Two AI studies dropped in the same week. Read together, they tell you exactly where to deploy AI hard,  and where to stop. 

Bottom line 

Half of heavy Claude users say AI already handles 50% of their work, and we’ve all seen it first-hand. From the LinkedIn posts to the email messages, to the overall engagement, despite the evolution of AI we’ve seen, it’s often easy to detect the difference between when a human composed that email vs. AI. But it goes further than that; a Princeton simulation just showed that AI agents, given 500 days to run a fictional business, mostly go broke and a simple rule-based heuristic with no AI in it beat nearly all of them. 

Both findings are true. The gap between them is where every real AI decision in your organization lives. 

If you blur that gap, you’ll outsource business judgment to a system that lacks that very capability. But if you respect it, you’ll get productivity gains and keep the steering wheel. This simply reaffirms a key position of LumenForge Advisors: AI is a tool that can make our lives simpler and HELP make decisions, but we never want to rely on it for those decisions.  

What’s happening: the optimistic signal 

In its June 2026 Economic Index [1][2], drawn from a survey of roughly 9700 Claude users across Chat, Cowork and Code, they reported: 

• Approximately 50% of respondents say AI can already handle 50% or more of their work tasks. 

• 26% expect AI to cover 60-90% of their work within 12 months. 

• 4% believe Claude could already do their entire job. 

• The power users, the heaviest, most experienced ones at least, are the most optimistic about their own career prospects. Early-career workers are the most anxious. 

Net: real productivity gains are landing at the task layer, and the people closest to the tool see their skills becoming more valuable, not less. This also reinforces the view that integration between human users and AI are the ideal solution.  

Grounding call: This is self-reported data from Anthropic’s own customers. People who paid for Claude, used it enough to form a strong opinion, and bothered to respond to a survey. That’s a stacked deck, a selection effect, not a representative sample of the workforce. The trend is real. The magnitude needs salting. 

What’s also happening: the grounding counterweight 

The same week, Princeton researchers released CEO-Bench, a simulation where AI agents have to run a fictional SaaS company called NovaMind for 500 simulated days [3][4]. The agents start with $1M in the bank, zero customers, and a Python API of 34 tools. The agents then set pricing, run ads, manage R&D, scale infrastructure, negotiate with enterprise clients, and read noisy signals from a simulated social network. Drop below zero in cash once and the company is dead. 

The results are the part you need to internalize: 

• Of 14 tested models, only three finished their best run above starting capital: Claude Fable 5 ($47.15M), Claude Opus 4.8 ($27.8M), and GPT-5.5 ($21.3M). 

• The theoretical ceiling for cash on hand is roughly $2.2 billion. The best agent hit about 2% of it. 

• Meanwhile, a simple rule-based heuristic comprised of fixed prices, fixed quotas, focused ad spend, capacity adjusted to recent usage,  and zero AI involvement finished at $15.76M, beating every model except the top three. 

• Models that survived without profiting (Claude Opus 4.7) did so by defaulting to cost-cutting and cash preservation instead of exploration. Survival is not business strategy. 

To reinforce: a static decision tree outperformed nearly every frontier model at running a small business. 

That is not a story about AI being bad. It’s a story about what AI is for; it’s NOT to do it all for you, it’s not the EASY button, it’s a tool for your toolbox. A powerful tool, yes, but still just a tool.  

Why the gap is the point 

We all know that task execution and business judgment are different problems, requiring different skills and methodologies. The Princeton authors say it plainly: tool-using agents handle short-horizon tasks with clear goals, brief actions, and fast feedback well. Real management is the opposite: long chains of decisions under uncertainty, hidden variables, delayed payoffs, shifting context, and competing priorities you have to weigh against each other. 

The four capabilities that correlated with success in CEO-Bench tell you exactly what’s missing from current agents at scale: 

1. Uncovering hidden information (e.g., which ad channel works for which segment) 

2. Predicting the future (cash forecasting over weeks) 

3. Adapting quickly to a competitor’s move 

4. Planning ahead with if-then scenario reasoning 

That list is not a description of “tasks.” It’s a description of management through the lens of HUMAN judgement. And current models, even the best ones, are inconsistent at it across long horizons. Compressing the simulation to 50 days didn’t fix the problem either; most models still struggled to coordinate decisions toward even a short-term goal. 

The most useful sentence from the Princeton write-up, for anyone evaluating an agentic AI pitch: when a rule-based heuristic with no LLM beats your $X/month frontier agent, the question isn’t whether AI is capable. It’s whether AI is the right tool for the decision you’re trying to delegate. And choosing the right tool is a human-centered decision. 

What this means for SMB leaders 

Two truths, held at the same time: 

Where AI is winning right now: drafting, summarizing, querying structured data, generating artifacts, coding, accelerating analysis. Anthropic’s own data shows blog writing, marketing content, and database queries as the top work-driven uses [1]. Stack those gains across a team, and you get real, compounding throughput. That’s the 50% offload number; that’s not hype, that’s a workflow win. 

Where AI is not winning is in the space that requires sustained judgment across hidden variables, delayed feedback, and shifting priorities. Pricing strategy. Resource allocation across quarters. Reading what a customer actually wants versus what they’re saying. Knowing which signal in the noise is the one that matters. That’s still a human job, and CEO-Bench just put a number on how much of a human job it still is. 

The anxious early-career worker and the optimistic power user are responding to the same tool differently. The difference is not the AI, it’s experience, context, and judgment. People with developed judgment see AI as a force multiplier on what they already do well. People without it fear being replaced by something that, frankly, also doesn’t have judgment. 

That asymmetry is the whole game. 

The move 

Here’s where I land with clients: 

• Push AI hard at the task layer. Audit your workflows. Find the drafting, summarizing, querying, and artifact-generation work that eats hours and doesn’t require sustained judgment. Find the routine stuff that can be reasonably automated and requires miminal judgement call; the data is what the data is. Offload that work; the data is on your side. 

• Hold the line at the judgment layer. Do not let “agentic” become a synonym for “unsupervised.” When a vendor pitches you autonomous business decision-making, ask them how their agent performs on the four CEO-Bench capabilities. If they don’t have an answer, you have the answer. 

• Build the human muscle. Train your people along two paths: 1) how to engage with AI in a manner that simplifies their tasks, and 2) to read AI output critically and understand when to trust it, when to override it, or when to ignore it entirely. That skill is the moat. Without it, the 50% offload becomes a 50% liability. 

• Govern the deployment. This is the gap LumenForge was built for. Most SMBs are buying AI faster than they’re building the governance to deploy it safely. They are deploying solutions that may, but most likely may not, address their real business needs. The OEA exists because the autonomous AI marketing message has outrun the operational reality, and someone has to do the unglamorous work of closing that gap. 

Closing thought 

AI is a tool. The most capable tool any of us has ever held. It crunches data brilliantly, but data does not exist in a vacuum; it exists alongside human thought, decision, and consequence. Experience, judgment, and the willingness to be wrong and adjust are still the job. The frontier models we have today are extraordinary at parts of that job. They are not, yet, qualified to do the whole thing. 

Use the tool. Don’t outsource the work to it. 

Sources 

[1] Bastian, M. (2026, June 27). Half of Claude users say AI can already handle half their work according to Anthropic survey. The Decoder. https://the-decoder.com/half-of-claude-users-say-ai-can-already-handle-half-their-work-according-to-anthropic-survey/ 

[2] Anthropic. (2026, June). Anthropic Economic Index — June 2026 Report. https://www.anthropic.com/research/economic-index-june-2026-report 

[3] Schreiner, M. (2026, June 28). Only three AI models finished above starting capital in a 500-day startup survival test. The Decoder. https://the-decoder.com/only-three-ai-models-finished-above-starting-capital-in-a-500-day-startup-survival-test/ 

[4] Chen, Narasimhan, & Liu. (2026). CEO-Bench: Measuring Long-Horizon Strategic Decision-Making in Language Model Agents. arXiv:2606.18543. https://arxiv.org/abs/2606.18543 

Dan Bond is the Founder and Principal AI Advisor at LumenForge Advisors, helping SMBs deploy AI as a force multiplier — without outsourcing the judgment that still belongs to humans. lumenforge.ai