AI Signal – February 2026

February's big finding: most organisations are using a fraction of the AI capability they've already paid for. Research published this month shows AI agents can handle tasks lasting five hours — but the median real-world use is forty-five seconds. That gap is what we're calling the deployment overhang, and it's this month's deep dive.

Elsewhere: AI costs collapsed again, always-on AI agents went from a hacker experiment to over a million users in days, revenue numbers settled the "is AI real?" question, and agent teams built working software with no human orchestrator.

TLDR

At a Glance

Model Releases This Month

---

Key Signals This Month

Emerging Signals

The "SaaSpocalypse." Investors are pricing in broad disruption of traditional software companies. Some major enterprise software stocks dropped 20–37% in early 2026. The concern: AI-native tools will replace entire software categories, not just augment them. The reality is more nuanced — large organisations move slowly, and most existing tools won't vanish overnight — but if you're renewing expensive software contracts, it's worth asking whether an AI-native alternative now exists.

AI's physical limits. Training a single frontier AI model will soon require gigawatts of electricity — equivalent to powering a small city. The energy and infrastructure demands of AI are becoming a real constraint. If sustainability is part of your organisation's mandate, the environmental footprint of AI tools is a question that's going to get harder to ignore.

A tightening knowledge-work market. Professional services job openings hit an 11-year low in February, according to industry data. Hiring rates in knowledge-work sectors have dropped significantly. Whether this is AI-driven, cyclical, or both is debated — but the practical implication is the same: the value of existing staff who can work effectively with AI is going up, not down.

A generational shift in career attitudes. Nearly three-quarters of Gen Z parents now say they'd prefer their children to become trade entrepreneurs rather than tech employees. Only 16% believe a university degree guarantees job security. The future workforce may be less concentrated in the knowledge-work roles AI is disrupting most.

AI as infrastructure, not just a tool. One of the month's most thought-provoking arguments: AI may be better for plumbers than programmers. The logic is that reduced software costs make previously uneconomic niche markets viable — scheduling tools for trades, inventory systems for small workshops, custom apps for local businesses. If this plays out, AI's biggest impact may not be in the industries that talk about it most.

---

The Framework: February Deep Dive

Each month we map signals to four tiers that cover everything from the tools themselves to how you actually use them. This is the structured analysis behind the headlines above.

Tier 1: The Landscape

What's available – and what changed

The cost of intelligence keeps falling. The same level of AI capability that cost a certain amount 21 months ago now costs 150 times less. That's not a typo. And it's not slowing down — in February alone, one major provider released a model that users prefer over their previous top-of-the-line option, at one-fifth the price. Google matched it with a model that doubled its own benchmark scores at zero price increase.

Revenue numbers have settled the "is AI real?" question. AI companies are generating revenue at a pace with no precedent in software. The fastest any traditional software company reached $10 billion was over 20 years. AI labs are doing it in about one year. Anthropic's coding tool — used by developers to write and edit code — now accounts for 4% of all public code on GitHub. That figure doubled in a single month. Whether or not your organisation uses AI today, the market has made its decision.

Always-on agents arrived — practically overnight. OpenClaw, a platform that turns AI models into persistent assistants running 24/7, went from a small community to 1.5 million agents in days. Dedicated hardware sold out. Early adopters are running teams of AI agents on small computers at home — handling continuous research, monitoring, and project management while they sleep. The highest-value use case so far isn't autonomous coding (which requires too much human feedback) but persistent research: agents that continuously scan, catalogue, and surface relevant information. This is a new category of AI use that didn't exist at scale a month ago.

The "best" model still changes every week. The leading AI companies — OpenAI (ChatGPT), Anthropic (Claude), and Google (Gemini) — are trading benchmark leadership on a near-weekly basis. The practical takeaway from January's guide still holds: stop trying to pick the right model. Build the team capability and workflows that work regardless of which model is on top.

Tier 2: The Foundation

What your organisation needs in place

AI adoption is becoming mandatory. Accenture — with over 700,000 employees — now ties AI usage to career progression. Three other major consultancies confirmed that senior managers are the hardest group to convert. The signal is clear: in large organisations, AI isn't optional anymore. If this pattern spreads (and it will), "I don't use AI" stops being a personal preference and starts being a career risk.

Trust is becoming a competitive differentiator. In a landscape where capability is converging — most AI tools can do roughly the same things — the companies behind those tools are differentiating on values and trust. Anthropic committed to keeping their consumer product permanently ad-free. As AI tools become more capable, the question of which providers you trust with your data and workflows is becoming a real business decision, not just a technical one.

Autonomy isn't something you buy — it's something you build. A detailed study on how people use AI agents found that the degree of autonomy an agent exercises isn't determined by the technology alone. It's co-constructed — shaped by the model's capability, the product design, and the user's skill level. You can't just purchase an "autonomous AI." You have to design the workflows, train the people, and calibrate the trust level. There's no shortcut.

Tier 3: The Practice

How your team should actually work with AI – this month's deep dive

This isn't about needing better tools. It's about using what you already have.

The January edition showed that how people use AI matters more than which AI they use — the delegation vs inquiry finding. February extends that insight with hard numbers. The gap isn't just about interaction style. It's about scope. Most people are asking AI to do small, safe, familiar tasks when it's capable of handling complex, multi-step, hours-long work.

Expert users don't approve more — they monitor differently.

The same data revealed a clear pattern in how experienced AI users differ from new ones:

The shift from new to expert isn't "hands off." It's "hands different." Expert users give more trust upfront but intervene more assertively when something matters. They monitor rather than approve. This is a learnable skill, not an innate talent — and it's the skill that closes the deployment overhang.

The AI is already better at knowing when to stop than most people expect.

One surprising finding from the same dataset: AI agents stopped themselves to ask for clarification twice as often as humans interrupted them — 16% of the time on complex tasks, compared to 7% human interruption. The safety concern that AI will "run away" with a task is less supported by the evidence than the concern that humans won't push AI far enough.

Tier 4: The Application

Where AI is being applied – and what's emerging

Agent teams can self-organise. In one of February's most striking experiments, a research team tasked sixteen AI agents with writing a working C compiler — 100,000 lines of code that boots an operating system. No human orchestrator. The agents coordinated using shared text files, essentially inventing their own project management. The reported cost: approximately $20,000. The lesson isn't about compilers — it's that well-designed AI teams can handle genuinely complex, multi-week projects. The key design decisions: write very thorough tests, structure output for AI to read (not humans), and give each agent a clearly defined speciality.

AI agents work beyond code — and the data proves it. Over half of all AI agent actions in a major study were in non-engineering functions. Back office operations (9%), marketing (4.4%), sales and CRM (4.3%), and finance (4%) are all active agent deployment areas. If your organisation has been thinking of AI agents as a developer tool, it's time to widen the lens.

Hard ROI numbers are appearing. Walmart's AI shopping assistant increased average basket size by 35%, with half of online customers actively using it. A senior engineering team at Spotify reported they haven't written a single line of code since December — all coding happens through AI, with engineers reviewing and directing rather than typing. These aren't pilot programmes. They're production deployments at scale.

---

Framework Check

---

What To Do This Month

---

AI Signal is published monthly by Pandion. We help organisations build real AI capability – the foundations, the practice, and the fluency that turn tools into results.

Have a question about something in this guide? Get in touch.

FAQs