AI is transforming productive work. Previously constrained by technical expertise and organisational capacity, individuals can increasingly deliver work that once required a larger team or an established service firm.
At OpenMercury, we have witnessed this transformation first-hand. Against our pre-AI baseline, we saw productivity rise to ~1.5x with ChatGPT in 2024, 8.3x with Claude in mid-2025, 14.8x with AI agents in late 2025, and 27.6x with more capable agents in the first half of 2026. AI agents now carry out much of our operational, financial, and legal work.
As agent capabilities continue to expand, should companies hire more experienced talent for better results or save on hourly rates? Should they build with agents internally or save by outsourcing?
This article examines those questions through one metric:
Here, execution rate is how much work gets done in a given period. Success probability is the chance that the work achieves the required outcome.
While we cannot give a universal answer for every industry, we use publicly available AI-usage data to identify four early signals: 1) judgement creates a wider performance gap; 2) the economics of cheap labour are reversing; 3) better economics for talent as entrepreneurs; and 4) agentic marketplaces foster better economics for both talent and businesses.
AI does not amplify everyone equally
People continue to decide what should be built and what counts as done. In a study, Anthropic found that people made about 70% of planning decisions in a typical session, while Claude made about 80% of execution decisions.
Expertise matters. Anthropic compared agent actions, output per prompt, and verified success across novice, intermediate, and expert users:
| User type | Agent actions | Agent output (words) | Success rate |
|---|---|---|---|
| Novice | 5 | 600 | 15% |
| Intermediate | 8 | 1,700 | 28% |
| Expert | 12 | 3,200 | 33% |
Expert sessions set off action chains more than twice as long, carrying five times the output. Their verified success rate also more than doubled.
Turning expertise into agentic capacity is itself a skill. In a separate study, OpenAI found that 80.6% of sampled Codex users had submitted work estimated above 30 minutes, but only 25.6% had crossed the eight-hour threshold.
Learning speed is almost as powerful as deep expertise. Most of the gain in agentic performance came from moving from novice to intermediate; between intermediate and expert, the slope decreased. Verified success rose from 15% for novice-rated sessions to 28% for intermediate sessions and approximately 33% for expert sessions. When sessions encountered trouble, abandonment also fell from 19% for novices to 5-7% for intermediate and expert users. Learning fast gives users similar levels of persistence and ability to steer the agent in the right direction.
Occupation is also becoming less decisive. In code-producing Claude Code sessions, users in software-related occupations achieved verified success approximately 34% of the time, compared with 29% for users from other professions.
Better talent is becoming cheaper to hire
Before AI, people remained constrained by human execution speed. An expert’s higher success rate often could not offset their higher hourly rate during budgeting. Businesses are often restructured by moving tasks across geographies and levels of seniority.
AI agents add a new execution multiplier. In Anthropic’s report, expert prompts produced five times the output of novice prompts, with verified success of approximately 33% versus 15%. Using output as a rough execution proxy, the successful-execution capacity becomes:
This 11x successful-execution advantage reverses the economics of hiring and corporate structure. It exceeds the 2026’s 8x gap in average base salary for a US software engineer ($138,000) versus Rs 14.5 lakh ($16,500) in India. It also exceeds the ~3x junior-to-senior total-compensation gap at FANG companies. AI-fluent talent is now a key source of cost savings.
Tokens also contribute to hourly cost. For a typical developer, Claude Code costs $3-7 per hour on API billing. This is lower than human labour, but not negligible. Failed work still consumes resources. In the same study, 19% of troubled novice sessions were abandoned, compared with 5-7% for experts.
AI-native workflows widen the gap further. Experts can encode repeatable knowledge in agent instructions, skills, and plugins instead of restating it for every task. By doing so, they shift much of the execution cost from human hours to tokens, creating another estimated 50-70% in cost savings. Consequently, an expert-led, AI-native service can reach a striking 20-30 times the cost efficiency.
Top talent is increasingly choosing entrepreneurship
While some are driven partly by corporate layoffs, top talent now has a clear economic incentive:
This refers to the gap between what an individual can produce and what they earn as an employee.
Wages adjust more slowly than productivity. From Q1 2025 to Q1 2026, US corporate profits rose 12.8%. By Q2 2026, workers’ inflation-adjusted pay was down 0.1% year over year, while the share of income generated by the economy going to workers fell to 52.9%, the lowest since 1947. It is a history-high incentive to become an owner and capture the surplus.
Starting fresh beats upgrading. AI execution is outpacing organisational coordination. OpenAI found that 43.5% of occupation-specific messages crossed into another occupation, but only 25% of organisations expect cross-functional agent coordination within two years. Data access, approvals, and integration are increasingly becoming the bottleneck.
Traditionally, leaving a corporation meant losing access to complementary labour. AI agents reverse the trade-off: independent talent with AI can assemble better complementary support and achieve a lower cost per successful outcome.
Hourly pay caps the reward for good judgement; ownership does not. On Stripe, more than 2x as many solopreneurs earned over $1 million in 2025 as in 2023, and nearly 3x as many crossed $5 million and $10 million. AI-influenced journeys now account for nearly 4x the share of Stripe sign-ups compared with last year. As AI makes execution cheaper, top talent can increasingly own the upside instead of selling the hours.
Your best talent is now a freelancer with agents
The reversal in hiring and talent economics changes the logic of what to keep in-house, what to hire externally, and how talent should be compensated. Superstar talent is much harder to beat merely by adding investment without a supportive structure.
Build permanent capabilities; hire defined outcomes.
| Build internally when | Hire externally when |
|---|---|
| The capability is continuous and core | The need is specialised or intermittent |
| The approach must evolve with strategy | The outcome and acceptance criteria are clear |
| You need daily control over execution | A provider can own delivery end-to-end |
Your best provider may not want another job. Experienced operators can now use AI agents to serve multiple clients without building a large firm. Companies can access their judgement and outcomes without employing every capability full-time.
At OpenMercury, we make outcome-based hiring work for businesses and individuals. We built RenX, an agent marketplace, for this mission. Buyers define the target outcome, standard, and price. AI-fluent providers use their own judgement and agents to deliver through pay-for-results contracts. Buyers pay for outcomes, while capable independent providers retain more of the value they create.
