where your data can physically live – and which model you’re even allowed to call in a given region.
“Move fast and break things” now comes with compliance officers, DPIAs and model risk committees attached.
4. Infrastructure quietly became the real power play
While social feeds argued about benchmarks, someone signed cheques worth hundreds of billions to build AI infrastructure.
- Investment in AI data centers and compute crossed the 300B USD mark, with mega‑campuses measured in gigawatts becoming a thing.
- Countries started talking seriously about sovereign AI: local models, local data, local compute as a strategic asset.
This trickles all the way down to product decisions:
- “Which model should we choose?” becomes “how do we design so we can swap models and regions without rewriting everything?”
- Architecture choices (multi‑cloud, on‑prem options, edge vs cloud) stop being purely technical and turn into product strategy tools.
If your PRDs don’t include at least a paragraph on “AI architecture assumptions,” future‑you will curse present‑you.
5. AI tools for PMs grew up a little
Q4 also brought more AI for product managers, not only AI in products.
- We saw more vertical copilots – including ones tailored to financial product design and regulatory‑heavy workflows.
- Generic “summarise this doc” quietly evolved into “help me design an experiment, generate a first PRD draft, and flag risk in this backlog.”
The interesting twist is adoption:
- A clear majority of PMs now use some form of AI in their daily work, but the spread is huge between “I ask ChatGPT to rewrite my emails” and “I run my discovery, roadmap scenarios and opportunity sizing through an AI stack.”
The gap between those two groups is likely to widen. Not because of tools – everyone has them – but because of workflow design.
6. So what do you actually do with this?
If you’re a product manager trying to make sense of Q4 2025, here’s the short version you can paste into your own notes:
- Treat AI as infrastructure, not add‑on. Identify the parts of your product where agents and automation can sustainably take over workflows, not just single clicks.
- Design with constraints in mind: regulation, data locality, model governance, observability. The boring stuff is now part of the core value proposition.
- Invest in agent‑native discovery: think in terms of goals, environments, and feedback loops, not prompts and screens.
- Upgrade your own stack: pick 2–3 AI tools and go deep, not wide. Your edge won’t be “I use AI,” but “my product practice is re‑architected around it.”
Q4 didn’t give us one big “AI moment.” It gave us something more dangerous and more interesting: AI becoming part of the background. The water we all swim in. And once something becomes the water, opting out stops being a strategy. It just means you’re the last one to notice you’re already underwater.