Someone asked me this question at a meetup last week. Then again in a team standup. Then a client brought it up in a strategy session. Everyone is asking the same thing right now:
“What are the AI trends we should be paying attention to? What do we need to do to stay relevant?”
I’ve been sitting with this question for a while. Here’s my honest answer — not the conference-slide version, but the one I’d give a close colleague over coffee.
First: Filter the Signal from the Noise
The AI landscape in 2026 is drowning in predictions. Gartner forecasts that 40% of enterprise applications will embed AI agents by end of 2026. McKinsey says 62% of organizations are already experimenting with agents. Multi-agent system inquiries on Gartner’s platform surged 1,445% from Q1 2024 to Q2 2025.
These numbers are real. But there’s also the noise: Gartner warns that over 40% of agentic AI projects will be cancelled by 2027 due to unclear business value and cost overruns. And of the thousands of companies claiming to offer “AI agents,” only about 130 actually deliver genuine agentic capability. The rest are “agent washing” — rebranding chatbots and rule-based automation with a new label.
So the first skill you need is discernment. Not every trend deserves your attention. Here’s what I think actually matters.
The 4 Trends Worth Your Real Attention
1. Agentic AI is graduating from lab to production
The shift from “AI as a feature” to “AI as a worker” is real and accelerating. We’re past the demo phase. Companies are deploying agents that don’t just answer questions — they plan, execute multi-step tasks, call tools, make decisions, and loop back to verify their own work.
The key word is autonomous. Not AI that helps you write an email. AI that monitors your sales pipeline, identifies at-risk deals, sends follow-up drafts, updates your CRM, and notifies your team — without you initiating each step.
This is not sci-fi. This is what production agent systems look like in 2026. The Gartner market number ($52 billion by 2030, from $7.8 billion today) reflects genuine traction, not hype. The teams shipping real value understand something critical: agents need guardrails, not just capabilities.
2. Smaller models will eat more of the market than large ones
The obsession with “biggest model wins” is fading. Gartner predicts that by 2027, task-specific small language models (SLMs) will be deployed at 3x the volume of general-purpose LLMs.
Why? Because a 7B-parameter model fine-tuned on your customer support transcripts beats GPT-class models on your specific task — at 10x lower inference cost, running on-device, with no data leaving your servers. For enterprises with strict data privacy requirements or latency constraints, SLMs aren’t a compromise. They’re the right tool.
The SLM market is projected to grow from $0.93 billion in 2025 to $5.45 billion by 2032 (28.7% CAGR). This is not a niche. This is where a huge portion of practical enterprise AI will live.
What this means for you: The answer to “which AI model should we use?” is no longer “the biggest, smartest one.” It’s “the smallest model that can reliably do this specific job.”
3. The entry-level squeeze is real — and it’s moving up
Entry-level software engineering positions have seen a 73% hiring drop in the past year (Ravio, 2025). This isn’t because AI replaced junior developers. It’s because one senior developer using AI tools now produces what used to require a team of three to five. The headcount math changed.
This is moving up the org chart. Mid-level engineers who do well-defined, scope-clear work — implementing a feature from a spec, writing boilerplate, porting code between formats — are facing pressure too.
But here’s the nuance: BLS projects 17% job growth for software developers through 2033, even accounting for AI. The job is not disappearing. The job is changing. The question is which version of the job you’re doing.
4. Multimodal + physical AI is coming — but this is a 3-5 year wave
Gartner projects 80% of enterprise applications will be multimodal by 2030. IBM’s researchers are explicitly flagging robotics and physical AI as the next frontier after language models hit diminishing scaling returns. Vision + language + action agents that can perceive and operate in the real world are not theoretical.
But be honest with yourself about timelines. Physical AI is a 2028-2030 story, not 2026. Focus on what’s deployable now: multimodal understanding (text + image + document) in customer-facing and internal products is very real and underused today.
What This Means If You’re a Technical Lead
I’ve spent a lot of time thinking about how this changes my own role. Here’s what I’ve landed on.
Your new job description: AI system designer
You are no longer primarily a code author. You are an AI system designer — someone who decides which tasks go to which model, how agents are orchestrated, where guardrails are placed, and what happens when things go wrong.
This requires:
1. Architecture judgment on AI costs As I wrote about in my post on the LLM tax: the engineers who thrive are the ones who know when NOT to use an LLM. Static analysis for a structural code question. A regex for pattern matching. An algorithm for sorting. Reserve LLMs for what they’re actually good at: semantic understanding, reasoning under ambiguity, and generation. Every token you save is money that can fund the complex use cases that actually need it.
2. Knowing the model landscape — not just one vendor Opus for deep reasoning. Sonnet for most production tasks. Haiku for high-frequency, low-complexity calls. Open source SLMs (Qwen, Llama, Mistral) for on-premise or privacy-sensitive workloads. Multi-model routing is a real engineering discipline now. You need to know the tradeoffs.
3. Observability and evals for AI systems You need to know how to measure whether your AI system is doing what you think it’s doing. LLM-as-judge patterns, golden dataset evals, latency + cost dashboards, hallucination detection — these are engineering fundamentals now. If you’re shipping AI features without evals, you’re flying blind.
4. Security and trust boundaries Prompt injection is a real attack surface. Agent permissions need to be as tight as any other API authorization. What can the agent read? Write? Delete? Who can override it? These are engineering decisions, not afterthoughts.
The skills to build right now
In order of priority:
- Agentic system design — how to architect reliable agent workflows with retries, fallbacks, and human escalation points
- Evaluation methodology — how to build benchmark datasets and measure model quality for your specific domain
- Infrastructure cost modeling — how to estimate, forecast, and govern AI infrastructure spend
- Prompt engineering + fine-tuning — not as a party trick, but as a production engineering discipline
What This Means If You’re a Product Lead
The PM role is being rewritten. The meeting-heavy, PRD-writing, Jira-managing version of the job is being automated. What survives — and what you need to double down on — is different.
You now own the human-AI boundary
The hardest product decisions of the next three years are not “should we add AI to this feature?” They’re: where exactly does human judgment end and AI autonomy begin?
- Should the agent just draft the email, or send it?
- Should the recommendation be shown to the user, or acted on automatically?
- Who reviews the agent’s output, and in which cases does it override itself?
These are not engineering decisions. They are product and design decisions that require deep user empathy, ethical reasoning, and business judgment. This is your territory.
Feature-first thinking is dead. Capability-first is the new model.
The old PM playbook: identify user pain, define feature, write spec, ship, measure.
The new playbook: identify pain, ask what capability an AI agent could develop to address it, design the human-agent collaboration model, define what “good enough” looks like for autonomous operation, build evals, then ship.
Notice what changed: you’re designing a behaviour, not a feature. The agent learns. It improves. The product is never “done” in the same way a static feature is done. This requires a different relationship with your engineering team and a different definition of done.
You need to prototype, not just spec
The old world: write a spec, hand it to engineering, wait three weeks for a prototype.
The new world: you should be able to string together an LLM call, a RAG pipeline, and a basic UI in an afternoon — not to write production code, but to validate whether your idea even works before spending engineering time on it. Tools like Cursor, Claude artifacts, and no-code AI builders make this possible. If you’re not prototyping your own AI feature hypotheses, you’re slower than you need to be.
The 3 skills Product Leads need right now
- RAG architecture literacy — understand how retrieval-augmented generation works well enough to ask the right questions and spot bad implementations
- AI evaluation design — be able to write a test plan for an AI feature, including what “good” looks like and how to measure degradation over time
- Agent UX patterns — understand how to design for progressive autonomy: show → recommend → auto-act with override → fully autonomous. Know when each mode is appropriate.
The Mental Model Shift That Matters Most
Both Technical Leads and Product Leads need to internalize one shift that underlies all of the above:
We are moving from a world where software does what it’s told, to a world where software makes decisions.
This changes everything about your accountability. When a deterministic system produces a wrong output, you debug the logic. When an AI agent makes a bad decision, you need to understand: was it a training data problem? A prompt design problem? An eval gap? A hallucination? A misaligned objective?
The answer requires a different kind of thinking. Not just “does the code work?” but “is the system behaving as intended across the full distribution of inputs?”
This is the real adaptation. Not learning a new tool. Not getting a certification. Developing the judgment to design, evaluate, and govern systems that are probabilistic, not deterministic.
A Practical Roadmap for the Next 6 Months
If I were starting from scratch today, here’s what I would do:
Month 1-2: Get your hands dirty Ship one real AI feature. Not a demo. Something in production with real users. Learn the full cycle: prompt → eval → deploy → monitor → iterate. Use the cheapest model that works.
Month 3-4: Go deep on agents Build one agentic workflow — even small. A pipeline that takes a trigger, calls multiple tools, makes a decision, and produces an output. Understand where it fails and how to make it robust.
Month 5-6: Build governance muscle Set up cost dashboards for your AI usage. Create an eval harness for your most critical AI features. Write a one-page AI decision framework for your team: what requires LLM, what doesn’t, and how you decide.
The teams that are 12 months ahead right now didn’t move faster because they had better tools. They moved faster because they treated AI as infrastructure to be engineered properly — not magic to be applied liberally.
The Honest Answer to “What Should I Do?”
Stop waiting for AI to stabilize before investing in it. It won’t. The landscape will keep shifting. The advantage goes to people who develop good judgment faster — judgment about what AI can and can’t do reliably, what it’s worth spending on, and how to design systems that keep humans in control of the decisions that matter.
That judgment only comes from building things, breaking things, and learning from both.
Get started. Stay critical. Keep learning.
References
- The trends that will shape AI and tech in 2026 | IBM — IBM’s predictions on agentic AI, SLMs, and physical AI for 2026
- AI agent trends 2026 report | Google Cloud — Google Cloud’s enterprise AI agent adoption analysis and market data
- 7 Agentic AI Trends to Watch in 2026 | MachineLearningMastery — Comprehensive breakdown of multi-agent systems and orchestration patterns
- 10 AI and machine learning trends to watch in 2026 | TechTarget — Enterprise ML trends including SLMs, multimodal, and governance
- Explore Agentic AI Market Trends 2025-2026 | Svitla — Gartner’s 1,445% surge in multi-agent inquiries and market sizing
- AI Trends 2026: Post-Hype Enterprise Impact | Instinctools — Analysis of “agent washing” and the 40% project cancellation risk
- AI Impact on Software Engineering Jobs 2026 | AIToolRanked — Job posting analysis: 40% increase in AI tool requirements, 73% entry-level drop
- Software developers are the vanguard of how AI is redefining work | WEF — World Economic Forum on developer roles in the AI era
- Future Outlook of Software Engineering in 2026 and Beyond | Lemon.io — 65% of developers expect role redefinition; BLS 17% growth projection
- The Future of Software Engineering with AI | Pragmatic Engineer — Gergely Orosz’s analysis of how engineering roles are splitting
- Essential AI Skills for Product Managers in 2026 | Amoeboids — Technical fluency in 78% of PM job postings; RAG and agent design skills
- AI Product Manager Roadmap 2026 | NextByRahul — Skills roadmap: agentic AI, MCP, systems thinking, prototyping
- How AI is Changing the PM Role | SkipLevel — Shift from feature-first to capability-first product thinking
- 7 AI-Proof Careers That Will Survive 2026 and Beyond | FinalRoundAI — WEF analysis: 15-25% growth in human-judgment roles through 2030
- How will AI Affect Jobs 2026-2030 | Nexford University — McKinsey: 57% of US work hours potentially automatable, 60% of occupations have partial exposure