The 'Act' Threshold

By Paul Chada, Co-Founder, DoozerAI

I've been building and selling automation platforms for a while now ;-). BPM, RPA, intelligent capture, you name it. I've seen plenty of "this changes everything" moments that didn't actually change much.

This one is different. And the stock market agrees.

On Monday, Anthropic published a blog post. Just a blog post. It explained how their Claude Code tool could modernize COBOL, the 67-year-old programming language that quietly powers most of the world's banking, insurance and government systems. IBM's stock dropped 13%. Worst day in 25 years. $40 billion in market cap, gone.

From a blog post.

That got my attention. And it should get yours.

The Expert on Speed Dial

Here's how I've been thinking about it.

You have two experts on speed dial. One charges $200/hr. The other charges $400/hr. You default to the cheaper one. Makes sense, right? Why pay double?

But every time you try the expensive one, the answers are sharper. Fewer blind spots. More business acumen. Eventually you stop calling the cheap expert altogether. Not because they're bad. Because you can't afford second-best answers when you're trying to move fast.

Now swap "experts" for AI models.

The cheaper expert is Sonnet 4.x. The expensive one is Opus 4.x.

Opus nails it first time. Code runs. Proposals land. Analysis is sharper. The back-and-forth shrinks to almost nothing. Conversations are shorter because the goal is accomplished quicker. It catches angles that Sonnet (or GPT 5) just doesn't see.

Once you feel the difference, there's no going back.

It's Not Just IBM

The IBM drop wasn't an isolated incident. It's part of something much bigger that's been unfolding all through February 2026.

Traders at Jefferies are calling it the "SaaSpocalypse." Salesforce is down 26% year-to-date. ServiceNow dropped 28%. Thomson Reuters plunged 16% in a single day after Anthropic released a productivity tool for in-house lawyers. Adobe, Intuit, Palantir, all hit hard. A major software ETF is down 27% this year, tracking toward its worst quarter since the 2008 financial crisis.

The pattern keeps repeating. Anthropic ships a new capability. A sector panics. Stock prices crater.

Some analysts say the sell-off is overdone. Jensen Huang called the idea that AI will kill software "illogical." JP Morgan says the market is pricing in worst-case scenarios. Maybe they're right about the speed. But the direction? That feels irreversible.

What the market is really reacting to is a realisation that AI models have crossed a capability threshold. And most businesses haven't caught up yet.

Something Broke Through in December

This isn't gradual progress. Something fundamentally shifted.

Andrej Karpathy posted something recently that stopped me in my tracks. For those who don't know him, he's the former Director of AI at Tesla, a founding member of OpenAI, and one of the most respected voices in AI research. He wrote that it's hard to communicate how much programming has changed due to AI "not gradually and over time in the 'progress as usual' way, but specifically this last December." His take: coding agents basically didn't work before December and basically work since.

Not a year-long trend. A two-month step function.

He shared an example. Over a weekend he needed a local video analysis dashboard for his home cameras. Instead of building it himself, he gave an AI agent a single instruction in plain English: set up SSH keys, deploy a model, build a server endpoint, create a web UI, test everything, configure system services, write a report. The agent ran for about 30 minutes. It hit multiple issues along the way, researched solutions online, resolved them one by one, wrote the code, tested it, debugged it, set up the services, and came back with a finished system.

He didn't touch anything. A weekend project, done in 30 minutes.

His conclusion: "Programming is becoming unrecognizable. You're not typing computer code into an editor like the way things were since computers were invented, that era is over."

Here's what matters for those of us in the business world. The improvements Karpathy is describing aren't just about coding. What changed was the models' reasoning quality, their coherence over long tasks, and their ability to power through problems without giving up. Those same improvements make AI dramatically better at analysis, research, strategic thinking, and any knowledge work that requires sustained reasoning across a complex problem.

In his year-in-review, Karpathy put it bluntly: the industry hasn't realized "anywhere near 10% of their potential even at present capability." We're not waiting for a future breakthrough. It already happened. Most organisations just haven't absorbed it yet.

What the Quality Gap Actually Looks Like in Practice

The conversation around AI models tends to focus on benchmarks and token pricing. For anyone actually using these models in daily work, the difference between tiers is visceral.

With a mid-tier model, you get a solid first attempt. Maybe 70% of the way there. Then you refine. You clarify. You go back and forth. By the fourth or fifth exchange, you have something usable. The model did the work, but you spent 30 minutes managing it.

With Opus, you get a result you can use on the first pass. The code compiles. The analysis accounts for the edge cases you were about to ask about. The strategic recommendation already considered the objection your CFO would raise.

That difference doesn't show up in benchmarks. It shows up in your calendar.

Think about it: if a better model saves you two rounds of iteration on every task, and you run 20 tasks a day, that's 40 exchanges you didn't need to have. Multiply that by the cognitive load of switching back and forth with a tool that almost got it right. The frontier model isn't a luxury at that point. It's a productivity multiplier.

This is why people don't go back once they've experienced it. Not because the cheaper model is bad. Because the time you save with the better model compounds in ways that are hard to measure but impossible to ignore.

Understand. Engage. Act.

At DoozerAI, we've built our entire approach around three phases: Understand. Engage. Act.

My co-founder Gavin and I both come from an RPA/BPM background. We've delivered meaningful automation to businesses over the years. Real results, real ROI. But the knowledge worker tasks were always out of reach. The judgment calls, the nuanced analysis, the work that required someone to actually think about the problem before acting on it. That stuff couldn't be automated. Until now.

For years, the AI industry has been stuck in the first two phases of our framework. Models could understand documents. They could engage with data, answer questions, summarise, extract. Impressive and genuinely useful, but fundamentally passive. The AI understood. It engaged. But it didn't act.

The "Act" was always the hard part. Acting requires reliability. Acting on a 70% accurate result is worse than not acting at all. It creates more work, more errors, more clean-up. This is what held back every RPA implementation I've ever been involved with. The brittle scripts broke the moment a UI changed or a document format shifted.

What's changed, and what the market is now violently pricing in, is that the "Act" phase has become reliable enough for real enterprise work. Not perfect. Not infallible. But reliable in the way that matters: consistent enough that you can build processes around it.

When Anthropic showed that Claude Code can map dependencies across thousands of lines of COBOL, document workflows, and identify risks that would take human analysts months to surface, that's not "understanding" anymore. That's acting. That's doing the work that used to require an army of consultants and a seven-figure budget.

IBM's stock didn't drop because investors suddenly learned COBOL was old. Everyone knew that. It dropped because the market realized that AI can now reliably take action on that knowledge at enterprise scale. The "Act" threshold got crossed, and it crossed faster than anyone's financial models predicted.

Where This Goes

December was two months ago.

That's how long it's been since Karpathy says the capability threshold was crossed. Two months. And in that time, IBM lost $40 billion in a day, software stocks entered a bear market, and the entire SaaS industry started having an existential crisis.

Most organisations haven't even updated their AI strategy since last quarter.

If the models that couldn't reliably act in November can reliably act in February, what do they look like in June? In December? The pace here isn't annual. It's monthly. And the gap between companies that are building on frontier intelligence and companies that are still "evaluating options" is widening every week.

The question isn't whether AI will transform how your business operates. The question is whether you'll be the one doing the transforming, or the one getting transformed.

Understand. Engage. Act.

The "Act" is here. And it's more reliable than most people realize.