The market is no longer debating whether AI can write code
PANews said the most successful application story in AI over the past two years has been coding. The progression has been clear: early code completion, then automated test generation, bug fixing, codebase migration, and now coding agents that can read an entire project, use the terminal, run tests, revise code repeatedly, and commit changes. The article argues that this was the first time AI proved it could do more than chat or generate content. It could enter a real production environment and perform work that companies are willing to pay for, with outcomes that can be verified and value that can be measured.
That success also created the next problem. Coding agents have moved from being an expectation gap to becoming a market consensus. The debate is no longer about whether AI can code, but about how many developers it can reach, how much software engineering work it can replace, and how large the revenue pool can become. PANews wrote that coding still has substantial fundamental room to grow, but the story of “AI can write code” is becoming less capable on its own of supporting ever-richer valuations across the broader AI supply chain.
The article says the market is now focused on two questions. First, is there another AI application category large enough to follow coding. Second, if that handoff does not happen soon, what will support today’s expanding spending on data centers, GPUs, networks, and power, along with the corporate debt tied to that expansion. Put differently: where will the revenue, free cash flow, and productivity gains come from to justify the capital spending, and will AI returns on invested capital actually hold up.
PANews said the second issue is already beginning to show up in early models for cloud-provider ROIC during the second-quarter earnings season, but the first question remains the core one. The tension is not whether AI is useful. It is that the first growth engine has already become consensus, while the second may take much longer to appear in revenue, profit, and cash flow. That lag, in the article’s framing, is the most important timing gap in AI applications right now. The next phase of pricing needs proof that AI can do more than reshape software development. It needs to show that AI can take over a wider set of economic activity.
Why coding became AI’s first breakout application
The article argues that coding did not emerge first simply because programmers are more willing to test new tools. It worked because software tasks line up with nearly every condition AI commercialization needs.
Code, logs, APIs, documentation, and error messages are already fully digitized. AI does not need to interpret the physical world first, and companies do not need to spend decades cleaning up paper records. Results are also highly verifiable. A code change either compiles or it does not; tests pass or fail; an API returns the right response or it does not; a page runs or it breaks; performance improves or degrades. That fast feedback loop lets AI make an attempt, observe what happened, revise the work, and try again.
Software development is also decomposable. Reading a repository, understanding a request, editing files, running tests, analyzing failures, revising the output, and generating a commit record form a workflow that naturally fits persistent agent execution. Labor costs are high as well, so even a 10% to 20% reduction in developer time can produce an ROI that companies can calculate quickly.
Another point in coding’s favor is that partial correctness still has value. AI does not need to replace an engineer from day one. If it can generate tests, write SQL, explain code, make front-end edits, or migrate APIs, companies already have a reason to pay. And coding products can spread from the bottom up. Individual engineers can subscribe, install, and use them without waiting for lengthy enterprise procurement, compliance review, data integration, or organizational redesign.
For those reasons, PANews describes coding as the first AI application to combine product maturity, defined budgets, measurable results, high-frequency usage, and a closed commercial loop. The article adds that enterprise application spending materials show coding is already the largest single department-level AI application. Even so, the combined scale of horizontal copilots, general assistants, vertical AI, and consumer AI remains larger than coding alone. In that sense, coding is the most commercially mature, the most intensely used, and the easiest-to-measure single AI workload, but not the sum of all AI demand.
The search for “the next Coding” may start with the wrong question
PANews says many investors and operators imagine the next phase as another product that looks like a coding agent: a clear category, very fast growth, major revenue scale, and something the capital market can identify quickly. The problem is that the non-coding world is not structured like software development.
Programmers around the world share code repositories, IDEs, version control, testing practices, and deployment patterns with strong commonality. That makes coding easier to define as a unified market. Non-coding work is far more fragmented. Banks need credit, KYC, and anti-money-laundering agents. Insurers need underwriting, claims, and anti-fraud agents. Hospitals need charting, coding, prior authorization, and patient-triage agents. Manufacturers need quality control, procurement, scheduling, and maintenance agents. Retailers need customer service, inventory, and replenishment agents. Finance teams need invoice, reconciliation, close, and collections agents.
Each of those categories may look smaller and less concentrated than coding in isolation. Together, though, they map to service industries, knowledge work, back-office operations, and professional services. PANews says the associated labor, outsourcing, and transaction budgets are far larger than the software development budget. That is why the article argues that AI’s second growth engine is unlikely to appear as a single product suddenly becoming “the second Coding.” A more likely outcome is dozens of industries and hundreds of business processes entering agent-based execution at the same time, with the combined total eventually surpassing coding. In that framing, the next stage of the AI revolution is a distributed workflow shift rather than one giant app.
Which use cases are closest to the next coding wave
The article proposes a simple test for whether a use case can be transformed quickly: AI business value is approximately task frequency multiplied by labor cost, multiplied by automatable share, multiplied by verifiability, multiplied by system execution capability, minus the costs of errors, integration, and governance. Under that framework, the best candidates are not strategy, creativity, or executive management. They are highly digitized processes with clear SOPs, measurable outputs, human backstops, and systems that AI can actually operate.
Customer service and voice agents
PANews describes customer service as the single scenario that looks most like a second coding market. It already has knowledge bases, service workflows, identity checks, ticketing systems, and clear success metrics. The value of a customer-service agent can be measured through first-contact resolution, handoff rates to humans, average handling time, repeat call rates, and customer satisfaction. As voice models improve, AI is moving beyond chat windows into phone booking, refunds, itinerary changes, collections, sales qualification, hotel service, healthcare front desks, and insurance inquiries.
But voice itself is not the business model, the article says. The real value is not whether AI sounds human. It is whether it can look up an order, verify identity, change a system, trigger a refund, update CRM records, and close a ticket. Only then does it become a business agent rather than a talking assistant.
IT operations and cybersecurity
IT operations and security could become a second major line inside technical departments. Like coding, they run on machine-readable data: logs, alerts, config files, network topology, vulnerability information, permission systems, and runbooks. PANews says AI can move from explaining alerts to locating faults, calling diagnostic tools, executing runbooks, verifying recovery, and generating incident reports, while sending high-risk actions to humans for approval. The largest bottleneck is not model quality, according to the article. It is whether enterprises are willing to give agents sufficient system permissions.
Finance, accounting, and procurement
The article calls this one of the most underestimated large markets. Financial processes are document-heavy, rule-heavy, repetitive, and naturally verifiable. Whether amounts match, debits and credits balance, approval rights are correct, three-way matches hold, and monthly close is complete can all be checked. That makes invoice processing, accounts payable, accounts receivable, expense review, bank reconciliation, sourcing, month-end close, cash-flow forecasting, and audit workpapers suitable for agent-based execution. These products may never become famous consumer apps, PANews says, but they could absorb a great deal of shared-services, finance-outsourcing, and back-office work.
Healthcare administration, legal, insurance, and compliance
These areas have high labor costs, dense documentation, and relatively standardized workflows. The article says the most mature healthcare AI applications today are often not direct diagnosis, but medical documentation, coding and billing, prior authorization, scheduling, patient triage, and claims material handling. Legal AI has entered search, contract review, due diligence, litigation material organization, and regulatory reporting. Insurance is a natural fit for document reading, underwriting, claims, anti-fraud, and collections. Their shared constraint is high error cost, which means professional review and final signoff will remain in place for a long time.
Supply chain and manufacturing
PANews says supply chain and manufacturing may offer enormous long-term value, but the pace of adoption will be slower. Demand forecasting, inventory replenishment, procurement, production scheduling, quality control, equipment maintenance, and logistics dispatch can create economic value that far exceeds general office assistance. Yet these workflows span multiple departments and legacy systems, and some outcomes happen in the physical world, with feedback cycles that may take weeks or months. As a result, the article argues that supply chain and manufacturing could become one of the largest long-term AI markets without productizing and scaling as quickly as coding did.
If a single answer is required, PANews says customer service and voice agents are closest to “the next Coding.” But the use cases that can exceed coding in aggregate will not come from customer service alone. The combined total from customer service, finance, IT, healthcare, legal, insurance, procurement, and supply-chain workflows is more likely to matter.
Microsoft’s numbers show three stages of non-coding AI adoption
The article points to three figures disclosed by Microsoft in FY2026 fourth-quarter reporting: 100,000 Foundry customers with revenue up by more than 100% year over year, more than 30 million paid Microsoft 365 Copilot seats, and nearly 40 million registered agents after the launch of Agent 365. All three support the idea that AI is spreading beyond coding, PANews says, but they do so with very different evidentiary strength. The numbers also cannot simply be added together, because a single large enterprise may appear in all three datasets at once.
Foundry shows enterprises are building
Foundry is not an end-user app. It is a platform for building, deploying, and running models, applications, and agents. According to the article, its users are often enterprise IT teams, data teams, AI platform groups, software companies, system integrators, and AI-native businesses. On Foundry, companies can choose models, make inference calls, connect RAG systems to enterprise data, orchestrate agents, call tools, evaluate outputs, monitor behavior, and manage safety and governance.
That means 100,000 customers and revenue that more than doubled show that enterprises are moving from general-purpose chatbots to constructing industry applications and business agents of their own. At the same time, Foundry supports both coding and non-coding scenarios. Microsoft did not disclose how many customers are still experimenting, how many have reached production, or what the absolute revenue base and use-case mix look like. For that reason, PANews treats Foundry as upstream and midstream evidence that the non-coding build pipeline is expanding, not as sufficient proof that end-market non-coding revenue has fully broken out.
Microsoft 365 Copilot shows non-coding workers are using AI
The 30 million figure refers to paid employee seats, not 30 million companies. Those users include developers, but the article says they more importantly include finance staff, sales teams, HR, legal professionals, doctors, consultants, bank employees, manufacturing managers, and general administrators. They summarize emails in Outlook, produce meeting notes in Teams, draft documents in Word, build presentations in PowerPoint, analyze data in Excel, and search enterprise knowledge bases.
In PANews’ reading, this is the most direct evidence that AI is moving from coding into non-coding work. Enterprises are no longer buying AI only for programmers. They are buying it for large populations of knowledge workers. Even so, paid adoption is not the same thing as process redesign. A worker using Copilot to summarize meetings, rewrite emails, or look up information shows AI has entered daily work, but it does not prove that a company has reduced headcount, lowered SG&A, or increased free cash flow.
Agent 365 shows enterprises are starting to govern agents
Agent 365 is framed in the article as an identity, permissions, safety, and audit control plane for enterprise agents. It discovers and registers agents, manages which data and tools they can access, tracks which actions they can take, and records lifecycle and compliance information. Nearly 40 million registered agents does not mean 40 million agents are working every day, the article stresses. Nor does it mean 40 million paid units, 40 million token-consuming agents, 40 million production agents, or 40 million revenue-generating entities.
Instead, the figure is closer to an enterprise inventory of agent assets. It may include test agents, personal agents, department-level agents, deactivated agents, auto-synced agents, and system-detected “shadow agents.” What it really signals is that internal agent counts and sources have become complicated enough to require centralized identity, permissions, safety, and audit governance. That makes it an important leading indicator, but not a metric of revenue, activity, or productivity.
The article’s bottom line is straightforward: Foundry shows enterprises are building, Microsoft 365 Copilot shows workers are using, and Agent 365 shows companies are starting to govern. Among the three, M365 Copilot is the most direct evidence of non-coding diffusion, Foundry is an upstream signal with revenue behind it, and Agent 365 is a leading indicator for architecture and governance.
How far enterprise AI has actually progressed
PANews splits enterprise AI diffusion into five stages. Stage one is the coding agent stage, where AI enters software engineering to handle completion, testing, debugging, refactoring, and deployment. The article says this phase has already moved from experimentation to scaled production and remains the most mature enterprise AI workload.
Stage two is the horizontal knowledge-work copilot: email, meetings, documents, spreadsheets, research, enterprise search, and content generation. Microsoft 365 Copilot’s 30 million paid seats are presented as evidence that this phase has reached scaled procurement.
Stage three is the department-level agent stage. AI begins entering customer service, sales, finance, procurement, HR, legal, IT operations, and security. The change here is that AI is no longer only generating content; it is beginning to call enterprise data and business tools. PANews says Foundry growth, Dynamics consumption growth, and a large body of industry cases show this stage is taking shape.
Stage four is the end-to-end business-process stage. An agent can query data, call systems, execute approvals, trigger refunds, generate orders, update records, and handle exceptions. The article gives the example of a refund agent that does not merely tell a support worker that the customer qualifies for a refund, but handles identity verification, order lookup, policy checking, payment refund, CRM updates, and customer notification. PANews says only some leading companies and some standardized processes have entered this stage so far.
Stage five is where enterprise profit and macro productivity show up. Revenue growth starts to run persistently ahead of employee growth, SG&A rates decline, unit costs fall, collection cycles shorten, and cash conversion improves. Only then do the changes start to feed total factor productivity, GDP, and other macro data.
The article’s assessment is that enterprise AI is already broadly in stage two, moving toward stage three, with a small number of advanced customers entering stage four. Stage five still lacks broad, continuous, auditable evidence. Microsoft case studies cited in the piece point in a direction: contract attachment handling compressed from 30 to 45 minutes to about 3 minutes, an HR certificate process cut from three days to 120 seconds, pilot sales teams raising revenue per person and shortening deal cycles, and rapid growth in usage-based customer-service consumption. PANews stresses that these are selected cases, not average results across the full corporate universe. The key distinction remains unchanged: using AI inside a business process is not the same as AI completing the process, and time saved by employees is not the same as higher corporate profit.
Non-coding diffusion is real, but still broad rather than deep
Looking at model-provider usage data, PANews says AI is clearly spreading from coding into other functions. Anthropic data show computer and mathematics-related tasks still account for roughly 35% of Claude.ai conversations, keeping coding as the largest single task category. But the share of the ten most common tasks fell from 24% in November 2025 to 19% in February 2026, which the article interprets as evidence of a broadening use-case mix.
OpenAI’s enterprise data point in a similar direction. Writing remains the most common use for ChatGPT. In agent-style usage, coding and system operations together account for nearly 75%. At the same time, about one-quarter to one-third of agent messages in recruiting, sales, policy, and enterprise communications already involve system or agent operations. Since February 2026, weekly active Codex users in legal departments among OpenAI enterprise customers have grown 108x, while sales and recruiting have each grown 41x, marketing 26x, and engineering about 5x. PANews notes that non-engineering functions started from much smaller bases, so those multiples are not directly comparable, but the direction is unmistakable: agent-style usage is moving beyond technical teams.
The U.S. Census Bureau data cited in the article offer a more cautious counterpoint. From late 2025 to early 2026, about 18% of U.S. businesses used AI in at least one business function, rising to about 32% when weighted by employment. Among AI adopters, 57% had deployed AI in three or fewer functions, and 65% used it for three or fewer tasks. Another 66% still used AI only to augment employees, while only about 2% reported reducing employment because of AI.
Put together, PANews says, those numbers lead to a precise conclusion: AI has entered the enterprise, but it has not broadly entered the enterprise cost structure. It has moved from trial to procurement, and from technical teams to sales, finance, healthcare, and administration, but most companies are still at the personal assistant and partial copilot stage. End-to-end workflows that actually reshape the income statement are still concentrated among a subset of leading firms. That helps explain why the non-coding market can be much larger than coding in total while still producing less standardized products and a slower expansion curve for any one company.
Consumer opportunity and the longer-term case for scientific discovery
On the consumer side, PANews argues that the likeliest platform-scale opportunity is not simple chat, but the personal action agent: a system that can actually complete tasks for a user. It may read emails, identify to-dos, arrange trips, compare flights and hotels, book doctors, manage bills, handle returns, buy goods, plan travel, fill out forms, and coordinate family schedules. Most current AI products still operate at the level of telling people what to do, the article says. The next step is an AI that knows a user’s preferences, has access to authorized accounts and tools, and can finish the work on the user’s behalf.
That direction could support business models richer than standard subscriptions, including transaction commissions, payment revenue, advertising, referral fees, financial services revenue, and membership fees. But the trust threshold is much higher than in coding. The pace of expansion will depend on whether users are willing to let AI read email, access bank accounts, use credit cards, handle medical information, or execute transactions in their name. For that reason, PANews says the personal action agent may be the biggest long-term consumer opportunity without being the easiest category to monetize quickly at scale.
The article then turns to what it sees as an even more important long-term frontier: AI scientific discovery. Coding is about producing software more efficiently. Enterprise agents are about completing existing business tasks more efficiently. Scientific discovery, in this framing, is about producing new knowledge, new technology, and new possibilities more efficiently. That is not a normal application category. It embeds AI in the production function of technological progress itself.
PANews points to a recent case: the personalized mRNA cancer vaccine intismeran autogene, previously known as mRNA-4157 or V940, jointly developed by Moderna and Merck, delivered positive results in a Phase III adjuvant melanoma trial. The article says AI played a major role in neoantigen selection and manufacturing scheduling. Moderna, in its official blog, called the outcome “A Testament to the Power of AI,” saying AI enabled speed and precision in the complexity of one-person-one-drug treatment design. PANews writes that this supports the viability of AI/ML-driven personalized neoantigen prediction combined with an mRNA platform, producing meaningful clinical benefit in a large randomized Phase III trial for high-risk solid tumors and surpassing Keytruda monotherapy as the existing standard treatment. In the article’s view, that validates the “computational biology plus precision immunotherapy” route, especially in tumors with high mutational burden.
The broader point is that AI scientific discovery uses intelligence to generate new knowledge, and that knowledge can then become new drugs, materials, energy systems, chips, robots, quantum devices, and industrial methods. PANews lays out a long chain: AI generates hypotheses, simulation and computation screen them, automated labs validate them, new drugs and materials emerge, new devices and robots appear, and those industries generate still more data, compute demand, and research needs. The article says this could move AI from a tool that improves current economic efficiency to a technology that raises the speed of technological progress across society. It also makes a timing distinction: in civilizational and long-term economic terms, scientific discovery may matter more than the next consumer app or even workflow agents, but in the next one to two years it may not monetize faster than customer-service agents, enterprise copilots, or coding.
Four different clocks are shaping the AI timing gap
PANews says the AI application timing gap comes from four clocks running at very different speeds.
The first is the capital-expenditure clock. GPUs, data centers, power, networks, land, and cooling have to be built in advance. If cloud providers wait until demand is fully proven, they may miss the capacity window and lose customers. The article says this phase can be understood as immediate to 18 months, though actual project timelines vary.
The second is the product-revenue clock. Cloud compute, model APIs, coding agents, and general copilots can usually start charging within a few quarters after capital is committed. PANews places this phase at roughly 6 to 24 months. The article references a pattern it has discussed before: cash is paid first, capacity goes online later, orders turn into revenue after that, and profit and free cash flow arrive last.
The third is the enterprise workflow profit clock. To get AI onto the income statement, companies have to complete data governance, permissions opening, system integration, process redesign, training, and organizational adjustment. The time saved also has to become slower hiring, less outsourcing, higher throughput, shorter delivery cycles, better conversion, or lower error losses. PANews says this process may take 18 to 48 months, and longer in complex industries.
The fourth is the macro productivity clock. Only when AI reaches enough companies, workflows, and industries, and when organizational and capital structures are redesigned around it, can the benefits show up clearly in total factor productivity and GDP data. The article cites year-over-year growth of 2.2% in second-quarter labor productivity for the U.S. nonfarm business sector and 1.4% growth in unit labor costs. It calls that a positive sign, but says labor productivity is not the same as total factor productivity, and current gains cannot all be credited to AI. This process is measured in years, perhaps 3 to 7 years or longer.
These are not precise forecasts, PANews says, but a framework for understanding industry rhythm. The key point is that the patience of capital markets, bond markets, and corporate management is usually shorter than the time needed for organizational restructuring and macro productivity realization. That is why the AI capital cycle is especially prone to volatility.
The real risk is duration mismatch between capex and end-market cash flow
The article says today’s AI demand is not fake. Cloud revenue, model usage, coding activity, and copilot procurement are all rising. But it is possible for demand growth and capital consumption to be true at the same time. Using the FY2026 framing cited in the article, Google Cloud is growing rapidly, but quarterly capex has already exceeded operating cash flow and pushed free cash flow negative. Amazon Web Services is still posting strong growth and generating real AI revenue, yet AI equipment spending has also materially compressed free cash flow over the last twelve months. Microsoft’s operating cash flow and order quality are relatively stronger, but quarterly capex still sits far above free cash flow. Meta continues to subsidize AI spending with advertising revenue, while its free-cash-flow cushion has also narrowed substantially.
PANews notes that the reporting periods and accounting scopes are not identical across these companies, so the numbers cannot be compared mechanically. The directional message is still consistent: AI revenue is growing, but AI capital intensity is growing faster. That is why, in the article’s view, the market no longer needs proof that AI has revenue. It needs proof of when incremental AI gross profit, growth in existing business lines, and cost savings will begin to outpace depreciation, energy costs, interest expense, lease obligations, and replacement costs for equipment.
The piece then describes the early AI funding loop as follows: model companies raise money, buy cloud compute and GPUs, cloud providers grow revenue, cloud providers keep building data centers, chip, optical communications, power, and data center companies expand capacity, and the whole supply chain raises more money. That loop can persist for a long time. Still, a meaningful share of the revenue inside it comes from spending enabled by capital markets rather than from operating cash flow created by traditional enterprises and consumers.
PANews breaks AI demand quality into three layers. The first layer is expansion financed by equity, debt, leasing, and project finance among model companies, AI cloud providers, and data center operators. The second layer is end demand from the tech sector itself: coding, model APIs, ad optimization, search, and recommendation systems. The third layer is non-tech industries such as banking, insurance, healthcare, manufacturing, retail, logistics, finance, and customer service using their own operating cash flow to buy AI and earning returns from real business processes.
Only when the third layer scales, the article argues, does the AI capital cycle move from capital-market-driven supply expansion to end demand generated by production activity across the broader economy. That is why non-coding use cases matter so much. They do not only add revenue. They shift the funding source for AI demand away from model-company financing and tech-company capex toward operating budgets across the wider economy.
No clear “second Coding” does not mean AI is broken
PANews argues that AI, as a general-purpose technology, does not need a single application to prove its commercial case. Personal computing did not change the world through one piece of software alone. The internet did not depend only on search. Cloud computing was not built around one workload. AI’s end state may not be another isolated super app, but a gradual emergence as the execution layer underneath software, data, workflows, and transactions everywhere.
The article highlights three underappreciated shifts. First, the budget pools linked to non-coding use cases are larger. Coding mainly touches software-engineering and IT budgets. Once non-coding agents start completing real work, they can tap enterprise software budgets, labor budgets, BPO and outsourcing budgets, consulting and professional services budgets, sales and marketing budgets, claims and fraud-loss budgets, and transaction and commission budgets. That could move pricing from seat-based subscriptions toward usage-based, task-based, and outcome-based models. A seat fee pays for a tool assigned to a worker; an outcome fee suggests the AI actually completed the work. Those are not the same markets.
Second, the article says vibe coding is becoming infrastructure for non-coding workflows. Many future non-technical employees may not think of themselves as coding, but they could still use natural language to generate reconciliation tools, sales dashboards, inventory warning systems, internal approval pages, customer data-cleaning scripts, automated reports, contract-checking flows, and department-specific agents. In the past, these long-tail workflows were too small or too specialized to justify a dedicated SaaS purchase or a custom IT project, so companies left them inside Excel, email, and manual operations. Once coding agents cut the cost of custom software, PANews says, many non-coding workflows become economically viable candidates for software automation and agent execution for the first time. In that sense, coding may end up hidden beneath non-coding business processes rather than standing apart from them.
Third, the article returns to the consumer answer: the personal action agent. Here, the platform battle will center on trust, account permissions, and transaction access. The winner is likely to be the system that can actually finish the task, not just suggest the next step. Still, from the perspective of supporting today’s AI infrastructure buildout, enterprise workflow agents remain more important than any single consumer application because enterprise demand is steadier, budgets are larger, and inference consumption is easier to sustain.
Three possible paths and the metrics that matter next
PANews sketches three paths from here. The first is a fast handoff. Customer service, finance, healthcare administration, IT operations, insurance claims, and legal compliance move rapidly into production over the next one to two years. Non-coding agent consumption accelerates, pricing shifts from seat-based to usage- and outcome-based models, AI demand spreads deeper into traditional industries, enterprise revenue growth starts running ahead of employee and cost growth, cloud-provider free cash flow bottoms out, and AI capital returns are validated.
The second is the base case. The direction is right, but the middle is uneven. Non-coding demand continues to grow, yet enterprise deployment, systems integration, and organizational redesign take two to four years. Coding keeps growing but slows over time. Non-coding grows rapidly from a smaller base but does not fully take over at once. Meanwhile, capex, depreciation, and interest keep rising. The market repeatedly questions AI returns, leading to valuation pressure on semiconductors, data centers, and other capital-intensive names, before demand, workflow, and productivity data improve again. PANews says this may be the most realistic path and also says the author personally leans somewhere between the base case and a faster handoff.
The third path is a slow handoff, or even capital clearing. If model improvements slow, non-coding tasks keep requiring heavy human oversight, enterprises remain unwilling to connect agents to core systems, and model price competition cuts unit revenue, then capital returns worsen. Even if AI retains large long-term value, one wave of spending may still fail to earn a reasonable return.
To judge whether non-coding demand is truly taking over, the article says the market should not rely on registered agent counts, model size, token growth, or capex alone. It points instead to active agents rather than registered agents; whether agents actually execute system actions; autonomous completion rates and human handoff rates; the shift from seat revenue to consumption and outcome revenue; the absolute scale and customer mix of Foundry revenue; whether enterprises are actually harvesting productivity; and whether incremental AI profit at cloud providers can cover the capital cost base. The real unit test, in PANews’ wording, is whether incremental AI gross profit, growth in existing businesses, and customer cost savings can cover inference costs, depreciation, energy, interest, and equipment replacement.
The article also lists three crossover points that matter. The first is when non-coding growth overtakes coding growth as coding inevitably matures. The second is when AI revenue overtakes AI-related depreciation, energy, and interest costs. The third is when enterprise time savings finally cross into enterprise profit, meaning revenue grows faster than headcount and expense.
The market is not looking for a second app. It is looking for a second demand engine
PANews closes by saying the market is not truly searching for a second application. It is searching for a second demand engine. Customer service and voice agents are the closest single candidate. The larger answer is a stack of enterprise workflow agents across finance, healthcare, legal, insurance, sales, procurement, IT operations, and supply chains. On the consumer side, the strongest platform-level possibility is the personal action agent with long memory, account permissions, and transaction power. Scientific discovery is the more important long-term frontier beyond both.
The article’s core conclusion is that AI’s second growth engine probably will not arrive in the neat form of a “second Coding.” It is more likely to arrive through hundreds of business processes shifting from manual work to AI execution at the same time. The absence of one new super app does not mean AI is failing. AI can still create large economic value through broad, gradual productivity gains.
For capital markets, though, it is no longer enough to show that many people are using AI. The next proof points must be tougher: that AI can complete real work, that enterprises will pay on a usage or outcome basis, that non-tech industries can generate durable end demand, that productivity improvements can reach margins and free cash flow, and that incremental gross profit can eventually cover depreciation, energy, interest, and equipment renewal.
That is the central contradiction in the current AI investment case, PANews says. The future may be large, but capex and debt are already here. Coding has moved from expectation gap to consensus, while non-coding revenue and broad productivity still sit on the way, not fully in hand. Coding has already run the first leg of the relay. Customer service, healthcare, finance, sales, supply chain, and personal agents are trying to take the baton for the second. The next driver will be whether those scattered use cases can add up to enough revenue, cash flow, and productivity before the first leg loses visible momentum. In the article’s view, that timing gap is the most important issue in the AI application layer today.
The piece ends by saying the market is likely to become more selective about who has real end demand and who only has internal supply-chain orders, who can produce free cash flow and who can only produce revenue, and who can charge against labor and transaction budgets rather than compete mainly through low-priced token supply.