The state of AI: our starting point
A visual, plain-English guide to AI in November 2026: capabilities, agents, costs, jobs, risks and India. Our starting point for tracking what changes.
A snapshot of where we stand on 11 November 2026 (India time).
I’m starting AI_First_World with a simple idea: let’s write down where we are before we get carried away with where we might be going.
AI can already help us write, build software, make images, work through documents and use other software. But “AI can do this” can mean very different things. Sometimes it means a carefully prepared demo. Sometimes it means a tool people use every day. Sometimes it means work you can trust enough to hand over.
This first post is a starting point for following those differences. I want to come back to it as things change, including the things we get wrong.
You don’t need a technical background to read it. I’ve kept the explanations short and put the bigger comparisons into graphics. Sources sit beside the claims they support.
An editorial guide to this article, not a numerical ranking. Baseline: 11 November 2026.
1. How we got here
AI was part of everyday life before chatbots became popular. Spam filters, recommendations and fraud detection are familiar examples. These systems usually did a particular job.
The recent change is that one system can help with many different jobs, and we can tell it what we want in normal language.
A model is the underlying system that has learned patterns from data. An AI product adds things around it: a chat window, access to documents, search, tools and rules about what it can do.
That difference matters. A model on its own may be unable to check today’s news. A product using that same model may search the web and read the results. The experience depends on the whole setup.
There are a few overlapping changes worth keeping in mind:
- From predicting to creating. Systems can produce text, images, audio and other material.
- From one format to several. Some can work across words, pictures, speech and video. You’ll hear this called “multimodal”.
- From answering to taking steps. With access to tools, they can search, write files, run code or operate software.
- From a quick response to more work on a problem. Some systems spend more computing effort checking or exploring an answer.
These changes haven’t replaced earlier AI. A simple forecasting system may still be the right tool for a shop deciding how much stock to order.
The World Bank’s 2026 report makes a useful point here: countries can benefit from smaller, practical AI applications, including tools that work through basic phones. The most useful application doesn’t have to involve the biggest model.
2. What can AI actually do?
The honest answer depends on the task, the information it gets and how carefully someone checks the result.
Here is a practical way to look at it. This is a guide to using the technology, not a ranking of every available product.
| Type of work | What you can ask it to do | What still needs attention |
|---|---|---|
| Writing | Draft, rewrite, translate and organise ideas | Facts, intent, tone and whether it says anything useful |
| Research | Find material, compare documents and explain unfamiliar subjects | Whether sources exist, support the claim and are current |
| Coding | Write code, investigate bugs and attempt changes across a project | Tests, security, integration and long-term maintenance |
| Images, audio and video | Generate or edit material from instructions | Exact details, consistency, consent and rights to use it |
| Computer tasks | Navigate software and carry out a sequence of actions | Permissions, unexpected screens and costly mistakes |
| Physical tasks | Help robots recognise objects and plan actions | Safety and reliability outside the tested setting |
One reason the discussion gets confusing is that these abilities don’t improve evenly.
Stanford’s 2026 AI Index describes systems doing very well on difficult maths while still struggling with some apparently simple visual tasks. Its examples draw on specific tests and earlier reporting periods. They aren’t a live scorecard for every model available this November.
For a more recent reference, Artificial Analysis changed its evaluation in October 2026, putting more weight on private tests and work that is harder to complete. Its results still leave substantial room for improvement. No single model leads every test.
That is why I’d be careful with statements like “AI is smarter than humans”. Smarter at which task? With which tools? How many attempts? And who checked the answer?
For a real job, the useful question is smaller: can this system complete this particular task to the standard we need?
3. What people mean by an AI agent
An assistant helps you do something. An agent can take some of the steps itself.
There isn’t a clean industry-wide dividing line, so it helps to use an example.
Suppose a customer asks where their order is. A chatbot might draft a polite reply. A fixed workflow might look up the order number and insert the delivery status. An agent might decide which records to inspect, investigate a delay and suggest what to do next.
- 1AssistantYou ask. It answers.Draft a reply for you to review.
- 2WorkflowSoftware follows a set route.Read a ticket, classify it, draft a reply.
- 3AgentAI chooses steps and uses tools.Look up the order, investigate, propose a fix.
Illustrative examples. More freedom needs clearer permissions, checks and a way to hand the work back to a person.
The extra freedom can be useful. It also creates more ways for things to go wrong.
Looking up an order and issuing a refund are different responsibilities. A sensible system should know which actions it can take, which need approval and when it should stop.
Memory helps it keep track of relevant information. Tools let it act. Checks help establish whether the work is right. None of these removes the need to decide who is responsible.
Anthropic’s June 2026 Economic Index gives one sign of this shift. The company changed how it studies usage because longer tasks in Claude Code and Cowork were no longer well described by looking only at chat conversations. That tells us something about its products, rather than the whole economy.
The question I want to track is how much useful work people can hand over without spending the saved time supervising and fixing it.
4. How much should we trust the scores?
A benchmark is a test used to compare systems. It can be useful, but it only tells us about the work it measures.
A high score on a maths test doesn’t tell us whether an AI can handle an unhappy customer. A coding test doesn’t automatically show that it can maintain a complicated business system.
Even two results on the “same” test can differ because one system gets more time, extra tools or several attempts.
There is also a problem when questions, or their answers, become familiar through training data. Private tests help reduce that risk, although they make independent inspection harder.
Artificial Analysis’s Index v5 announcement, dated 27 October 2026, says 60% of its index uses private test sets or held-out answers. A score from an earlier version shouldn’t be compared directly without checking what changed.
Another widely discussed measure is METR’s “time horizon”. It describes the difficulty of tasks an agent can complete, using the time a human expert would take.
A two-hour task at 50% success does not mean two hours of dependable, unattended work. It means the system is predicted to succeed half the time on tasks of that measured difficulty.
METR’s methodology page also says the tasks mainly involve software, machine learning and cybersecurity. Its page was last updated on 8 September 2026 and is no longer actively updated. It cannot tell us the current performance of every new release.
When I use a score in this series, I want the test name, date and important conditions to travel with it.
5. Who is using it, and how deeply?
Trying a chatbot once is different from using it every week. Both are different from a business depending on it to serve customers.
An employee might use AI regularly while their employer is still deciding how to handle private data. A company might buy thousands of accounts while only a few teams use them well.
We should keep those situations separate.
Two different survey questions. These are reported experiences, not measured gains or stages of a funnel.
McKinsey survey · May–June 2026 · 1,719 respondentsThe graphic uses McKinsey’s 2026 survey, conducted from May to June with 1,719 respondents across 97 countries. Eighty percent said AI improved their individual productivity. Thirty-seven percent reported some contribution to company operating profit, meaning profit before interest and tax.
These are answers to different questions, based on people’s reports. They aren’t independently measured gains, and the sample isn’t a census of all businesses.
Still, the distinction is useful. Someone finishing a presentation faster doesn’t automatically change the company’s finances. The saved time may be spent elsewhere. Other steps may still be slow. Extra checking may absorb some of the gain.
For consumers, I’d track repeat use and whether it solves a problem worth returning for. For organisations, I’d look for work that runs regularly, has an owner and produces an outcome someone measures.
6. Does it make people more productive?
Sometimes. How much depends on the person, the job and the setup.
That sounds less exciting than one big percentage, but it is much more useful.
METR’s earlier experiment found experienced open-source developers took 19% longer on their tasks when allowed to use the AI tools studied in early 2025.
It would be a mistake to turn that into “AI makes developers slower” in November 2026.
In its February 2026 follow-up, METR said its newer experiment could not give a reliable estimate. Developers were increasingly unwilling to work without AI, the tasks submitted could be selective, and measuring time became harder when people used several agents. The researchers thought speedups had probably improved, but the study provided only weak evidence of their size.
That is a useful example of research being honest about what it can establish.
If I were checking an AI workflow, I’d measure:
- Time from starting the job to an accepted result.
- Errors and the time spent correcting them.
- Whether quality improved or declined.
- How much attention a person still had to give it.
- Whether the change helped a customer or the business.
A tool can make the first draft faster and the whole job slower. It can also make work possible that nobody previously had the time or money to attempt. Both effects belong in the picture.
7. What does useful AI cost?
There are two large buckets: building the model and running it.
Training is the work of teaching a model from data. Inference is using it to produce an answer or take a step. For most people buying an AI tool, the second bucket is closer to the bill they see.
But the model bill is only part of the cost.
There may also be software subscriptions, document preparation, search tools, storage, monitoring and human review. A failed attempt can cost money without producing anything usable.
Average model cost per Intelligence Index task at this dated snapshot, among its ten highest-scoring models. Different models, different results. This excludes human review and is not a price quote for your work.
Artificial Analysis · Index v5 launch · 27 October 2026The comparison above is deliberately narrow: the model cost of a task in one dated evaluation. It shows why model choice matters. It does not establish which model is best value for your business.
A cheaper model that needs repeated attempts may cost more per useful result. An expensive model may be unnecessary for a simple sorting job.
For a practical comparison, I’d use:
Cost per accepted result = total cost of the work ÷ results that meet the required standard.
Include review and failed attempts in that total. Compare quality too.
There is a wider business question here. Who keeps the value: chip suppliers, cloud providers, model companies, application builders or the organisations using them?
My expectation is that the answer will vary by market. Owning a model is one advantage. Having trusted distribution, useful data and a product that fits someone’s daily work can be advantages too. Those are possibilities to investigate, not a settled list of winners.
8. Open models, closed models and dependence
Some models are available only through a provider’s service. Others release their weights, the learned numerical settings needed to run the model.
Released weights can give organisations more options about where and how to run a system. That may help with control, adaptation or dependence on a supplier.
It doesn’t automatically mean the model is cheap to operate, fully transparent or free of restrictions. The licence, computing requirements and available expertise still matter. “Open weights” also doesn’t necessarily mean the training data and full development process are open.
For this baseline, I’d keep three questions separate:
- Can we access a capable model?
- Can we use it affordably and reliably?
- Can we change providers or run it ourselves if we need to?
A country or business can do well on the first question while remaining dependent on the other two.
The World Bank’s 2026 development report highlights the need to adapt AI to local languages, institutions and data, while avoiding unnecessary dependence on one supplier. That is especially relevant for places whose needs aren’t the main priority of global technology companies.
9. The physical world behind the chat window
AI runs in buildings full of computers. Those computers need chips, electricity, cooling and network connections.
A software release can happen quickly. Adding power supply or building a data centre takes longer.
Annual growth, not shares of global electricity. AI-focused centres are part of the total; do not add these bars together.
IEA · Key Questions on Energy and AI, 2026 · 2025 dataThe IEA’s 2026 report estimates that electricity demand from all data centres grew 17% in 2025. AI-focused centres grew faster, at 50%.
At the same time, the IEA reports improving energy efficiency per task. Those statements can both be true: each task can become more efficient while more people run more demanding tasks.
This is why a single “energy per prompt” number is a poor description of the entire system. A short text answer, a generated video and a long tool-using task are different workloads.
Data also matters. A useful system needs suitable examples to learn from and dependable information when it is used. Public text alone won’t tell it whether a particular invoice has been paid.
For the next stage, I want to follow the practical constraints: access to chips, power connections, local energy and water pressures, reliable data and the money needed to keep expanding. Announced investment and operating capacity should be counted separately.
10. Jobs, learning and the people doing the work
A job is a collection of tasks.
A salesperson may research companies, write messages, speak with buyers, negotiate, update records and decide where to spend time. Automating one part changes the job. It doesn’t tell us by itself whether the job disappears.
Workers worldwide in occupations with some exposure to generative AI, in the 2025 estimate.
The ILO–NASK study published in May 2025 estimated that one in four workers worldwide was in an occupation with some exposure to generative AI. That is a task-based estimate of potential exposure. It is not a count of people who lost work, or a prediction that a quarter of jobs will vanish.
We still need to watch actual hiring, pay, working hours and job quality. A company may respond to better productivity by doing more work, lowering prices, hiring differently or reducing staff. Several responses can happen together.
I’m particularly interested in the learning path for people starting their careers. If AI handles much of the beginner work, how do beginners get enough practice to develop judgement? That is an open question worth following.
There is a similar issue in education. Finishing an assignment and understanding it are different outcomes.
An October 2026 Microsoft Research preprint studied 297 workers doing AI-assisted writing. Support that encouraged planning improved writing quality and showed evidence of better recall, even though participants rated the tool as less helpful. It is a limited experiment, not a verdict on all learning.
The interesting possibility is that a little effort can help us learn, even when removing that effort feels more convenient.
11. Science, medicine and robots
Some of AI’s most useful work may be far from the chat window.
For example, the AlphaFold 3 research paper, published in 2024, described improved predictions of structures involving proteins and other molecules. That can help researchers decide what to investigate.
A prediction still needs to survive the next steps. A promising molecule is not an approved medicine.
The same distinction matters when reading claims about medical tools or robots. I’d place each claim on a simple ladder:
| Evidence stage | What it tells us |
|---|---|
| A demonstration | It worked in the situation shown |
| A controlled test | It worked under specified conditions |
| An independent evaluation | Someone else checked it |
| A real-world deployment | People used it outside the test |
| Measured outcomes | We know something about benefits, failures and costs |
A system can be valuable at an early stage. We just shouldn’t quietly move it up the ladder.
For a medical tool, outcomes might include missed cases, false alarms and whether patients actually receive better care. For a robot, they might include successful tasks, interventions, downtime and safety in the intended environment.
I want this series to follow those steps. A striking video is a starting point for questions.
12. Trust, harm and the rules around AI
Errors are only one part of the problem.
A system can produce misleading information confidently. It can expose private material if access is poorly managed. Generated media can be used to impersonate people. An agent reading an untrusted page can encounter instructions intended to make it do something the user never asked for. That last problem is often called prompt injection.
The seriousness depends on what the system can access and do. A wrong draft and an unauthorised payment have very different consequences.
The UK AI Security Institute’s Frontier AI Trends Report examines capabilities and safeguards across areas such as cybersecurity and biology. Its stated model coverage ends in October 2025. Those tests are useful evidence about tested systems and conditions; they are not a live assessment of every November 2026 product.
I’d keep three kinds of claims separate: harm observed in practice, behaviour found in a test, and possible future risks. Each deserves attention, but they support different conclusions.
Rules are also becoming part of everyday product decisions. The European Commission’s current AI Act timeline lists transparency rules among the provisions applying from August 2026, with other requirements phased in. There is no single global rulebook, and the details depend on the use and jurisdiction.
Copyright, consent and payment for creative work remain important questions too. Being able to generate something does not settle whether someone has the right to use it.
For ordinary users, trust should include the ability to inspect important evidence, correct a result and reach a person when the system cannot help.
13. Where India fits
India is worth watching both as a builder of AI and as a place where useful applications could reach a very large number of people.
Those are related ambitions, but they need different measures.
Government-reported figures with different reporting dates. GPUs are specialised computing chips. Selected proposals are not completed models.
Government of India · PIB factsheet · 13 August 2026The government’s August 2026 factsheet reports shared computing capacity, subsidised access and selected model proposals under the IndiaAI Mission. These show investment and activity. They don’t yet tell us how much people’s lives have improved.
For that, I’d look closer to the user.
Can a small business owner get useful help in the language they normally speak? Does it understand a mixed Hindi-English conversation? Can it cope with a noisy phone call or an incomplete document? Can someone afford it, and can they reach a person when it gets something wrong?
Consider a small manufacturer trying to respond to an enquiry. An AI tool could help read the request, prepare a draft quotation and find missing information. The result only becomes useful when it fits the actual product catalogue, prices and approval process. That is an example of a possible application, not a claimed deployment.
In public services, I’d ask whether people complete an application more easily, receive clearer information or spend less time travelling between offices. A chatbot conversation count wouldn’t answer those questions.
For education, agriculture and healthcare, local expertise and a way to handle mistakes matter as much as language support.
There is also a question for India’s technology workforce. If routine digital work becomes easier to automate, what services will clients still pay for? Understanding a messy problem, connecting systems and taking responsibility for the result may become more valuable. That is my working hypothesis, and it needs evidence over time.
I’ll explore these questions more closely in Ideas for India. For this series, India will remain part of the global picture.
14. What happens from here?
I don’t think this starting post needs a date for “AGI”, a term people use for broadly capable AI without always agreeing on its meaning.
It needs a way to recognise meaningful change.
Several paths seem possible. More capable agents could take on larger pieces of work. Smaller models could make useful AI cheaper and more widely available. Infrastructure costs could favour a few large providers. Trust failures or disappointing results could slow adoption in some areas. Different paths may happen at the same time.
I’d look for evidence that helps distinguish them: lower costs for accepted work, fewer interventions, stronger results outside demos, wider access and changes in how people spend their time.
A proposed checklist for future posts. No invented overall “AI progress” score.
My starting view is that AI is already useful enough to deserve serious attention. Turning that usefulness into dependable everyday results still takes work around the model: information, permissions, checks and people who understand the job.
The next posts will go deeper into individual parts of this picture. When the evidence changes, I want to be able to say exactly what changed.
For now, this is where we started.
A note on the sources
This is the first published draft of a baseline dated 11 November 2026. The date of the article is not the date of every dataset. Annual reports, experiments and government updates cover different periods; those dates are stated where they matter.
Links beside the claims lead to the research, evaluation methods or official reports used. Company usage data describes that company’s products. Surveys describe their respondents. Test results describe the tested setup. The graphics are original presentations of those sources, with the important limits included.
The examples and future questions are editorial explanations, not additional measured findings. This post does not claim to list every model or independently test each product. Corrections to the baseline should be labelled; later developments belong in follow-up posts.