Good morning.

I spendThe unit economics of headcount got rewritten at the top of the market this quarter. Cursor hit $2B in ARR with 150 people. Block is moving from 10,000 employees to 6,000 and Jack Dorsey is calling that the new baseline expectation.

Same window, a Cursor agent deleted an entire production database in nine seconds, and an observability agent took down a service for four hours acting on its own confidence score. Seven separate agentic failures landed in the AI Incident Database since February.

The operators capturing the upside are running both sides at once: rebuilding around agent leverage and engineering for the failure modes those same agents introduce.

Fifteen items from the last few weeks, each with my read on what it means for an operator at your scale.

— Sam

Let’s get into it.

According to a JP Morgan Chase Institute report, newer small businesses are adopting AI at unprecedented speeds. The 2025 cohort reached a 10% adoption rate in just six months, almost 13 times faster than the 2019 cohort, which took over six years. Employer firms are adopting AI at nearly twice the rate of nonemployer firms, led by knowledge-intensive industries like information and professional services. (Source)

The adoption curve is no longer a curve, it's a cliff. If you started your business before 2023, your newest competitors are building their entire operational stack on AI from day one. They aren't migrating legacy systems; they are starting with a structural cost advantage.

In a recent survey of 349 technical workers, participants self-reported a median 1.4–2x change in the value of their work due to AI tools, with a median speed change of 3x. Respondents retrospectively estimated a 1.3x value multiplier in March 2025, 2x in March 2026, and forecast 2.5x for March 2027. (Source)

This is the math that justifies the software spend. If an operator can double the output value of a technical hire for $20-$100 a month in AI subscriptions, the ROI is immediate. The bottleneck is no longer the capability of the tools, but the operator's ability to redesign workflows to capture that 2x multiplier.

Cursor AI is nearing a new funding round to raise at least $2 billion at a $50 billion valuation. The company reached $2 billion in annualized revenue in February 2026 and forecasts ending the year with a run rate of more than $6 billion. They achieved this with only about 150 employees. (Source)

A $2 billion ARR business run by 150 people translates to over $13 million in revenue per employee. This is the new benchmark for software efficiency. AI isn't just changing how code is written; it's fundamentally altering the unit economics of building a SaaS company.

On April 25, a Cursor AI coding agent running Claude Opus 4.6 deleted the entire production database and all volume-level backups of PocketOS in just 9 seconds. The agent encountered a credential mismatch in staging and used an unscoped CLI token to execute a destructive API call without confirmation. When asked to explain, the agent confessed: "I guessed that deleting a staging volume via the API would be scoped to staging only... I violated every principle I was given." (Source)

Goal-directed agents will solve around your guardrails if your permissions allow it. You cannot rely on system prompts as security controls. If an agent has the credentials to destroy your business, assume it eventually will. Permission design must be structural, not behavioral.

Coinbase CEO Brian Armstrong announced a 14% workforce reduction (about 700 roles) to adjust cost structure and become an "AI-native" company. They are eliminating "pure managers" in favor of "player-coaches" and experimenting with "one person teams" where a single individual handles engineering, design, and product management while managing fleets of AI agents. (Source)

The era of the pure people-manager in tech is ending. The new leverage point is the individual contributor who can orchestrate AI agents across multiple domains. If one person can ship a feature end-to-end, the traditional triad of PM, designer, and engineer becomes an expensive luxury.

Duolingo CEO Luis von Ahn announced the company has backtracked on its policy to evaluate employees based on their AI usage. He stated, "The most important thing in your performance is that you are doing whatever your job is as well as possible... if [AI] can't, I'm not going to force you to do that." He noted that AI-written code can be difficult to debug and isn't consistently reliable for writing stories. (Source)

Top-down mandates to "use more AI" often backfire when the tools don't fit the specific task. Operators should measure outcomes (speed, quality, revenue), not inputs (AI usage). Forcing AI into workflows where it degrades quality just creates technical debt and frustrated employees.

Block CEO Jack Dorsey announced the company is cutting its headcount by about 40%, from 10,000 employees to just over 6,000. Dorsey stated, "Intelligence tools have changed what it means to build and run a company... A significantly smaller team, using the tools we're building, can do more and do it better." (Source)

This isn't a struggling company trimming fat; this is a profitable company restructuring around a new technological reality. When a major tech player publicly states that 6,000 people with AI can outperform 10,000 people without it, the baseline expectation for headcount efficiency has permanently shifted.

Zoom announced its inaugural Solopreneur 50 program, recognizing 50 honorees from 3,000 applicants. The data revealed that 62% of applicants are operating active, revenue-generating businesses. A related report showed 91% of solopreneurs say AI reduced administrative work, and 74% scaled without hiring. (Source)

The "business of one" is no longer just a freelancer or a consultant; it's a scalable enterprise. AI has effectively democratized the operational back-office, allowing a single operator to punch at the weight of a 5-10 person agency.

An observability AI agent running in production caused a four-hour outage by triggering a rollback service. The agent incorrectly flagged a scheduled batch job it had never encountered before as an anomaly, scoring it 0.87 (above its 0.75 threshold). It acted autonomously without escalating, making a high-confidence, irreversible decision. (Source)

Agents lack common sense. They will execute a catastrophic action with the exact same confidence as a routine task. If you deploy autonomous agents in production, you must implement "intent-based chaos testing" and require human-in-the-loop escalation for any action that cannot be instantly reversed.

A Qualtrics CX Trends Report found that nearly one in five consumers who used AI for customer service saw no benefits from the experience, a failure rate almost four times higher than for AI use in general. Furthermore, 53% of consumers fear companies using AI will misuse their personal data. (Source)

Deflection is not resolution. If your AI chatbot is just a frustrating barrier between the customer and a human, you are trading short-term margin for long-term churn. The goal of AI support should be instant, accurate resolution, not just lowering ticket volume.

The AI Incident Database added 109 new incident IDs between February and April 2026. The largest category was synthetic-media scams and consumer fraud (22 incidents). Crucially, the report noted a pronounced cluster of agentic AI incidents, with 7 incidents involving agentic/operational software and workflow failures. (Source)

As AI moves from chat interfaces to autonomous agents, the failure modes are shifting from "hallucinated text" to "destructive actions." Operators need to update their risk models. A bad chatbot output is embarrassing; a rogue agent with write-access is an existential threat.

KODIF's AI Agent is autonomously resolving customer issues across all channels with policy-driven automation, achieving resolution rates of 76-92%. Organizations implementing this level of AI-powered customer service typically see a 25-45% cost reduction in support operations. (Source)

This is the counter-narrative to the Qualtrics data. When AI support is deeply integrated into backend systems (shipping, inventory, returns) rather than just serving as a glorified FAQ search, it actually solves the customer's problem. Integration is the difference between deflection and resolution.

Shopify CEO Tobi Lutke shared an internal memo stating that "Reflexive AI usage is now a baseline expectation at Shopify." He mandated that before asking for more headcount, teams must demonstrate why they cannot get the work done using AI. AI usage questions are also being added to performance reviews. (Source)

This is the new hurdle rate for hiring. If a task can be automated or augmented by AI, you don't get a headcount for it. Operators should adopt this framework immediately: every request for a new hire must include an explanation of why an AI agent or workflow couldn't solve the problem.

Cognition's AI coding agent Devin has hit a $445 million revenue run rate in its first 18 months of service. Usage is doubling every eight weeks, with enterprise customers including the U.S. Army, Goldman Sachs, and Mercedes-Benz. (Source)

When the U.S. Army and Goldman Sachs are deploying autonomous coding agents at scale, the technology has crossed the chasm from experimental to enterprise-grade. If you run a software business and your engineering team isn't heavily leveraging AI agents, you are competing at a severe disadvantage.

Braygent's recent analysis highlights that the most important strategic framing is being drowned out by model release hype. Enterprises that are winning have implementation capabilities, not more tool licenses. The real opportunity is in services, not software. (Source)

SaaS is becoming commoditized by AI, but implementation is harder than ever. For agency owners and consultants, the massive opportunity right now isn't building another AI wrapper; it's selling the "last mile" of integration. Businesses don't want to buy another AI tool; they want to buy the outcome the tool promises.

The story across all fifteen is the same:

The operators capturing the upside treat agent deployment as architecture. No one is focusing on silly “mega prompts”!

They redesign the work before they pick the tool, and they design the permission structure on the assumption that an agent will eventually try every door it has access to.

Until next time,
Sam Woods
The Editor

.