This Week in AI & Automation

Week of March 3, 2026

OpenAI dropped GPT-5.4 on Wednesday with a feature that changes the game: native computer use. Accounting startup Basis became a unicorn at $1.15 billion by deploying AI agents that file tax returns autonomously. Lio raised $30 million to let AI agents run enterprise procurement. And Jensen Huang promised a chip reveal at GTC next week that will "surprise the world." This week's theme: AI stopped being a tool you use. It became a worker that uses tools.

The Big Story

OpenAI Ships GPT-5.4 — The First Model That Uses Your Computer

OpenAI released GPT-5.4 on March 5, and the headline feature isn't benchmarks — it's computer use. GPT-5.4 is the first general-purpose model from OpenAI with native ability to operate across applications on your machine autonomously. It can open files, navigate software, fill forms, and execute multi-step workflows without human hand-holding.

The model ships in three variants: standard GPT-5.4, GPT-5.4 Thinking (replacing GPT-5.2 Thinking for Plus/Team/Pro users), and GPT-5.4 Pro for high-performance workloads. The API version supports context windows up to 1 million tokens — the largest OpenAI has ever offered.

Performance matters too. Individual responses are 33% less likely to contain errors compared to GPT-5.2, and the model solves problems with significantly fewer tokens than its predecessor. That means lower API costs for the same quality output.

Source: OpenAI, TechCrunch, 9to5Mac

Our Take: Computer use is the bridge between "AI that answers questions" and "AI that does work." This is what turns a chatbot into an agent. For enterprises running complex workflows across multiple systems, the implications are significant. But the real test isn't capability — it's reliability. Can GPT-5.4 navigate your ERP without breaking something? That's where production-grade implementation still matters more than model capability.

Notable Developments

Basis Hits $1.15 Billion — Agentic Accounting Gets Its First Unicorn

Basis closed a $100 million Series B led by Accel and GV (formerly Google Ventures) at a $1.15 billion valuation — a 4.6x jump from its $250 million Series A just six months ago. Former Goldman Sachs CEO Lloyd Blankfein joined the round alongside Khosla Ventures.

What Basis built is genuinely different from traditional accounting software. Instead of giving accountants better tools, Basis deploys AI agents that perform accounting tasks autonomously — tax preparation, audit trails, compliance monitoring, and financial reporting — with human oversight only at critical decision points. The company recently demonstrated the first AI agent to complete an end-to-end 1065 tax return without human intervention.

About 30% of the top 25 US accounting firms and 20% of the top 150 now use the platform. Total funding: $138 million.

Source: Bloomberg, Yahoo Finance

Our Take: This is what agentic AI looks like in a specific vertical. Not a general-purpose assistant — a domain expert that handles the entire workflow. The finance function is where AI agents deliver the clearest ROI because the work is rule-bound, high-volume, and error-prone. Basis proved there's a billion-dollar market in automating what accountants spend 70% of their time doing: data entry and compliance checks.

Lio Raises $30M to Deploy AI Agents in Enterprise Procurement

Lio (formerly askLio) raised a $30 million Series A led by Andreessen Horowitz, with SV Angels, Harry Stebbings, and Y Combinator participating. Total funding: $33 million.

The Berlin-based company builds AI agents that handle enterprise procurement — sourcing, vendor management, purchase orders, and contract negotiation. Since 2023, the platform has managed billions in enterprise spend for Global 2000 and Fortune 500 companies including Munich Re, Brose, and Novozymes. The new capital funds US expansion and product development.

Source: TechCrunch, PR Newswire

Our Take: Procurement is one of those back-office functions where AI agents shine because the work is repetitive, data-heavy, and expensive to do manually. The a16z bet validates what we see in the market: enterprise AI implementations that target specific operational workflows outperform horizontal AI tools by 3-5x on ROI. Lio's challenge will be the same as every enterprise AI company — integrating with the messy reality of legacy systems.

Quick Hits

• Apple ships MacBook Air with M5: Pre-orders started March 4, availability March 11. The M5 chip brings expanded on-device AI capabilities — more models run locally without cloud dependency. Apple Newsroom
• AI agents hit 57% production deployment: New data from G2 and PwC shows 57% of companies now have AI agents in production, 22% in pilot. Up from under 5% of apps embedding agent capabilities in 2025 to 40% in 2026. G2 Report
• Samsung targets 800M AI devices by end of 2026: Samsung doubles down on Google Gemini integration, aiming to put generative AI in mid-tier and budget phones — not just flagships. The mass-market AI hardware play is accelerating.
• Mega seed rounds keep getting bigger: Humans& (an AI lab from ex-Google/Anthropic/OpenAI researchers) raised $480M in seed financing. Over 40% of seed and Series A funding in 2026 has gone to rounds of $100M or more. The early-stage bar has moved permanently.

Numbers of the Week

What We're Watching

NVIDIA GTC next week will set the hardware agenda for 2026. Jensen Huang delivers his keynote March 16 in San Jose and has explicitly promised a chip reveal meant to "surprise the world." With 30,000 attendees from 190 countries, GTC is where the AI infrastructure roadmap gets written. Whatever Huang unveils will determine what's buildable — and affordable — for the next 18 months.

Agentic AI is splitting into two tiers: horizontal and vertical. OpenAI's GPT-5.4 with computer use represents the horizontal play — a general agent that can navigate any application. Basis and Lio represent the vertical play — purpose-built agents for specific domains. Both approaches work, but for enterprise buyers the build-vs-buy calculus now includes a third option: deploy vertical agents that already understand your domain. Expect more vertical agent unicorns in Q2.

The 1-million-token context window changes what's possible. GPT-5.4's API context window means you can feed an entire codebase, a full quarter of financial statements, or a complete contract library into a single prompt. For enterprise AI implementations, this eliminates one of the biggest technical constraints: the need to chunk, retrieve, and reassemble context via RAG. It doesn't kill RAG — but it makes many RAG implementations unnecessary.

The Bottom Line

This week, AI stopped being something you prompt and started being something you deploy. OpenAI gave models the ability to use computers the way humans do. Basis proved that autonomous AI agents can handle an entire professional function — not just assist with it. Lio showed that a16z is betting real money on AI agents replacing manual procurement workflows. And next week, Huang will show us the hardware that makes all of it faster.

The shift is structural. AI isn't getting incrementally better at answering questions. It's learning to do the work. For enterprises still running proof of concepts, the window between "experimenting with AI" and "competing against companies that deployed AI agents" is closing fast.

Get This Week in AI & Automation delivered every Saturday.

Previous Editions

• This Week in AI & Automation: The 40% Cut (Feb 28, 2026)
• This Week in AI & Automation: The $100 Billion Week (Feb 21, 2026)
• This Week in AI & Automation: The Age of Agent Teams (Feb 14, 2026)
• This Week in AI & Automation: The $470 Billion Question (Feb 1, 2026)

Have a story we should cover? Contact us.