AI in 2031: What Will Actually Change in the Next Five Years

Start with five years ago

In October 2021, there was no ChatGPT. It launched at the end of November 2022. According to Stanford's 2026 AI Index, generative AI reached 53% adoption among people within three years, and 88% of organisations now use AI somewhere in the business.

So when someone asks what AI will look like in 2031, the honest starting point is humility. Five years is long enough for the whole conversation to change shape. But it is not a mystery either. A handful of trends have been measured carefully for several years, and they are the best guide we have. This article follows those trends, says where they point, and is clear about where the evidence runs out.

The number worth watching: how long AI can work on its own

Most headlines compare models on exam scores. A more useful question is: how long a task can an AI system complete without a person stepping in?

METR, an independent research group, measures exactly this. They time skilled people on hundreds of real software and research tasks, then check which tasks each AI model can finish. The result is a "time horizon": the length of task, measured in human working time, that a model completes successfully about half of the time.

The trend is striking. Across 2019 to 2025, that horizon doubled roughly every seven months. In METR's updated January 2026 analysis, models released since 2023 have been doubling roughly every four months. The strongest models they measured could handle tasks that take a skilled person around five hours.

Now the careful part. Extending a line on a chart is not a forecast. METR itself notes that its measurements get thin at the top end, because few of its long tasks have been timed with real people, and that the trend is sensitive to which tasks are included. But suppose progress slows to one doubling every fourteen months, half the long-run pace. Five more years would still make the horizon about twenty times longer: from an afternoon's work to a couple of weeks'. That is the difference between an assistant that answers questions and a colleague that takes on a project.

Intelligence is getting cheap, fast

The second trend gets less attention, but for most businesses it matters more.

Epoch AI tracks what it costs to get a given level of performance from a language model. In its 2025 analysis, the price of reaching a fixed level of capability fell between 9 and 900 times per year, depending on the benchmark. For GPT-4-level performance on PhD-level science questions, the price fell about 40 times in a year.

Think about what that does over five years. Tasks that are too expensive to automate today, such as reading every customer call, checking every invoice line or giving every student feedback on every answer, become cheap enough to do routinely. Many of the biggest changes of the next five years will not come from a breakthrough model. They will come from yesterday's capability becoming almost free.

From chatbots to agents, unevenly

The step from answering to acting is already visible. On OSWorld, a benchmark where an AI has to complete real tasks on a computer by clicking, typing and switching between apps, success rates rose from about 12% to about 66% in a year, according to the 2026 AI Index. On a widely used coding benchmark, SWE-bench Verified, performance went from 60% to near the top of the scale over the same period.

Yet the same report describes a "jagged frontier". The best model it tested read an analogue clock correctly only about half the time. AI systems that write working software can still fail at things a child finds easy.

That jaggedness is the most practical thing to understand about the next five years. Agents will take over whole chains of work in some areas while staying unreliable at surprising steps in others. The organisations that do well will be the ones that know exactly where that line falls in their own work, and design checks around it.

What it means for work

The World Economic Forum's Future of Jobs Report 2025, based on data from more than 1,000 employers, expects 170 million jobs to be created and 92 million displaced between 2025 and 2030. That is a net gain of 78 million, but it also means change for about 22% of today's jobs. Employers expect nearly 40% of the skills their jobs need to change in the same period.

The more useful way to read those numbers is in terms of tasks, not job titles. Few roles will vanish whole. Most will be rebuilt: the routine parts handed to software, the judgement, relationships and responsibility left with people. A teacher who no longer spends evenings checking practice sheets, an engineer whose first draft of an estimate is ready in minutes, an accountant who reviews exceptions instead of every entry.

There is also a gap in expectations. Stanford's report found that 73% of AI experts expect a positive effect on jobs, compared with 23% of the public. Closing that gap will depend less on new technology than on whether people see the gains shared with them.

The brakes nobody puts in the headline

Three things could slow all of this down.

Electricity. The International Energy Agency estimates that data centres used about 415 TWh of electricity in 2024, around 1.5% of the world's total, and that this will roughly double to about 945 TWh by 2030, with AI the main driver. Power plants and grid connections take years to build. In many places, the limit on AI will be measured in megawatts, not in ideas.

Reliability. The AI Index counted 362 documented AI incidents in 2025, up from 233 the year before. As AI systems act on their own, a mistake is no longer a wrong sentence on a screen; it can be a wrong payment or a wrong email to a customer. Testing, permissions and audit trails will decide how much autonomy anyone is willing to grant.

Rules. Regulation is arriving, if more slowly than first planned. In the European Union, the AI Act's obligations for high-risk systems, such as those used in hiring, education and credit, were pushed back in 2026 and now apply from December 2027. In India, the Digital Personal Data Protection Rules were notified in November 2025, and the main obligations, including consent, user rights and stricter handling of children's data, take effect in May 2027. Any AI product that touches personal data in India needs to be designed for those rules now, not retrofitted later.

India's particular five years

India comes into this period with some real advantages. It has one of the world's largest pools of software talent, and under the IndiaAI Mission the government has made tens of thousands of GPUs available to startups and researchers at subsidised hourly rates, so building with AI no longer requires owning a data centre.

The bigger opportunity is that cheap AI fits India's needs unusually well. Explaining a concept patiently in Hindi or Bhojpuri, reading a handwritten form, answering a farmer's question by voice: these are exactly the language-heavy, high-volume tasks where AI is strongest and where human time has always been too scarce. The constraint will be trust: whether people's data is handled with care, and whether the answers are right often enough to rely on.

Our own predictions for 2031

Everything above comes from published measurements. This section is different: it is our judgement, as a team that builds AI software, about where those trends lead. We have said how confident we are in each, and we would rather be specific and occasionally wrong than vague and never useful.

• Very likely: you will talk to most business software. Menus and forms won't disappear, but "just ask it" will become the usual way into an app. Instead of hunting through five screens for last month's pending payments, people will ask, check the answer and act on it.
• Likely: whole back-office processes will run end to end. An invoice arrives, gets read, matched, checked and queued for payment; an admission enquiry turns into a counselling call, a form and a fee receipt. People will approve the exceptions rather than push every step. By 2031 this will be ordinary in mid-sized companies, not just large ones.
• Likely: the next hundred million Indian users will meet AI by speaking, not typing. Voice in Indian languages will become good enough for everyday use: a shopkeeper checking stock, a parent asking about a child's progress, a farmer asking about a crop disease.
• Likely: every student will have a patient personal tutor, and good teachers will matter more, not less. When explanations are free and endless, what becomes scarce is someone who motivates, notices when a student is stuck and checks real understanding. We expect more assessment to move into the classroom, through oral questions, projects and supervised work, because take-home work will no longer prove much.
• Likely: a lot of AI will run on your own device. As capable models get smaller and cheaper, many everyday tasks will run on a phone or laptop without sending data anywhere. That will matter in India, both for privacy and for places with patchy internet.
• Plausible: "checked by a person" will become a selling point. As AI-made text, images and video become impossible to tell apart, people will start to value proof of where content came from and who stands behind it.
• Unlikely by 2031: AI replacing doctors, teachers or lawyers outright, or household robots in ordinary Indian homes. The software will do much of their paperwork and first drafts. The responsibility, and the physical world, will stay mostly with people for longer than the headlines suggest.

Our biggest bet sits underneath all of these. By 2031, the gap between organisations will not be which AI model they use, because good models will be cheap and everywhere. It will be whose data is clean, whose processes are clear, and whose customers trust them with their information.

What a sensible organisation does now

No one can tell you what the best model of 2031 will be. You don't need to know that to prepare. A few steps hold up under almost any version of the future:

• Pick one workflow that eats hours every week, measure it today, and redesign it with AI. Learn from a real result, not from a demo.
• Keep your data clean, organised and owned by you. Every future AI system will only be as useful as the information you can give it.
• Build with the assumption that models will change. Keep the model behind a layer you control, so you can switch to something cheaper or better without starting again.
• Put a person where the stakes are high. Let AI draft, sort and check; let people decide what has real consequences.
• Treat privacy as part of the design. With India's data protection rules arriving in 2027, it is cheaper to get this right at the start.

What we honestly don't know

Trend lines bend. Progress could slow if the gains from larger models flatten out, if power runs short, or if a serious failure leads to tighter limits. It could also speed up in ways that are hard to picture today. That includes our own predictions above: they are reasoned bets, not certainties, and anyone who claims to know exactly what 2031 looks like is selling something.

What the evidence does support is this: AI will be able to work for longer on its own, it will cost far less, and it will be held to clearer rules. The interesting question for most organisations is not whether that happens, but which of their own tasks will look completely different when it does.

Sources: METR · Stanford AI Index 2026 · Epoch AI · WEF · IEA · EU AI Act changes · India DPDP