🌊 The 10 tasks driving Anthropic's $10B run
Where AI value is concentrating (from millions of real conversations)
👋 I’m Ivan. I write about venture capital waves.
This week’s sponsor is LaurenAI by Flippa.
LaurenAI helps you buy online businesses. Tell LaurenAI what you want (SaaS, app, ecom, content site, YouTube + budget). It returns a target list - including businesses not listed for sale, and lets you email owners from one place:
Set your criteria in 60 seconds
Get matched with businesses within your budget
Send personalised messages that actually get replies
Hello there!
Anthropic’s revenue has exploded from $1B to $10B in 2025:
Which begs the question:
What is AI actually doing all day in the real world?
I dove into Anthropic’s Economic Index Report, which is the cleanest answer I’ve seen so far because it looks at anonymised Claude usage in November 2025 across both their consumer chat (which I’ve been using a lot lately) and their first-party enterprise API.
So as usual, here are my top key takeaways:
1. AI usage is still shockingly concentrated
A small number of tasks still explains a huge chunk of AI value.
The details
Anthropic sees over 3,000 unique work tasks on Claude.ai, but the top 10 tasks account for 24% of sampled conversations. That concentration also shows up even harder in enterprise API usage, where the top 10 tasks are 32% of traffic.
This is the opposite of the “AI does everything now” vibe. It’s more like early SaaS where lots of possible use cases popped-up everywhere, but only a few pulled $$.
Also worth noting that the single most common task is still basically “fix the bug”.
So what
This is how real tech diffusion looks where it typically starts narrow where the ROI is more obvious and then it fans out. The money today seems to still be in a small set of painful frequent jobs where AI already works reliably enough.
2. Coding is still the core “AI job”
Software work dominates, especially in enterprise.
The details
Across Claude usage the biggest block is still Computer & Mathematical work. On Claude.ai it’s roughly a third of usage, and in the first-party enterprise API it’s closer to 50%.
Another interesting trend to follow is "Office and Adminstrative Support”, which is super interesting considering what is happening this very week with the parallel rise of ClawdBot / MoltBot / OpenClaw (stay tuned).
So what
The first wave of AI adoption is the economy paying to speed up people who already work with systems. If you are building, this translates to build where outcomes can be proved (i.e. “We saved engineers 30%” > “we improved thinking”).
3. Most people use AI like a co-worker, not an autopilot
Collaboration is the default behavior in chat.
The details
On Claude.ai, Anthropic classifies usage as either “automation” or “augmentation”. In November 2025, augmentation is 52%, while automation is 45%.
Augmentation” is back-and-forth: edit, refine, rewrite, sanity check, iterate. Not “do the whole thing and I disappear”.
And the trend is interesting because Anthropic shows automation had risen earlier, then came back down, with augmentation becoming dominant again. They mention product changes like file creation and workflow skills as possible drivers.
So what
This is why AI feels like productivity and not replacement (mostly, and for now). Most value right now comes from people who stay in the loop and use AI as a force multiplier like yours truly over here.
4. AI saves the most time on hard thinking work
AI saves more time on complex work than simple work.
The details
Anthropic estimates “speedup” by comparing how long a task would take a human alone vs human plus AI. When they plot that against task complexity (proxied by years of schooling needed to understand the prompt), the curve goes up.
In Claude.ai conversations, tasks around high school level show about 9x speedup, while college-level prompts show about 12x speedup.
So what
This is one of the most founder-relevant facts in the report because it suggests that AI is not only “junior labor” but also clearly “senior leverage”, because it can collapse time for high-skill tasks like analysis plans, architecture thinking, and complex writing.
That pushes a lot of advantage to teams that already have taste and good judgment because they can run more cycles per week.
5. But harder work also fails more often
Speed comes with fragility.
The details
As task complexity rises, success rates fall.
Simple tasks succeed roughly in the high-60s to low-70s percent range.
More complex tasks dip below that.
This is the gap behind most Christmas-table conversations that start with
“AI is amazing, but…”. The model can move fast, but it is not always right. And the cost of being wrong rises quickly as the work gets harder.
So what
This is why the next wave is not just “bigger models” and you see big rivers of capital going into tooling, evals, monitoring, guardrails, workflows, and product design that makes failure cheap and detectable.
6. AI breaks when a job runs for too long
AI struggles to stay on track for a long time without stopping.
The Details
They also look at how often AI succeeds as a job gets longer (which is different from our previous point, hard ≠ long, not always):
Short jobs usually work.
As jobs run longer, success drops.
For API-based automation, success falls below 50% after about 3.5 hours.
So what
If you ask AI to run one long process from start to finish, it will often break, and to make this work, you either:
keep the job very small, or
build a system that can stop and recover.
7. Breaking work into steps makes AI work much better
Iteration beats autonomy.
The details
The same chart tells a second story. On Claude.ai, success rate declines much more slowly with task duration. Anthropic extrapolates that Claude.ai would hit 50% success at about 19 hours.
Their explanation is that multi-turn chat breaks complex tasks into steps and each turn becomes a feedback checkpoint.
So what
The agentic future is not “more autonomy” but better scaffolding with memory, checkpoints, validation, retries, and human review at the right moments.
Stepping back, the picture is pretty consistent: we are seeing AI doing real work, but mostly in narrow, repeatable, high-skill tasks. It tends to work best with humans in the loop, breaks on long horizons, and creates the most value where problems are clearly defined.
👉 Try an AI agent that helps you find businesses you can buy:
🤖 LaurenAI
🌊 Top Research Reports
CB Insights - Physical AI Models Market Map
McKinsey & Company - Age of AI report
CB Insights - Top 14 2026 Tech Trends
Dealroom - European Spinouts Report
Anthropic - Technology Adolescence
Anthropic - Economic Index
A16Z - State of Markets
KKR - 2026 Outlook
🌊 VC Wave Tracker:
A list of 30 startups that just raised Series B, all likely hiring:
Upwind Security - cloud runtime security (US) $250M
Harmattan AI - industrial vision AI (FR) $200M
JetZero - all wing aircraft (US) $175M
Defense Unicorns - defense software (US) $136M
Railway - app hosting platform (US) $100M
Cambium - defense materials (US) $100M
Tandem - AI healthcare ops (US) $100M
Ethernovia - vehicle networking (US) $90M
Superstate - onchain funds (US) $82.5M
Apella - AI clinical copilot (US) $80M
Emergent - AI agent platform (US) $70M
Listen Labs - AI user research (US) $60M
Accelsius - data center cooling (US) $55M
Aikido Security - app security scans (BE) $60M
Hydrosat - water satellite data (US) $55M
Orbital - legal due diligence (UK) $60M
Ivo - AI contract review (US) $55M
Zocks | AI for Advisors - advisor meeting AI (US) $45M
Benepass - employee benefits (US) $40M
AnswersNow - autism telehealth (US) $40M
Articul8 AI - enterprise genAI (US) $35M
hashtag#Compa - pay benchmark data (US) $35M
GovDash - gov contract tools (US) $30M
Unbox Robotics- warehouse robots (IN) $25M
Visitt - property ops tools (IL) $20M
VelaFi - nonprofit banking (US) $20M
MontyCloud - cloud governance (US) $18M
CloudSEK - cyber threat intel (SG) $20M
Klir - water utility software (US) $20M
Eat App - restaurant ops software (US) $16M
AppliedAI - enterprise automation (DE) €12M
3 clear waves show up in this list:
AI for real work: Security, contracts, research, healthcare, ops.
Infrastructure for AI scale: Cooling, networking, chips, cloud control.
Regulated markets modernizing: Defense, water, government, finance.
That’s it!
If you want to help me out, the best thing you can do is share this newsletter with someone who’d find it useful. 🙏🤙
$1 Billion to $10 Billion ARR in 12 Months 🤯
Brilliant breakdown of the Anthropic data, the concentration insight is eyeopening. The finding that augmentation beats full automation matches what I've seen in practice—teams that keep humans in the loop tend to iterate faster and catch edge cases. I'm particularly intrested in the task duration curve you showed; it suggests the real opportunity isn't longer autonomy but smarter chekpoints. Thanks for pulling this together so clearly.