The tools are 20% of it. The other 80% is harder.

The companies winning at AI are changing what they expect from people when using tools.

Ramp's CPO published their adoption playbook this week. 99.5% of employees use AI. Non-engineers are submitting 12% of production code changes. They shipped 1,500 internal apps in six weeks. Impressive numbers.

The first instinct is to ask "which tools did they use?" Skip that question. The tools are Claude, ChatGPT, the usual suspects. What matters is that Ramp made AI proficiency a performance expectation, removed every approval bottleneck on tokens and connectors, and built a public leaderboard so builders got recognized.

The second question, maybe more relevant: are those 1,500 apps actually useful, or is Ramp shipping for the sake of shipping? A leaderboard that rewards volume... gets you volume. The same headaches that made internal software painful before AI (maintenance, documentation, knowledge transfer when someone leaves) don't disappear because the tool was written in an afternoon (though, still cool!). Eventually, the culture needs to reward building, but it also needs to draw a line between "solves a real problem" and "impressive demo that nobody uses two weeks later."

So the Ramp numbers are real but, if you ask me, the framing around them has a bit of press release energy (remember the Klarna AI doing most of support that was later backtracked?). The companies that will pull ahead won't just remove friction for builders. They will also decide what "good" looks like once the app is shipped and define how that app is kept relevant. And that's a whole other playbook.

Read their playbook here

Claude Opus 4.7 Is Out Today and the Non-Developer Story Is Worth Knowing

What happened. Anthropic released Claude Opus 4.7 today across all Claude products, including the paid tiers at claude.ai. No API required to use it. The headline improvements are in complex, multi-step tasks: the model now checks its own reasoning before reporting back, catches logical errors during planning (not just after), and handles longer autonomous runs without going off the rails. Vision got a meaningful upgrade too: it reads charts, diagrams, and dense documents at higher resolution, which matters for anyone feeding it reports, financials, or slides.

The reality. Most of the launch testimonials are from software companies testing it on code. I've been running it for a few hours and still undecided if I feel a difference. On the non-developer angle: two things worth noting. First, the improved vision should be helpful (my initial tests are rather positive!), second a model that says "I'm not sure, let me verify" instead of confidently filling gaps with plausible nonsense is worth having - let's see if that's the case here in practice. Same price as the previous version on Claude.ai paid plans.

Ramp Published Its Internal AI Playbook. The Numbers Are Absurd.

What happened. Ramp's CPO shared the fintech company's full AI adoption strategy publicly. The company-reported results: 6,300% increase in internal AI usage over the past year. 99.5% of employees actively using AI tools. Non-engineers now write 12% of production code. These are self-reported figures from Ramp's own leadership, not independently audited.

The reality. Ramp is an engineering-heavy fintech company. Replicating 99.5% adoption in a services firm will be harder. But the playbook's 8 levers are organizational, not technical: proficiency ladders, visible leaderboards, removing approval bottlenecks, baking AI into hiring. Even with my disclaimers on if this output level with AI-generated tools is sustainable, this is worth watching and thinking through (tools don't matter as much as we might think).

Canva Launches AI 2.0 With Agentic Design Features

What happened. Canva unveiled its AI 2.0 overhaul as a research preview. The design tool is positioning itself to let users direct multi-asset campaigns using natural language while agentic AI (AI that takes actions on its own, not just answering questions) handles execution. You describe a campaign and Canva builds the assets across formats.

The reality. This launched as a research preview, so not all Canva users will see the new interface immediately. Canva's advantage is its 220M+ monthly users. The AI features land inside a tool they already know. Whether output quality matches a skilled designer on real brand work is to be seen but design is certainly poised to see the same acceleration coding did.

Snap Cut 1,000 Jobs and says AI Writes 65% of Their Code

What happened. Snap laid off 16% of its workforce, roughly 1,000 people. CEO Evan Spiegel explicitly cited AI-driven efficiency as part of a broader restructuring push targeting $500 million+ in annualized savings. AI now writes over 65% of Snap's new code (per Snap's own investor presentation), letting smaller teams maintain the same output. Activist investor Irenic Capital had pushed for a 21% workforce cut, citing AI replacement and Snap delivered 16%.

The reality. On one end, I'm not entirely sure this is as much about AI as it sounds. It's a lot nicer to say you reduced workforce because of AI than to say you did so because you're still trying to get to "solid profitable" territory or you overhired in the covid years. On the other, Snap is a tech company with highly automatable coding workflows. The 65% figure won't translate to client services or knowledge work. But the precedent matters: a public CEO cited AI as a factor in reducing headcount, on the record. One of the clearest public examples of a company framing layoffs as partly AI-driven.

“Companies investing seriously in AI are seeing real returns.”

PwC's 2026 AI Performance Study, out April 13, found that 74% of AI's financial gains go to just 20% of organizations. But the same study found that 56% of companies report no significant financial benefit from AI at all.

The headline is getting circulated as "AI works if you do it right." But my read on the actual finding is bleaker: most companies are spending on AI without seeing returns, not because the technology doesn't work, but because the majority of AI activity is "vibe experiments": ad-hoc use, without a coherent strategy, without the organizational changes to support sustained adoption and without the incentive structure to make people care.

PwC found that the technology itself delivers roughly 20% of an AI initiative's value. The remaining 80% comes from redesigning work. Most organizations invest heavily in the 20% (tools, models, access) and underinvest in the 80% (workflow redesign, governance, reskilling, outcome measurement).

The Goldman Sachs read is similar: "We still do not find a meaningful relationship between productivity and AI adoption at the economy-wide level" — though they do find a 30% productivity gain in two specific localized use cases (coding and customer service).

What we still don't know

Whether the 20% of winners are structurally different companies (larger, better-resourced, more technically mature) or whether their practices are genuinely replicable. PwC says it's the practices. I don't necessarily know that to be true. Change takes time, a lot more than technologists (me!) estimate and certainly less than what consultancies sell.

Source: PwC

Paste the full report into Claude or ChatGPT. Add this prompt:

Follow up with:

Spot-check the top 3 claims against the original document. Add your own judgment. Send.