The daily working routines of software engineers in Silicon Valley have shifted considerably over recent years. The result remains unchanged: line after line of programming code. Instead of typing, they compose their work by speaking aloud into microphones. Those directions are then converted into text via artificial intelligence and passed to AI agents — automated programs capable of interacting with software and, in this instance, producing it.

Colin Jarvis of OpenAI has noted that many white-collar workers have yet to integrate this technology into their professional lives. Whether that shift materialises will determine the trajectory of his organisation and the broader artificial intelligence expansion. To recoup their substantial investments in data centre infrastructure, cloud computing companies and AI developers such as Anthropic and OpenAI depend on continued broad uptake of the technology. Rough estimates indicate that AI-related annual revenue would need to climb to approximately $US2.5 trillion ($3.5t) by 2030 — compared with roughly $US150 billion today — to offset the expenditure involved in constructing data-centre infrastructure.

To date, adoption has been most pronounced among software developers. According to Stack Overflow, a developer community platform, roughly 80% of respondents reported using an AI-assisted coding tool. By mid-2025, the collective annual recurring revenue across four AI coding ventures — Cognition, Cursor, Lovable and Replit — had reached approximately $US800 million. The figure has since reached $US6 billion.

According to estimates from SemiAnalysis, a research firm, coding-related work represented more than half of the combined annual recurring revenue generated by Anthropic and OpenAI during the second quarter of this year. This pattern is consistent with the observation that artificial intelligence has spread more rapidly across sectors where software developers represent a higher proportion of the workforce.

Legal services, financial advising, and customer support are among the sectors most frequently discussed as prime candidates for AI-driven automation. Each is recording expansion. Three AI legal technology firms — Harvey, Clio and Legora — have seen their combined annualised revenue reach approximately $US1 billion, nearly twice the figure recorded a year ago. According to the company, attorney usage of the Harvey platform has been growing at a rate of roughly 100% month over month.

Financial institutions and investment professionals are increasingly adopting artificial intelligence across their operations. Rogo, a financial-services-focused artificial intelligence company, reported a 50 per cent rise in annualised recurring revenue alongside the onboarding of 100 enterprise clients over the most recent quarter.

Artificial intelligence is gaining traction across the customer service sector. By June, Sierra — a developer of purpose-built AI agent tools — had reached $US200 million in annual recurring revenue, twice the level recorded the previous November. Private equity investors are expressing optimism about the outlook. Investments in AI startups within this sector have reached $US3 billion so far this year, outpacing every other category of AI application. Funding for AI-coding ventures has plateaued, partly due to the expanding reach of proprietary coding tools developed by AI laboratories themselves.

All three of these tools share certain characteristics with programming. These markets hold considerable commercial potential, though the output tends toward standardised production. Four distinguishing characteristics place coding in a separate category: the volume of available training data, the relative ease of verifying model outputs, the degree of human involvement in the work, and the nature of software engineers as a professional group.

The widespread availability of publicly shared open-source code across the internet has contributed to the strong performance of AI tools in software development tasks. In certain instances, it has been incorporated into training datasets, representing close to 20 percent of the total input. Publicly accessible material on topics such as customer grievances or contract negotiations exists in far smaller quantities.

Verifying AI-generated code is generally considered a straightforward process. In engineering, automated validation tools routinely verify the behaviour of other code. This raises the probability that mistakes within AI-produced code will be identified before they create issues. Establishing whether a client's dissatisfaction was genuine, or whether a financial model accurately reflected the nuances within a company's accounts, is a notably complex undertaking.

One benefit of code-writing tools is that a substantial portion of the data needed to carry out a given task is already embedded within a programme's existing codebase, or accessible via the standardised digital connections between software systems known as application programming interfaces. Yet in numerous other situations, such data exists only in analogue form — held within the expertise of a call-centre operative or conveyed through the unspoken signals of an opposing party's legal representative.

Those firms are actively addressing a range of obstacles tied to artificial intelligence development. Winston Weinberg, chief executive of Harvey, has noted that developments in artificial intelligence now allow the company to train its systems partly through the use of data generated by AI models themselves. Companies like Mercor are bridging the gap between AI firms and domain specialists, enabling those professionals to channel their knowledge into training models.

Ping Wu, chief executive of Cresta, argues that analysing high volumes of call-centre transcripts enables the AI customer-service firm to surface unspoken employee knowledge — including workflows around handling refund requests. Some AI firms have begun embedding engineers directly at customer locations to troubleshoot problems and improve their offerings.

Among the elements driving AI's advances in programming, the role of software engineers themselves may prove the most difficult to replicate. Their familiarity with technical disciplines and routine exposure to evolving programming environments positions them among those who tend to embrace new tools readily. According to Anish Acharya of the venture-capital firm Andreessen Horowitz, the uptake of AI coding tools has largely been driven from the ground up, with software engineers independently discovering and trialling them on their own initiative.

Call-centre staff are generally observed to lack both the motivation and the autonomy to trial new tools independently, tending instead to receive technology through management-led rollouts. Those working in law and finance are required to navigate a complex web of professional regulations governing their respective fields. Broader adoption of AI across various domains is expected to expand over time, with the pace of that expansion likely to depend on advances in underlying models and the degree to which organisations grow willing to delegate tasks to automated systems.