2026: The New Race to Build Smarter Businesses

Every business has been told, repeatedly, that it is in a race to become intelligent. The message arrives from software vendors, consultants, conference speakers and the occasional board member who has just read an article. After spending the first five months of 2026 reading the evidence behind the message, I think the race is real but is mostly misdescribed. It is not one race but three running at once, on different tracks, at different speeds, and with different winners. The first is a race in the cost and capability of the intelligence itself, which is running faster than almost anyone predicted. The second is a race to integrate that intelligence into how a company actually operates, which is running far slower than the vendors say. The third is a race for capital and computing capacity, which is the most visible and the least understood. A company can be winning one track and losing another, and most of the confusion in the public conversation comes from treating the three as one. This study takes them in turn, then looks at what has changed in the software businesses use, how competition among software companies is being rearranged, and what a small or medium sized business should do. I have drawn on the Stanford AI Index, McKinsey's global survey, Menlo Ventures' enterprise report, company filings and named surveys available up to the end of May 2026, and I have flagged where figures come from vendors or aggregators. Where I offer opinions or predictions, I say so.

Begin with the first track, the cost and capability of intelligence, because it sets the pace for everything else. The Stanford AI Index for 2026, published in April, documents a year of rapid improvement. On a benchmark of real software engineering tasks, the best systems moved from roughly sixty percent of the human baseline to close to it within a single year. On tests of whether an agent can complete real tasks on a computer, success rates rose from about twelve percent to roughly two thirds, which means that agents now complete a majority of such tasks and still fail about one attempt in three. The gap between American and Chinese models has, by the Index's account, effectively closed on standard measures, and industry produced more than ninety percent of the notable frontier models. Corporate investment in AI more than doubled in 2025, to about 582 billion dollars by one estimate cited in coverage of the Index, and generative AI reached roughly 53 percent adoption among the population within three years of its mass market launch.

Costs moved just as fast. By the end of May, published rate cards for a leading Chinese open weight model listed prices of about fourteen cents per million input tokens and twenty eight cents per million output tokens for its fast tier, and a promotional discount of seventy five percent on its higher tier ran through the end of May. A competing Chinese lab offered a strong coding model at a few tens of cents per million tokens. Analysts tracking inference prices have described declines of roughly an order of magnitude per year at constant capability, and the price of output for a general frontier capability has fallen from about sixty dollars per million tokens in 2023 to a few dollars in 2026. Some of these figures come from consultancy and aggregator summaries, so I would check them against the providers' own pricing pages, but there is no serious dispute about direction. Whatever intelligence costs a business today, it will cost much less in a year, and what a company can do with a given budget is growing quickly.

What does this mean for a business that is not a technology company? The first implication is that capability is not the scarce thing. In nearly every domain, a business with a modest budget can now buy access to systems that would have required a research laboratory three years ago. The second is that competitive advantage cannot rest on access to the tools, since competitors have the same access. The third is that the economics of anything built on top of them are in flux, because prices are falling while demand and ambition rise. I will return to the strategic consequences below. For now, the point is that on the first track nobody is winning in the sense of owning an advantage, because the advantage is being competed away as fast as it appears.

Turn now to the second track, integration, where the evidence is much less flattering to the enthusiasts. The most useful statistics come from McKinsey's global survey, published last November. It found that 88 percent of respondents say their organisations regularly use AI in at least one function, up from 78 percent a year earlier, and that 72 percent report using generative AI. But only about a third say their organisations have begun to scale AI across the enterprise, only 39 percent report any impact on earnings before interest and tax, and most of those say the impact is under five percent. Around six percent qualify as high performers, meaning they attribute more than five percent of earnings to AI and report significant value. On agents, 62 percent say their organisations are at least experimenting and 23 percent are scaling agents somewhere, but in any single function no more than ten percent are scaling them. The Stanford Index reports the same picture for 2025, with agent deployment in the single digits across nearly every business function.

Surveys by vendors tell a consistent story in different words. A survey published by the software company WRITER in the spring found that almost every executive reported deploying AI agents in the past year, yet only about three in ten reported significant returns, and three quarters of executives admitted that their AI strategy was more for show than for guidance. A Capgemini study of financial institutions, published in November, found that although banks and insurers named customer service, fraud detection, loan processing and onboarding as top areas for agents, only about one in ten had implemented agents at scale. These surveys have obvious limitations, since they are self reported and sometimes sponsored, but their agreement with the independent academic and consultancy data gives me some confidence in the picture. Widespread use, narrow depth, and results that depend on something other than the technology.

That something, according to McKinsey, is the redesign of work. High performers are nearly three times as likely to have fundamentally redesigned workflows when they deployed AI, 55 percent against 20 percent, and they are far more likely to have defined human review processes, 65 percent against 23 percent. They also invest more, with many spending a fifth or more of their digital budgets on AI, they have senior leaders who own the effort, and they use AI for growth and innovation, not just for cutting cost. Microsoft's 2026 Work Trend Index, based on a survey of twenty thousand workers who use AI across ten countries and on telemetry from its own products, puts the same finding in the language of organisations: roughly one in five workers is in a frontier zone where personal capability and organisational support reinforce each other, about one in ten is blocked by an organisation that has not caught up, and about half sit in between. It describes a paradox in which the majority fear falling behind without AI, while few feel rewarded for reinventing their work. I treat the Microsoft numbers as a vendor describing its own users, but they agree with the independent findings about where the bottleneck lies.

The message I draw from this second track is that integration is a management problem and not a technology problem. Companies that treat AI as a tool to be installed get modest results. Companies that treat it as a reason to rethink how a process works, who does what, how quality is checked and how success is measured get much larger ones. This is slow, because it requires cooperation among people, changes in incentives and a willingness to disturb routines that work well enough. It is also why the race on this track is so uneven. A small firm with a decisive owner can redesign a process in a month. A large firm with layers of approval may take years. And it is why the technology giants, who are good at changing their own processes, appear to be gaining on companies that are slower to adapt. A sign of this is the revenue per employee of the newest software companies. Reporting in March, drawing on research from the venture firm Redpoint, put the typical AI native start up at two to four million dollars of revenue per employee against about three hundred thousand for the average public software company, and cited a company that had reached 400 million dollars of annual recurring revenue with 146 full time employees. Such figures come from the companies that succeeded and ignore the many that did not, so I would not generalise them, but they illustrate what a business built around the tools from the start can look like.

Now consider the money flowing into enterprise AI, since it shows where companies are actually spending. Menlo Ventures, in its survey of 495 American enterprise decision makers conducted in November, estimated that enterprise spending on generative AI reached about 37 billion dollars in 2025, up from about 11.5 billion in 2024 and 1.7 billion in 2023, which would make it the fastest growing software category on record and about six percent of the global software market. Slightly more than half of that, about 19 billion dollars, went to applications and the rest to infrastructure and model access. Within applications, coding tools were the standout, at about four billion dollars, followed by smaller amounts for IT operations, marketing and customer success, while the horizontal category was dominated by general purpose assistants and copilots, with agent platforms accounting for a small share. Menlo also estimated that one model provider had captured around forty percent of enterprise spending on large language model access, up sharply from two years earlier, and that three providers together accounted for the great majority. These are survey based estimates by an investor with a stake in the sector, so I would use them for orders of magnitude and not for precision.

The lesson from the money is consistent with the lesson from the surveys. Enterprises are paying for things that deliver measurable value in structured work, with coding the clearest case, and are spending much less on the grander agent platforms that vendors promote most loudly. That is a sensible pattern. It suggests that buyers are more disciplined than the hype implies, concentrating on use cases where a result can be checked, and that the companies that will earn the most in the next two years are those that solve concrete, checkable problems in particular departments and industries.

This brings me to the third track, the race for capital and computing capacity, which dominates the headlines. The five largest technology spenders are planning capital expenditure for 2026 that analysts put at roughly 660 to 690 billion dollars, nearly double the previous year, with most of it directed at AI infrastructure. In April, Alphabet reported that revenue at its cloud division grew 63 percent to 20 billion dollars in the quarter and that its backlog of contracted future revenue nearly doubled from the previous quarter to more than 460 billion dollars, and said that paid users of its enterprise AI product grew 40 percent in a quarter. Microsoft has reported that its cloud backlog stood at 392 billion dollars at the end of last September and that Azure was growing around forty percent, and trade reports said it had more than twenty million paid seats for its AI assistant by April, up from fifteen million in January. In late May, Nvidia reported quarterly revenue of 81.6 billion dollars, up 85 percent from a year earlier, with data centre revenue of 75.2 billion dollars and a gross margin around 75 percent.

It is tempting to treat these numbers as proof that the intelligent business era has arrived. I would say more carefully that they prove that the largest technology companies and their suppliers are earning large amounts from the build out, and that customers are committing large sums to future capacity. They do not tell us how much of that spending will be matched by value delivered to the businesses that use the computing, which is the question the second track addresses. The two tracks are connected in a way that matters to any company thinking about its own position: the build out is being financed on the assumption that businesses will find enough valuable uses for the intelligence it produces, and the survey evidence suggests that most businesses are still working out what those uses are. A business that moves faster than its peers on integration is therefore not merely keeping up with a trend. It is helping to determine whether the investment turns out to have been wise.

With the three tracks described, the next step is to look at what has changed in the software that businesses use.

The most important change in business software over the past year is the movement from tools that help a person do a task to systems that carry out the task. The older generation of AI features answered a question or drafted a paragraph and then waited. The newer generation, usually called agents, takes a goal, breaks it into steps, uses other software to carry them out and reports back. The Stanford Index's finding that agents complete roughly two thirds of real computer tasks in benchmark conditions is the capability side of this shift. The adoption side is more modest, as I have described, with deployment in the single digits across most business functions. Gartner, in a forecast published last August, predicted that about forty percent of enterprise applications would include task specific agents by the end of 2026, up from less than five percent in 2025. Microsoft's Work Trend Index reports that the number of active agents on its platform grew roughly fifteen times in a year, though it does not give the base. Taken together, these data suggest that agents are moving from experiment to ordinary feature quickly in the products people already use, even if few companies have deliberately built agent based processes of their own.

A second change is the rise of vertical software, which wraps a general model in the knowledge, permissions and workflow of one profession or industry. The funding data of the past year show where investors see value. The legal AI company Harvey raised 200 million dollars in March at an 11 billion dollar valuation and said its customers were running more than 25,000 custom agents on its platform. Customer service agent companies have raised very large rounds, with Decagon raising 250 million dollars in January at a 4.5 billion dollar valuation and Sierra, founded in 2023, reported at about 150 million dollars of annual recurring revenue in January and announcing a raise of close to a billion dollars in May. Glean, which connects an enterprise's applications so that agents can find the right information while respecting who is allowed to see it, announced on 28 May that its annual recurring revenue had passed 300 million dollars, fifteen months after crossing 100 million. Euclid Ventures' annual vertical report found large rounds in healthcare and legal software, and it noted that legacy software firms in each profession are becoming AI powered through acquisition. Each of these figures comes from the companies or from investors, so I would treat them as indicative. But they point to where the buying is happening: in specific workflows where the value can be measured.

The third change is in how software is priced, and it is where the competitive rearrangement is easiest to see. For two decades the dominant model charged by the user. As agents do work that people used to do, the number of users a customer needs falls even as the value delivered rises, and the pricing model starts to work against the vendor. The early months of 2026 saw a sharp sell off in the shares of established software companies as investors reassessed how much of the sector's revenue depended on the headcount of its customers. A pricing study reported by several trade sources found the share of software companies charging purely per seat falling from 21 percent to 15 percent in a year while hybrid models, which combine a platform fee with usage or outcome charges, rose from 27 to 41 percent. Atlassian was reported to have posted its first ever decline in enterprise seat counts. HubSpot cut the price of its customer agent to fifty cents per resolved conversation in April, and others charge between roughly one and two dollars per resolution. Gartner is quoted as expecting at least forty percent of enterprise software spending to move toward usage, agent or outcome based pricing by 2030, and IDC as expecting seventy percent of vendors to move away from pure per seat pricing by 2028. As with all forecasts, treat these as informed guesses, but the direction matches what I observe.

The fourth change concerns who stands between a business and its customers. The first five months of 2026 saw the arrival of open protocols that let software agents browse, compare and buy on a consumer's behalf. OpenAI and Stripe's protocol powered an instant checkout inside ChatGPT that began in September 2025 and was reported to have run until March, and in January Google and Shopify, with a group of large retailers, announced a Universal Commerce Protocol, with Google beginning to roll out agent checkout in its search and Gemini products. At the same time, Stripe, which processed 1.9 trillion dollars of payments in 2025, launched its own payments blockchain, Tempo, in March, after acquiring the stablecoin company Bridge, and Mastercard announced the acquisition of the stablecoin infrastructure firm BVNK for 1.8 billion dollars. I regard these as moves in a contest over who controls the point where an intention to buy turns into a payment. The businesses that sit at that point earn a fee on everything that passes through it, and the businesses on either side lose some bargaining power.

What does the competitive picture look like when we put the pieces together? The AI native newcomers have the advantage of building around the tools from the start, with small teams and no legacy to protect. The incumbents have customers, data, distribution and licences, and they are using them, either by building AI into their products, by buying newcomers or by partnering. The model providers are moving up the stack into specific industries, with healthcare offerings announced by two of them in January according to a vendor summary, which threatens vertical companies built on top of them. The infrastructure companies earn handsome returns from all of them. And the customers enjoy falling prices and rising capability, with the choice among suppliers growing larger. In such a structure, advantage tends to belong to whoever holds something that is hard to copy: proprietary data, a regulatory permission, a relationship of trust, a distribution channel or a deeply embedded workflow. The model itself is the one ingredient that is not on that list.

The fifth change is inside the organisation. If an agent can do the execution and a person must direct and review it, the shape of a team changes. Teams get smaller for the same output, which is the logic behind the revenue per employee figures. The roles within them shift toward specifying work, evaluating results and handling exceptions, and the pace of work quickens because the cycle from idea to result shrinks. I discussed in the evidence on integration that the redesign of workflows, not the purchase of tools, separates high performers from the rest. It is worth stressing the human side of this, since the Work Trend Index finds that workers are more ready than their organisations, that the biggest enabler is a manager who makes experimentation safe, and that few employees feel rewarded for reinventing their work. A company that wants to win the integration race should therefore ask whether its incentives, its managers and its review processes support change, because buying more software will not fix a culture that punishes experiments.

Any honest account of the race must include trust and risk, which the enthusiastic accounts leave out. The Stanford Index reports that while nearly sixty percent of respondents believe AI's benefits outweigh its drawbacks, up from 55 percent, about half also express concern about AI powered products and services, and it observes that capability is advancing faster than the safety measures around it. The security picture has worsened in a specific way. The Cloud Security Alliance reported in March that studies find AI generated code introducing security flaws in a large share of tasks, that vulnerabilities formally attributed to AI generated code rose from six in January to thirty five in March by one tracker, and that the true numbers are likely several times higher. In February, researchers found that a social network for AI agents built entirely by generating code had left its database open, exposing about 1.5 million authentication tokens. A business that races to deploy without review is racing toward an incident. The firms that win will be those that pair speed with controls: limits on what agents may do, logs of what they did, review by people who understand the domain and the discipline to stop a release that is not ready.

The race is also being run in a harsher economic environment than the one many strategies assumed. The war in the Middle East that began at the end of February has raised energy and shipping costs and unsettled inflation, and the International Monetary Fund's April projection of global growth of 3.1 percent for 2026 sits below the pace of recent years. For a business planning an AI programme, this matters in two ways. It raises the bar for returns, since a company under cost pressure will scrutinise any project that does not pay back quickly, and it favours the use cases that cut cost or recover revenue visibly in the first months over the grand transformations that promise benefits years away. It also raises the importance of resilience: a company whose operations depend on a single supplier of models, computing or logistics is more exposed to a shock than one with alternatives.

What should a small or medium sized business do in this race? I would suggest a plan built on realism about where it can win. The first step is to stop treating AI as a strategy and start treating it as a set of improvements to specific workflows. List the repeated tasks that consume the most staff time, such as answering common customer questions, preparing quotations, reconciling payments, reporting and onboarding new customers, and choose two or three where the output can be checked easily. The second step is to clean the information those tasks depend on, because an agent is only as good as the data and the documented procedures it can see, and the best predictor of a successful deployment I have seen is that a process was already well understood and written down. The third step is to run a small, measured trial with a clear owner, a baseline taken beforehand, and a human review of the results, extending only what works. The fourth is to keep the choice of tools flexible, so that the business can switch to cheaper or better models as prices fall, and to keep control of its own data and customer relationships. The fifth is to invest in people, since the evidence says that organisations where workers feel safe to experiment and are rewarded for improving how they work get the best results. And the sixth is to decide, honestly, what the business will not automate: the judgement calls and relationships that customers pay for.

From a Bangladeshi perspective, the race looks both encouraging and demanding. The encouraging side is that the falling cost of intelligence lowers the barrier for a small company or a single developer to build useful tools for local needs, in Bengali and for local workflows, which global vendors serve poorly. The cost of computing and software is priced in dollars, while revenue is largely in taka, so a business that uses these services must watch its unit costs closely and avoid being locked to a single provider. The strongest local opportunities lie where knowledge of language, law, payments and customers is the advantage: school and clinic management, accounting and compliance for small businesses, reconciliation across mobile wallets and gateways, export documentation, agriculture advice and customer service in Bengali. The weakest strategy, in my view, is to try to compete with global firms on general capability. The strongest is to use their capability as a component in a product that fits a local problem better than they can.

Now for my own predictions, stated as judgement and not as forecasts from any institution, and made from the vantage point of the beginning of June without knowledge of what the coming months will bring. Over the next twelve months I expect adoption statistics to keep rising while the share of companies reporting significant returns rises only slowly, with the gap explained by integration and not by technology. I expect the average price of intelligence to keep falling and the proportion of enterprise workloads running on open weight models to keep growing. I expect hybrid pricing to become the default for software with agent features and for several well known software firms to restructure how they charge. I expect the large providers to keep moving into vertical markets, leading to a wave of acquisitions of vertical start ups and to some failures. I expect the first serious public incident caused by an autonomous agent acting beyond its authority, at a company of consequence, to produce a tightening of controls and of customer questions about oversight. And I expect the companies that are widely described as winners a year from now to include a number that are quiet, specialised and profitable, and which are not mentioned in the conversation about the race today.

It may help to turn the three races into a simple scorecard that a business owner can apply to their own situation, because the abstractions are easy to nod at and hard to act on. On the first track, ask whether you know what the tools can and cannot do in the work you actually perform. That means having someone in the business who has tried the leading tools on real tasks from your own files, not on demonstrations, and who can tell the rest of the team honestly where they work and where they fail. A business that has this knowledge can take advantage of each improvement as it arrives. One that does not will either overreact to hype or ignore real change.

On the second track, ask whether any process has actually changed. Not whether people have been given access to a tool, but whether a specific workflow now runs differently, with a named owner, a baseline measurement and a review step. If the honest answer is that nothing has changed except that some employees use a chatbot, the business is still at the start of this race, however many licences it has bought. If one or two processes have been redesigned and measured, it is ahead of most of its peers, and the next step is to repeat the method rather than to buy more tools.

On the third track, ask what you are dependent on and what it would cost to change. For most small businesses this track is not about building computing capacity, which is the business of the giants, but about not being caught out by it. If your product or operations rely on one provider's models, one cloud, one payment gateway or one marketplace, you are exposed to that provider's prices and decisions. The practical test is to imagine that the provider doubled its prices or changed its terms next month, and to check whether you could switch within weeks. A business that passes the test has bought itself freedom to benefit from falling prices wherever they appear.

It also helps to see how the three races interact with a concrete example. Imagine a regional distributor of consumer goods with a few dozen staff, whose sales team spends much of its time preparing quotations, chasing payments and answering questions about orders. On the first track, the distributor does not need to build anything: it can buy access to capable tools cheaply. On the second, it needs to decide which of those tasks to redesign. A sensible choice is quotations, because the inputs, a price list, a customer record and a product catalogue, are structured, and the output can be checked by a salesperson before it goes out. After a few weeks of measured use, the time to produce a quotation falls, errors are caught in review, and the sales team spends the saved time on customer visits. On the third track, the distributor makes sure its price lists and customer data are held in a form it controls, so that it can use whichever tool is best next year. Nothing in this story required a large budget or an advanced team. It required a clear choice, a measurement and a decision to keep control of its own information, which is a description of most of the successful small deployments I have seen described.

A final point concerns the pace of decision making. The evidence suggests that the cost of waiting is rising, because competitors that redesign processes compound their advantage, learning faster about what works and building the habits that allow further change. But the cost of rushing is also real, as the security and quality failures of hastily deployed systems show. The balance I recommend is to move quickly on small, reversible experiments and slowly on anything that is hard to undo, such as handing control of money, customer records or regulated decisions to an automated system. A series of small experiments, each measured and reviewed, builds both the knowledge and the confidence to take larger steps when the evidence supports them.

The limits of this study are worth stating plainly. Much of the evidence on adoption and returns is self reported, which tends to flatter the respondents, and a number of the sources are vendors or investors with an interest in the story. Estimates of spending are surveys or projections and differ among providers. Studies of productivity measure specific tasks and may not generalise. Prices and valuations in this market move monthly, so figures quoted here will date quickly. Anything I say about the future is a judgement. I have tried to prefer primary publications and named surveys and to flag weaker sources, but readers should return to the originals and check the definitions and dates before relying on a number.

The short version is this. The race to build smarter businesses is really three races. On the first, the cost and capability of intelligence, progress is rapid and nobody can hold an advantage for long. On the second, integration into how a company actually works, progress is slow and uneven, and the winners are those who redesign their processes and support their people, not those who buy the most software. On the third, capital and computing capacity, the largest companies are spending at unprecedented scale on the assumption that businesses will find enough valuable uses, which the evidence suggests most are still working out. Within software, agents are replacing tools, vertical specialists are rising, pricing is shifting from seats to usage and outcomes, and control of the point of sale is being contested. For a business of any size the practical conclusion is the same: pick a few concrete workflows, measure the result, keep your choices flexible, protect your data and customers, and treat people as the thing that determines whether any of it works. The race rewards focus and discipline more than it rewards speed alone.

Sources: Stanford HAI, AI Index Report 2026, economy chapter (https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_4_economy.pdf); SAPinsider on the Stanford AI Index 2026 (https://sapinsider.org/blogs/stanford-ai-index-2026-enterprise-ai-readiness-governance-risk/); McKinsey, The state of AI in 2025 (https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai); WRITER, 2026 AI adoption survey (https://writer.com/blog/ai-adoption-survey-2026/); Capgemini Research Institute, World Cloud Report in Financial Services 2026, via Business Wire (https://www.businesswire.com/news/home/20251112597684/en); Menlo Ventures, 2025: The State of Generative AI in the Enterprise (https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/) and press release (https://www.globenewswire.com/news-release/2025/12/09/3202258/0/en/menlo-ventures-2025-state-of-generative-ai-report-enterprise-investment-hit-37b-in-2025-tripling-in-one-year.html); Microsoft, 2026 Work Trend Index (https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization); Futurum Group on AI capital expenditure (https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/); Alphabet, first quarter 2026 results (https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000043/googexhibit991q12026.htm); NVIDIA, first quarter fiscal 2027 results (https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000051/q1fy27pr.htm); Forbes on revenue per employee, 31 March 2026 (https://www.forbes.com/sites/paulbaier/2026/03/31/ai-native-firms-lead-in-revenue-per-employee/); Long Yield, The Token Economy (https://longyield.substack.com/p/the-token-economy-what-every-cfo); Deluair Consultancy on inference economics (https://deluair.com/consultancy/insights/ai-inference-economics-2026); Deepak Gupta on outcome-based pricing (https://guptadeepak.com/ai-native-outcome-based-pricing-2026/); Pickaxe on AI agent pricing models (https://pickaxe.co/post/ai-agent-pricing-models); Euclid Ventures, The Vertical Report 2026 (https://insights.euclid.vc/p/the-vertical-report-2026-full-version); Google, Universal Commerce Protocol announcement (https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/); CoinDesk on Stripe's annual letter (https://www.coindesk.com/business/2026/02/24/stripe-s-bridge-sees-stablecoin-volume-quadruple-as-utility-insulates-from-crypto-winter) and Tempo mainnet (https://www.coindesk.com/tech/2026/03/18/stripe-led-payments-blockchain-tempo-goes-live-with-protocol-for-ai-agents); Cloud Security Alliance, vibe coding security research notes, March 2026 (https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-generated-code-vulnerability-surge-2026/); company announcements from Harvey (25 March 2026), Decagon (January 2026) and Glean (28 May 2026).