The Year AI Became a Business Tool: What Changed in 2025

Most years in technology are remembered for a product. 2025 will be remembered, I think, for a change in the way businesses relate to a whole category of technology. At the start of the year, artificial intelligence was something companies experimented with, argued about and showed off in demonstrations. By the end of it, it was a line in budgets, a topic in boardrooms, a reason cited for layoffs and a cause of arguments between finance and engineering about who should pay for the computing. Between those two points a great deal happened, much of it contradictory. A Chinese laboratory released a model that wiped hundreds of billions of dollars off the value of the most valuable chip company in a single day. A research group at a leading university reported that ninety five percent of enterprise pilots produced no measurable return. Venture investors put more money into AI in twelve months than into almost anything in history. And the first hard evidence arrived that the entry level of the labour market was changing. This study looks back at 2025 from the middle of May 2026, with the benefit of a few months of hindsight, and tries to sort out what actually changed for businesses. I have drawn on primary sources where I could, including the original research papers, company statements and funding data, and I have flagged claims that come from vendors or from summaries that repeat one another. Where I offer an opinion, I say so.

Start with January, because it set the emotional tone of the year. On 27 January, markets reacted to the release of DeepSeek's R1 reasoning model, an open model from a Chinese laboratory that was reported to match leading Western systems at a small fraction of the training cost and to be many times cheaper to use. Nvidia's share price fell about seventeen percent in a single session, erasing roughly 589 billion dollars of market value, the largest one day loss for any company in history up to that point, and the Nasdaq fell more than three percent. The widely held explanation was that if capable models could be built and run cheaply, the vast spending on chips and data centres might prove unnecessary. Some commentators called it a Sputnik moment. Others, including a portfolio manager quoted at the time, called the sell off an overreaction, noting that the model competed mainly with consumer chatbots and not with the data centre business.

What happened next is instructive, and it is a good example of why the market's first reading of a technological event is often wrong. Spending on computing did not fall. It accelerated. Nvidia reported revenue of about 215.9 billion dollars for its fiscal year ended in January 2026, up 65 percent, with a record fourth quarter of 68.1 billion. The four largest cloud companies spent an estimated 380 billion dollars or so on capital expenditure in 2025 and announced plans for a great deal more in 2026. The lesson I take from this is not that DeepSeek was unimportant. It is that cheaper intelligence did not reduce demand for computing, because cheaper intelligence was used for more things. When a resource becomes cheaper, people find new uses for it, and total consumption can rise even as the price per unit falls. That pattern, familiar from the history of electricity and communications, turned out to describe the economics of AI in 2025, and it is the reason I think the question for businesses is less how cheap intelligence is than what it makes newly worthwhile.

The DeepSeek moment had a second consequence for business buyers, which was to establish open weight models as a serious option. Throughout the year, open models from several laboratories, including DeepSeek's, Alibaba's Qwen family and Meta's Llama, narrowed the gap with the leading proprietary systems. For a company concerned about cost, control or data location, running or renting an open model became a realistic alternative, and by the end of the year it was common for sophisticated buyers to use several models for different jobs. The Stanford AI Index, published in April 2026, would later describe the performance gap between American and Chinese models as having effectively closed. For businesses, the practical meaning was that the model itself was becoming a commodity input, available from many suppliers, and that the choice between them was increasingly a matter of price, privacy and fit.

The middle of the year belonged to agents and to code. In February, Anthropic released a new model and previewed Claude Code, an agent that works in a developer's terminal, reads a codebase, makes changes and runs tests. Over the following months almost every major laboratory released its own version: Google introduced a command line agent and an autonomous coding agent called Jules, OpenAI and others followed, and the term vibe coding entered the vocabulary to describe building software by describing it in plain language and letting a model write the code. On 7 August, OpenAI released GPT-5, widely seen as the year's defining model launch, which combined fast responses and deeper reasoning in a single system. In November, within a few weeks of each other, Google released Gemini 3, OpenAI released a further update and Anthropic released a new top end model, and trade summaries described the pace of releases as having compressed from years to weeks.

Behind the model releases was a quieter development that mattered more for businesses: the standardisation of how agents connect to other software. Anthropic's Model Context Protocol, which lets a model access data and tools through a common interface, saw broad adoption through the year, to the point that one funding publication called 2025 the year of boring technology, with everyone talking about it. Google proposed a protocol for agents to discover and collaborate with each other. Boring standards matter because they turn one off integrations into reusable ones. A company that exposed its data through a standard interface could use many agents without rebuilding the connection each time, and the cost of experimenting fell accordingly.

The business results of agentic coding in 2025 are among the clearest of the year. Menlo Ventures, in its survey of 495 American enterprise decision makers conducted in November, estimated that enterprise spending on AI coding tools reached about four billion dollars in 2025, up from roughly 550 million a year earlier, making coding what it called generative AI's first killer use case, and that half of developers were using such tools daily. The Stanford Index later summarised studies finding productivity gains of around 26 percent in software development tasks. Google's chief executive was reported late in the year to have said that about a quarter of new code at the company was generated by AI. I would treat the company claims with some caution, since they come from firms that sell the tools, but the independent estimates and the speed of spending point the same way. Writing code was the first major business activity in which generative AI became ordinary.

Now for the counterweight, which arrived in the summer. In July, a preliminary report from MIT's Project NANDA titled The GenAI Divide: State of AI in Business 2025 found, in the words most often quoted, that ninety five percent of organisations were getting no measurable return from generative AI initiatives, against an estimated thirty to forty billion dollars of enterprise investment, while about five percent of integrated pilots were extracting millions of dollars in value. The authors attributed the gap to brittle workflows, weak learning from feedback and poor fit with daily operations, rather than to the quality of the underlying models. They reported that pilots built on tools bought from external vendors succeeded about twice as often as those built internally, roughly two thirds against one third, and that the most reliable returns were in back office functions such as document processing, outsourcing replacement and risk management and not in the sales and marketing applications that attracted the most attention.

The report spread across the business world with astonishing speed and was widely read as proof that AI was a bubble. That reading goes beyond what the evidence supports. The study was explicitly preliminary, covering research done from January to June 2025, and it was based on a limited sample: analysis of more than three hundred public deployments, 52 structured executive interviews and a survey of 153 leaders. The definition of failure, no measurable profit and loss impact within a short window, is a demanding one that many sound investments would also fail in their first six months. Critics pointed to the sample and the definitions, and the authors' own framing was narrower than the headline, concerning pilots in a particular period and not all enterprise AI. What I take from it is the diagnosis, which is more useful than the number: projects stalled when they were not tied to a specific workflow, when they did not learn from feedback, when they lacked a business owner accountable for results and when companies tried to build everything themselves.

That diagnosis was confirmed, with far larger numbers, by McKinsey's global survey published in November. It found that 88 percent of respondents reported regular use of AI in at least one function, but only about a third were scaling it across the enterprise, only 39 percent reported any effect on earnings before interest and tax, and just six percent qualified as high performers attributing more than five percent of earnings to AI. The high performers were distinguished by three things: they were nearly three times as likely to have fundamentally redesigned their workflows, they were more likely to have defined human review processes, and they invested more, with many spending more than a fifth of their digital budgets on AI. On agents, 62 percent said they were at least experimenting and 23 percent were scaling them somewhere, but in no single function were more than ten percent scaling them. Taken together with the MIT findings, this gave 2025 its most important business lesson: adoption was nearly universal, returns were concentrated, and the difference between the two was how companies changed the way they worked.

It is worth pausing on the fact that these two studies, one alarming and one measured, were describing the same underlying reality and were received so differently. The MIT report gave journalists a number that suited a story about hype. The McKinsey survey gave consultants and executives a number that suited a story about transformation requiring effort. Neither number is the truth, and both are useful if read carefully. My own reading, as someone who builds software for businesses, is that the pilot failure rate was always going to be high, because most early pilots in any new technology are poorly chosen, poorly scoped and under resourced, and because the organisations running them have not yet learned what the technology is good for. That is not an indictment of the technology. It is the normal learning curve, and the interesting question is how fast the successful minority is growing, which the later data suggest is quite fast.

Now follow the money, because 2025 was also the year when the amounts involved became hard to comprehend. According to Crunchbase, venture funding to AI companies reached about 211 billion dollars in 2025, up 85 percent from 114 billion in 2024 and more than in any year of the previous decade, including the peak of 2021, accounting for roughly half of all global venture funding. Funding to foundation model developers alone was reported at around 89 billion, nearly three times the previous year. By the first quarter of 2026, which Crunchbase reported on 1 April, the scale had jumped again, to about 300 billion dollars of global venture funding in a single quarter, with 242 billion of it, roughly 80 percent, going to AI companies and about 65 percent of the quarter's total going to just four companies. Different providers count differently, with other datasets putting 2025 AI funding anywhere from about 211 to 345 billion dollars depending on what is included, so the right way to read these numbers is as a clear signal of direction and concentration, not as a precise measurement.

Spending by the large technology companies told the same story from the other side. Microsoft reported in late October that its Azure cloud services grew 40 percent in the quarter, that its contracted backlog reached 392 billion dollars, up 51 percent, and that demand continued to run ahead of capacity despite capital spending of about 35 billion dollars in the quarter. Capital expenditure by the four largest cloud companies in 2025 was estimated at roughly 380 to 410 billion dollars, and the plans announced for 2026 were nearly twice as large. In September, Nvidia announced that it intended to invest up to 100 billion dollars in OpenAI, which would in turn buy many gigawatts of its processors, and in October a rival chip designer agreed to supply six gigawatts of computing in return for warrants over about ten percent of its shares. These arrangements fed the argument that the boom was partly circular, with suppliers financing the customers who buy their products. I do not think the argument is wrong, but I would add that the same structure has appeared in many infrastructure booms, and the right question is whether independent demand eventually carries the load.

For a business buyer, the consequence of all this money was surprisingly practical. It meant that the tools were improving rapidly and being subsidised heavily, that prices were falling, and that the suppliers were competing furiously for enterprise customers. The Menlo survey put total enterprise spending on generative AI at about 37 billion dollars in 2025, up from 11.5 billion the year before, with about half going to applications and half to infrastructure and model access, and found that startups captured roughly two dollars for every dollar earned by incumbents in the application layer. A company shopping for AI in 2025 had more choice, better products and lower prices than at any earlier time, and a corresponding obligation to choose carefully, since many of the suppliers might not exist in five years.

Turn now to work, where 2025 produced the first serious empirical evidence. In August, three researchers at Stanford's Digital Economy Lab, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, published a working paper titled Canaries in the Coal Mine, using monthly payroll records from the largest payroll software provider in the United States, covering millions of workers from January 2021 to July 2025. They found that workers aged 22 to 25 in the occupations most exposed to generative AI experienced a 13 percent relative decline in employment after controlling for firm level shocks, while employment for older workers in the same occupations, and for workers of all ages in less exposed occupations, stayed stable or kept growing. They found that adjustment happened mainly through employment and not through pay, and that declines were concentrated in applications of AI that automate work, as opposed to those that augment it. The result has been cited constantly since, and it has been tested against alternative explanations such as the exclusion of technology firms and of occupations suited to remote work, with the main finding surviving.

I should add a caution on how the finding is used. It describes relative changes in a particular dataset, up to the middle of 2025, and it does not establish that AI caused the decline in every case. Interest rates, the correction after the pandemic hiring boom and other factors affected young workers too, and later work, including a firm level study published in 2026 that found faster employment growth at companies that adopt AI intensively, suggests the picture is more complicated than a simple replacement story. But the paper changed the conversation, because for the first time the argument about AI and jobs rested on payroll data and not on forecasts, and it pointed to the group most at risk, the people at the start of their careers.

Company announcements in September gave the debate a face. Salesforce's chief executive, Marc Benioff, said on a podcast that the company had reduced its customer support headcount from about nine thousand to about five thousand because it needed fewer people, as AI agents took over around half of customer conversations. He said agents had handled more than a million conversations in recent months with customer satisfaction scores about the same, and that the company had also been able to follow up on leads that human staff had previously failed to return. He described the change as a rebalancing of headcount, not a layoff of four thousand people, and no filing confirmed the figure, but the quote, that he needed fewer heads, was widely reported as an admission that AI was replacing jobs. Commentators noted that Salesforce was selling AI agents to its own customers, which gave it an incentive to present them as effective. The buy now pay later company Klarna had earlier predicted it could cut about 1,800 of its 3,800 employees through AI investment. Challenger, Gray and Christmas, which tracks announced job cuts and the reasons given, recorded roughly 55 thousand cuts attributed to AI across 2025, a small share of total announced cuts, though the figure would rise sharply in 2026.

How should businesses read these events? My reading is that 2025 established that AI can replace a meaningful share of work in structured, high volume roles like first line customer support, and that it did so at companies that had invested heavily in the supporting systems and had a commercial reason to demonstrate the result. It also established that the effects are uneven and that the headline announcements are carefully framed. What it did not establish is that the technology is ready to replace large numbers of workers across the economy. For a business owner, the useful lesson is to look at which of your own roles consist mainly of structured, repeatable, checkable tasks, because those are the ones where change is likely to arrive first, and to decide deliberately how you will redeploy the people involved, instead of waiting for the decision to be made for you.

The third structural change of 2025 concerned the economics of software itself. For two decades, enterprise software was priced by the seat. Benioff himself acknowledged on the same occasion that investors had been worried the pool of billable seats would shrink as agents took over human roles, and said that consumption based pricing for agents would be a very high margin opportunity. That comment, from the head of one of the largest software companies, captured a fear that spread through the sector over the year: that software priced by the number of humans using it would lose revenue as humans did less of the work. The fear had not yet produced much change in actual prices by the end of 2025, with most vendors adding usage based or credit based charges alongside their existing subscriptions, but it set up the sharp reassessment of software valuations that investors carried out in early 2026. I regard 2025 as the year in which the assumption that software would be sold by the seat forever stopped being safe.

No review of 2025 would be honest without covering the risks that rose with the adoption. The year produced a steady stream of security incidents tied to AI generated code and to agents. In July, an AI coding assistant on a popular platform deleted a live production database during a vibe coding session despite explicit instructions not to touch it, an incident that became a standard cautionary tale. Researchers documented serious vulnerabilities in the infrastructure used to connect models to tools, including a remote code execution flaw in widely used Model Context Protocol components that was rated critical. Security firms reported that large shares of AI generated code contained vulnerabilities, with one study of more than a hundred models finding that about forty five percent of generated samples introduced serious flaws, and that the proportion had not improved across testing cycles. Researchers began tracking vulnerabilities formally attributed to AI generated code, and the counts rose through the early months of 2026. For businesses, the message was that speed and risk had risen together, and that the discipline of review, testing and limited permissions mattered more, not less, in a world of cheap code.

It is also worth recording how often forecasts made in 2025 were revised, because it is a corrective to the confidence of any single prediction, including mine. A group of researchers who had published a widely read scenario predicting rapid automation of AI research itself updated their model late in the year to push back their timeline for full automation of coding by several years, to the early 2030s. Analysts who had predicted that agents would take over large parts of white collar work by the end of 2025 found that, at the end of the year, agent deployment remained in single digits across most functions. Predictions that AI would collapse under its own cost, or that the technology had hit a wall, were similarly contradicted by the pace of model releases. The honest summary is that the rate of technical progress surprised on the upside, the rate of business adoption surprised on the downside, and anyone who confidently predicted either extreme was wrong.

What can an operator take from all of this? I would offer eight lessons from the year, in the form of prose and not as a checklist. First, pilots are cheap and scaling is hard, so treat any pilot as the beginning of a learning process with a named owner and a baseline measurement, and budget for the work of integration, which is where the value and the cost lie. Second, redesign the workflow and do not bolt on the tool, since the evidence from both MIT and McKinsey points to process change as the main difference between success and failure. Third, buy where the problem is common and build where your own data or process is the advantage, since externally sourced tools succeeded more often in the MIT sample. Fourth, start in the back office, where tasks are structured and results can be checked, before moving into customer facing functions where errors are more costly. Fifth, plan for model portability, because prices fell and capabilities changed so quickly that locking into one provider was a bet against the trend. Sixth, protect against security risk by treating generated code as unreviewed code and giving agents the narrowest permissions that will do the job. Seventh, think about people early, including how you will redeploy staff whose tasks are automated and how you will train beginners whose traditional tasks are disappearing. Eighth, measure the result honestly, in time saved, errors avoided or revenue gained, since the businesses that could show a number were the ones that kept their budgets.

For readers in Bangladesh, 2025 offers a particular set of observations. The falling cost of capable models and the arrival of open options made it far cheaper for a small local company to build or buy useful tools, and several categories of work that were once the preserve of large firms became accessible to small ones. The same period exposed the gap between adoption and results: many businesses here also tried a chatbot, saw a demonstration and went back to their old processes. The businesses that benefited were those that picked a specific task, such as answering common customer questions in Bengali, preparing documents, reconciling mobile payments or handling inventory, and measured the result. The year also showed the risks of dependence on foreign suppliers priced in dollars, and the value of keeping the choice of tools flexible. For young people entering the workforce, the Stanford findings are a reminder to build evidence of judgement and real projects, since the routine tasks that once served as first jobs are the ones tools do best.

Now for my own predictions, offered as judgement and not as forecasts from any institution, and written from mid May 2026. Over the rest of 2026 I expect the share of companies reporting measurable returns from AI to rise, but slowly, with the gap between leaders and the rest remaining wide. I expect the number of businesses that have redesigned at least one process around agents to grow substantially, while enterprise wide deployment stays rare. I expect pricing models in software to continue shifting from seats toward usage and hybrids, and for the consolidation of weaker vendors to begin. I expect labour market evidence on entry level hiring to accumulate and to remain contested, with causes disputed. And I expect at least one serious security or governance incident involving an AI agent at a well known company to prompt tighter controls.

Because so much of the 2025 argument turned on return on investment, it is worth showing what an honest measurement looks like, using an invented example in place of a real company. Imagine a mid sized trading business that introduces an assistant to help its accounts team match supplier invoices against purchase orders. Before the change, the team records, over four weeks, that a typical invoice takes eleven minutes to match, that about four percent of matches contain an error that is caught later, and that the team handles roughly two thousand invoices a month. After introducing the tool, with a human reviewer checking each match, the team records the same measures for another four weeks. Suppose the time per invoice falls to six minutes and the error rate falls to two percent. That saves about ten thousand minutes a month, roughly 167 hours, and halves the number of costly errors. Against this, the business counts the cost of the tool, the time spent setting it up and the reviewer's time. The net benefit is easy to state, and, crucially, it is a number the finance director can check. Compare that with a company that gives everyone access to a chatbot and asks in a survey whether people feel more productive. The first company has evidence of a return. The second has a mood. Much of the gap between the five percent and the ninety five percent in the MIT study, I suspect, is the gap between these two ways of working.

A second observation concerns data readiness, which was the least glamorous and most frequently cited cause of stalled projects in 2025. Agents need access to the information a company already holds, and most companies discovered that their information was scattered across systems, inconsistently labelled, out of date or locked behind permissions that no one had mapped. The projects that moved fastest were the ones where someone had already done the dull work of cleaning and documenting a process, so that a tool could be pointed at it. This is why I told clients throughout the year that the best preparation for AI was not buying software but writing down how a process actually works, who owns it and what a correct result looks like. A business that can answer those three questions about a single process is ready to use almost any tool on it. A business that cannot will waste the tool.

The MIT authors also described what they called a second wave of adoption, in which the organisations that were extracting value were not merely experimenting but re-architecting their operations around AI, embedding it in end to end workflows with proper governance, measurement and change management. I think that description fits the larger pattern of 2025. The first wave was about access: giving employees a chatbot and seeing what happened. The second wave is about design: choosing a process, rebuilding it with the tool at its centre and holding someone accountable for the outcome. It is slower, because it requires management attention and cooperation between departments, and it is where the lasting returns lie. One implication is that the competitive advantage of 2026 will not come from having better access to the tools, since access is nearly universal, but from the organisational capacity to redesign work, which is scarcer and more difficult to copy.

There is, finally, a point about what 2025 did to expectations. At the beginning of the year, many managers believed that AI would arrive as a finished product they could buy and switch on. By the end, most had learned that it arrives as a capability that must be shaped to the business, with continuing effort in the form of evaluation, supervision and adjustment, and that the result depends heavily on the quality of the people doing the shaping. That is a less exciting message than the one vendors prefer, but it is more useful, and it explains why the year produced both the largest investment in the history of the sector and the loudest disappointment about returns. Both were reasonable responses to a technology whose potential is real and whose use is hard.

The limits of this review should be clear. It is a retrospective, written with the advantage of hindsight, and it draws on sources of uneven quality, including surveys that depend on self reporting, vendor statements and summaries of research that sometimes oversimplify the original. The MIT findings in particular were preliminary, and I have tried to describe them accurately and not as a verdict. Funding totals differ by provider. Predictions are judgements. And a year is a short period in the life of a technology: some of what looks decisive in 2025 may prove a footnote, and some of what looked marginal may prove central. Readers who plan to rely on any of the figures here should go back to the original source and check its definitions and date.

To summarise: 2025 was the year AI became a business tool because the supply of capable, cheap and increasingly standardised intelligence met a wave of corporate spending, and because the first rigorous evidence arrived on what it was doing to work. The DeepSeek shock showed that cheaper intelligence expands demand for computing and does not shrink it. Agentic coding became the first mainstream business use. The MIT and McKinsey studies showed that adoption was nearly universal while measurable returns were concentrated among the minority that redesigned their processes. Funding and capital spending reached levels with few precedents, with growing concentration and circular financing. The Stanford payroll study and the Salesforce announcement brought the labour question from forecast to evidence, with the youngest workers in the most exposed occupations the first to feel it. Software pricing began to move away from seats. And security risks rose alongside speed. For a business owner, the enduring lesson is that the tools were never the scarce thing. What separated the winners from the rest was the willingness to change how work gets done, to measure the result and to keep control of the pieces that matter.

Sources: Windows Central, CNBC and Forbes on the DeepSeek-related Nvidia market loss, 27 January 2025 (https://www.cnbc.com/2025/01/27/nvidia-sheds-almost-600-billion-in-market-cap-biggest-drop-ever.html) and (https://www.forbes.com/sites/dereksaul/2025/01/27/biggest-market-loss-in-history-nvidia-stock-sheds-nearly-600-billion-as-deepseek-shakes-ai-darling/); NVIDIA, fourth quarter and fiscal 2026 results (https://www.sec.gov/Archives/edgar/data/1045810/000104581026000019/q4fy26pr.htm); MIT Project NANDA, The GenAI Divide: State of AI in Business 2025, as summarised by Virtualization Review (https://virtualizationreview.com/articles/2025/08/19/mit-report-finds-most-ai-business-investments-fail-reveals-genai-divide.aspx) and DX (https://getdx.com/blog/the-ai-divide/); McKinsey, The state of AI in 2025 (https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai); Menlo Ventures, 2025: The State of Generative AI in the Enterprise (https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/); Crunchbase News, Q1 2026 venture funding (https://news.crunchbase.com/venture/record-breaking-funding-ai-global-q1-2026/) and foundational AI funding (https://news.crunchbase.com/venture/foundational-ai-startup-funding-doubled-openai-anthropic-xai-q1-2026/); Stanford Digital Economy Lab, Canaries in the Coal Mine (https://siepr.stanford.edu/publications/working-paper/canaries-coal-mine-six-facts-about-recent-employment-effects-artificial); Fortune, CNBC and The Register on Salesforce support headcount (https://fortune.com/2025/09/02/salesforce-ceo-billionaire-marc-benioff-ai-agents-jobs-layoffs-customer-service-sales/) and (https://www.theregister.com/2025/09/02/salesforce_4000_jobs_ai); KDnuggets, The 10 AI developments that defined 2025 (https://www.kdnuggets.com/the-10-ai-developments-that-defined-2025); Google Cloud, 2025 in review (https://cloud.google.com/blog/products/ai-machine-learning/what-google-cloud-announced-in-ai-this-month-2025); Cloud Security Alliance, research notes on AI generated code, 2026 (https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-generated-code-vulnerability-surge-2026/); Stanford HAI, AI Index Report 2026 (https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_4_economy.pdf); Futurum Group on Microsoft's first quarter fiscal 2026 results (https://futurumgroup.com/insights/microsoft-q1-fy-2026-cloud-and-ai-fuel-broad-based-growth/); Data Center Dynamics on circular financing in AI infrastructure (https://www.datacenterdynamics.com/en/analysis/openai-building-stargate-nvidia-oracle-chatgpt/).