AI Futures — Dr Greg Baker Editorial reconstruction, September 2026. The next few years of AI ======================== Suppose intelligence becomes much cheaper over the next five years. What happens to the value of your business? That is the question I want to work through. There are plenty of demonstrations of impressive software. What matters here is whether any of that changes a decision you have to make: an investment, a product, a supplier contract, or the way your people work. We will look at the evidence for changing capability, then follow the consequences through a business. Some activities become cheaper. Other things become the constraint. A competitor might get there before you do. Keep one actual decision beside you as you work through the course. By the end, you should have a clearer strategic choice, a small experiment and a result that would make you change your mind. Greg Baker ========== I work across artificial intelligence, software, teaching and business. That gives me several ways to be wrong about the future, which is probably useful in a session like this. I have built software businesses, worked as a consulting chief technology officer, and taught people who will have to earn a living with these tools. I am also a composer. So when somebody says that creative work is safely outside the reach of automation, I tend to ask which part of the work they mean. The examples here come from research, from things I have tried, and from reasoning about how businesses work. Those are different kinds of evidence. A controlled experiment deserves different treatment from my own experience. We will keep the distinction visible. What decision brought you here? =============================== Choose a decision that is expensive to reverse. It could be hiring a team, buying a business, developing a product, taking a lease, or signing a long supplier contract. You do not need to disclose anything confidential. Write down the assumption that makes the decision attractive. Perhaps customers will keep paying for specialist analysis. Perhaps your distribution network will remain difficult to reproduce. Perhaps a physical installation will take competitors years to match. Then put a date against it. A decision that pays back in six months is exposed to a different future from one that must keep earning for ten years. We will return to this decision repeatedly. If your answer changes, that is useful. The point is to discover which assumption needs checking before you commit more money. Expertise is becoming easier to reach ===================================== One of my own examples is preparing and presenting an employment case. I used AI for research, drafting and editing, and I remained responsible for checking what went into the documents. The useful business question is what changes when an individual can prepare work that previously required a larger budget or a team. Customers, employees, small suppliers and competitors all gain access to that capability. Your organisation may encounter much better prepared people on the other side of a negotiation. There is a public decision in my case, and there is my account of how I used the tools. The decision does not establish that AI caused the result. Keep those two things separate. For a business owner, the consequence is worth considering anyway. Expertise becoming easier to access changes who can challenge an institution, and how carefully that institution has to explain its decisions. Find the constraint =================== Imagine a factory with one slow machine. Improving the machines on either side of it may give you a larger pile of unfinished work. The customer still waits for the slow machine. A services business has the same problem. The queue might sit in an inbox waiting for approval, in a specialist's review list, or with a customer who has not supplied a missing fact. This is why a demonstration that saves somebody twenty minutes can be commercially irrelevant. If that activity was not constraining delivery, the improvement might never reach the customer. Look at your chosen decision. Where does work accumulate? What is the weakest part of the customer's experience? Ask whether AI can improve that particular step, and what becomes the next constraint when it does. This gives you a much better starting point than buying a tool and looking for something to do with it. Three questions. Three kinds of evidence. ========================================= There are three questions that often get mixed together. Can the system do the task? Are people using it? And does using it improve the business? A benchmark can help answer the first question. A product's usage data can help answer the second. Neither automatically answers the third. For example, a high proportion of software-related conversations with an AI service tells us something about that service's users. It does not tell us the proportion of all programmers who use AI, or the number of jobs that have disappeared. When somebody presents an impressive number, ask what was counted and what the denominator was. Then ask which of these three questions the number actually answers. This simple habit prevents a surprising amount of confusion in board papers and sales presentations. Longer tasks are becoming tractable =================================== The chart measures task difficulty using the time a human expert takes. A time horizon is the duration at which an agent is predicted to succeed at a specified reliability level. The two series show fifty and eighty per cent success. The vertical intervals show estimation uncertainty. These are selected models from the published May 2026 data, rather than every release. The vertical scale is logarithmic, so equal distances represent multiplication, not equal additions of minutes. The trend is useful. The scope matters: these are mainly well-specified software, machine learning and cybersecurity tasks. The measure does not mean an agent can replace the same number of hours in any employee's day. For your business, the next step is a representative task with an outcome you can check. Compare useful completion, review time and failure cost against your current process. Reliability changes the decision ================================ A success rate has no business meaning until you know the consequence of failure. If I am generating possible names for a product, I can discard the bad ones. If I am changing a customer's account, an unnoticed error has a different cost. The higher reliability threshold in the previous chart corresponds to shorter tasks. You cannot use the fifty per cent figure and mentally treat it as a dependable working day. Nor should you assume that repeated attempts are independent. A system may make the same mistake each time. A second model can share the same blind spot. A practical evaluation includes the person who checks the work. Record their time, what errors they caught and what reached the next stage. The economic result is what remains after checking, rework and consequences, not the speed of the first draft. Software travels faster than machinery ====================================== A software improvement can be distributed very quickly. Installing equipment in a pharmacy or a factory requires space, integration, training, maintenance and somebody taking responsibility when it stops. That creates uneven adoption even when the underlying reasoning capability improves at the same rate. The physical installation may be the expensive part. It can also protect the value of good integration and operating knowledge. Do not turn this into a claim that physical businesses are immune. AI can change scheduling, quoting, design, inspection and customer acquisition before a robot performs the central physical task. For your business, separate the cost of intelligence from the cost of deployment. Which would still be difficult if the software became free tomorrow? That is often a more useful question than asking when a robot will do the whole job. Forecasting becomes a routine input =================================== Scott Alexander's article about AI superforecasters raises a useful question: what happens when forecasting becomes affordable for ordinary management decisions? The Forecasting Research Institute reported in July that several systems were statistically indistinguishable from its superforecaster reference. These systems can search, use tools and combine forecasts. The comparison extrapolates from human forecasts collected in 2024, and uncertainty intervals overlap. Treat this as evidence of parity on that benchmark. The management opportunity is to ask more explicit questions about the future. Will a supplier miss a deadline? Will a market reach a specified size? What external event would change an investment? A forecast is useful when it can affect an action. Start with a resolvable question and a deadline, then decide what probability would make you act differently. Keep a record so you can learn whether the forecasts helped. Prepare for faster, slower and uneven change ============================================ You do not need a single date for artificial general intelligence to make a business decision. You need to know how your decision behaves under plausible differences in progress and adoption. In a faster scenario, capable systems become cheap and customers change their behaviour quickly. In a slower scenario, reliability or implementation holds things back. In an uneven scenario, one part of your value chain changes while the rest stays awkward. These are scenarios, not assigned probabilities. Their purpose is to expose a fragile assumption. An investment that only works if nothing much changes deserves a different structure from one that pays under all three. Look for a staged commitment: something that buys information now while preserving a larger choice later. Also identify the cost of waiting. Caution can be expensive if a competitor secures the scarce resource first. AI is entering the process of building AI ========================================= One reason progress could accelerate is that AI tools can help with research and engineering for the next generation of AI systems. They can propose experiments, write code, examine results and help researchers explore more alternatives. That creates a possible feedback loop. It does not establish that the loop runs without limits. More proposed experiments are only useful if you can run them and evaluate them reliably. Compute, energy, physical infrastructure and access to good evidence can still constrain progress. The business consequence is to avoid treating the current pace as fixed. A plan should survive some acceleration and some disappointment. For your chosen decision, ask which external improvement would matter most. Better reasoning may be less important than lower cost, private deployment, dependable tool use or a system that works with the information you already have. What forecast would change your decision? ========================================= Here is a deliberately simple decision. You can spend twenty thousand dollars now to prevent a fifty-thousand-dollar loss if a particular event happens. If that is the complete set of consequences, the break-even probability is forty per cent. Move the probability and watch the decision change. Then change the cost or the loss. The point is to identify where better forecasting would matter. If you would take the same action at ten per cent and ninety per cent, refining the estimate may have little value. Real decisions can include partial protection, timing, cash constraints and opportunities you would give up. Add those before relying on the calculation. For practice, SignalStorm asks you to interpret signals and make a forecast before learning the outcome. Record your estimate before the reveal. Afterwards, explain what evidence should have changed it. When cognition gets cheap, what becomes scarce? =============================================== A business combines inputs to produce something customers value. If one input becomes much cheaper, demand may shift towards the things needed alongside it. Suppose analysis becomes plentiful. You may still need access to the customer, permission to act, reliable information, a physical installation and somebody responsible for the result. Which of those limits output? That is an argument about complements and constraints. It does not tell us that the owner of every scarce asset becomes richer. Customers might find a substitute. A competitor might capture the benefit. The price you pay for the asset might already assume a wonderful future. Write down one input in your business that might become cheaper, then the next input that would limit growth. The interesting commercial question is who can capture the value of relaxing that constraint. The parts that do not speed up ============================== A string quartet takes four musicians and roughly the same performance time that it used to. Other parts of the economy can become dramatically more productive while that particular activity barely changes. Baumol's cost disease describes how labour-intensive activities can become relatively more expensive as wages and opportunities change elsewhere. It helps explain why making some things cheaper does not make everything cheaper. For a business, the useful question is which activities remain difficult to accelerate. They might involve human attention, a physical process or a customer who values the experience itself. This is not a promise that wages or profits automatically rise in those activities. Customers still have budgets and alternatives. Use the idea to examine the cost structure and the next constraint, rather than to declare an occupation safe forever. You have 100,000 free workers ============================= Imagine that you and every competitor can call on a hundred thousand inexperienced white-collar workers for free. This is a thought experiment. It deliberately ignores some real costs so that we can see what else matters. What would you attempt? Could you examine every prospect, translate every document, produce many versions of a design or test ideas that were previously too expensive to explore? Now give the same resource to a competitor. What happens to your prices? Which part of your service becomes ordinary? Finally, put management back into the picture. Somebody has to define the work, provide context, check the output and act on the result. The labour may be free in the experiment, but coordination is not. Notice whether that becomes the new limit. What changes when labour is plentiful? ====================================== Work through the four questions with your chosen business. Start with what becomes possible, then what becomes affordable, then what can be personalised. Finish with what might stop earning a premium. A software business could produce more features, but customers may not want to manage more features. A manufacturer could produce more quotes, but still have one machine with a queue. A tourism business could personalise every itinerary, while the location itself remains unchanged. The answers should name a customer and an outcome. Producing more material is only valuable if somebody wants it or it improves a decision. If you cannot find an effect, try the same exercise from a supplier's or competitor's perspective. Difficulty imagining a change in five minutes does not establish that the business is protected from it. Scarcity is only the first question =================================== There are three different questions here. Is something scarce? Can its owner capture value from that scarcity? And is it a good investment at the price being asked? A location can remain unique while customer demand falls. A distribution relationship can matter while a platform takes a larger share of the margin. A well-positioned business can still be a poor purchase if the price assumes every optimistic scenario. For the decision you wrote down, separate these questions. Describe the scarce input. Explain how cash reaches its owner. Then test the assumptions embedded in the price and the commitments you must make before learning more. This is a way to organise your investigation, rather than a recommendation to buy a particular asset. It should help you find the next fact you need. Test the business, not the job title ==================================== Consider three businesses rather than three occupations. In software, cheaper development may reduce the value of producing code while increasing the importance of distribution and integration. It can also allow a small firm to build a product that previously required a large team. In a physical operation, better reasoning can improve planning while installation and operating reliability remain difficult. The economics depend on the whole system. For a place-based business, AI cannot duplicate the location. It can change how people discover it, what they expect and which alternatives they consider. Each example contains both opportunity and pressure. Avoid deciding that a whole sector wins or loses. Follow the money through the particular business: what customers pay for, which costs change, what remains difficult, and who has bargaining power. Stress-test the investment ========================== Take your investment through the three scenarios. For each one, describe customer demand, margins and the input that constrains growth. Do this in sentences before trying to make a detailed spreadsheet. Then ask how much capital must be committed before you learn whether the assumptions hold. Some decisions can be staged. Others put most of the money at risk before the first useful result arrives. A good next step often reduces the uncertainty that matters most. That might be a paid customer trial, a technical integration test, or evidence about whether a supplier can support the system at your scale. Use the workbook to record the assumptions and the next evidence. Put a date beside the review. A scenario exercise becomes useful when it changes the sequence of commitments, rather than merely producing three descriptions of the future. What would make you change your mind? ===================================== Choose one indicator you can actually observe. It might be a competitor's price, customer conversion, time spent checking output, or the cost of delivering a specific outcome. Then write the commitment you will make now and the condition for making the next one. Be specific enough that another person could tell whether the condition was met. Finally, state a result that would make you reconsider. If every possible result confirms your original plan, you have probably written a justification rather than a test. Keep this short. One useful indicator that somebody checks is better than a dashboard nobody owns. The purpose is to connect learning to a decision while there is still time to change it. Build what differentiates you ============================= A business contains components at different stages of maturity. Treating them all the same is expensive. Some activities are uncertain and distinctive. You may need to experiment and build knowledge because the market cannot yet supply what your customer needs. Other activities are standard enough that building them yourself buys little advantage. A Wardley-style map puts the customer's need near the top and the supporting components underneath. Across the page, components move from novel to commonplace. The map is a conversation about dependencies and maturity, not a precise measurement of value. The useful decision is at component level. You can buy a standard service and use it to support a distinctive customer experience. That is why a single slogan telling a whole company to build or buy is not enough. Different stages need different management ========================================== The management question changes as a component matures. At the beginning, you are asking whether it can work and whether anybody needs it. A detailed efficiency target may be premature. When the solution is custom, close contact with a specific user helps you discover what matters. When products are available, comparison, integration and repeatability become more useful. When something is a utility, dependable supply and cost control matter greatly. A single organisation may contain all four stages. That is why imposing one method everywhere can be frustrating. A research team and a payroll operation should not need the same kind of uncertainty. Place one AI-related component from your business on this progression. Then ask whether you are managing it as though it were at a different stage. The mismatch may explain the argument you keep having about it. Lower cost here. Add customer value there. ========================================== There are two useful moves on this map. One is to make a supporting component more standard, reliable and inexpensive. The other is to use what that makes possible to offer the customer something better or different. They can happen together. Buying a routine capability can release attention for a difficult customer problem. Standardising an internal process can make a personalised service affordable. Neither move guarantees a competitive advantage. A supplier may capture the savings, and competitors may copy the new offer. The map helps you state the mechanism clearly enough to examine it. Pick one component you would like to move towards routine supply, and one part of the customer experience where you would invest the capacity that becomes available. Explain how the second benefits from the first. Map one customer outcome ======================== Start with an outcome a customer pays for. Avoid beginning with your organisation chart or a list of software products. List the capabilities needed to deliver that outcome. For each one, ask how standard it has become and what depends on it. You can draw this with a pen; the conversation matters more than the diagramming tool. Look for two mistakes. One is spending heavily to recreate something readily available. The other is outsourcing the learning that makes your customer offer distinctive. Now connect the map to the investment decision. What would you buy, build, stop or test? If the answer is unclear, identify the piece of evidence that would help place a component. A rough map with an explicit uncertainty is more useful than a polished map that pretends everyone agrees. Match the supplier to the bet ============================= A supplier's funding announcement does not tell you whether its product will work in your business. Ask for evidence on a representative task with your constraints. Include the costs that tend to sit outside the demonstration: integration, permissions, checking, support and changing the workflow. Ask which assumptions make the quoted economics work. Then consider continuity. If the supplier changes direction, what can you export? Who understands the system? What rights and dependencies would you need to keep operating? A copy of the source code may help, but it does not automatically provide maintainers, hosting or access to another company's service. Match the commitment to the evidence. A short trial can justify a different level of uncertainty from a system your business will depend on for years. Exceptional capability can become ordinary ========================================== A capability can begin as something only a specialist can assemble, become available through a paid service, and later become cheap or open enough for many more people to use. Some capabilities eventually fit on local hardware. This is a pattern to watch, not a timetable. The lag varies with the task, model size, hardware, licensing and the effort needed to make the system useful. Availability also differs from adoption. A capable tool sitting on a website does not mean a business has the data, confidence or processes to use it well. Think about an advantage you currently get from privileged access to a tool. How long would that advantage last if customers and competitors gained similar access? The more durable advantage may be the context you provide, the integration you maintain or the responsibility you take for the result. Your customers and staff get these tools too ============================================ Return to the individual preparing a case. The same change appears in less formal settings. A customer can analyse a contract. A small supplier can prepare a detailed proposal. An employee can organise evidence and ask a much more precise question. That can improve accountability, but it can also produce confident errors at scale. Your organisation needs a way to respond to the quality of the evidence, rather than the polish of the document. Make important decisions explainable. Preserve the records that support them. Give people a route to correct mistakes. These are useful practices regardless of which model is currently fashionable. The strategic point is that AI adoption does not happen only inside your company. It changes the capabilities of the people with whom you deal, including people who have much less money than you do. AI can change who contributes expertise ======================================= The Cybernetic Teammate study examined seven hundred and ninety-one professionals at Procter and Gamble working on innovation tasks. In that experiment, individuals using AI produced solutions of comparable quality to unaided pairs. The tools also helped people contribute beyond their usual functional expertise. That makes the result interesting for how work is organised: some handoffs may exist because one person previously lacked access to a particular kind of knowledge. There is a second question, though. Generating an idea and identifying the best idea are different tasks. The study found difficulties with selection in the AI conditions. This was a one-day experiment in one company. It is not evidence that every team should be halved. Use it to ask which expertise barriers and handoffs are worth testing in your own work, while preserving responsibility for the final choice. Which handoff still earns its place? ==================================== Draw the path from a customer's need to a delivered result. At each handoff, ask why the work changes hands. Some handoffs provide independent challenge, legal authority or responsibility. Others exist because the previous person did not have access to the required expertise or software. AI may change the second group before it changes the first. Test a complete outcome rather than celebrating a faster draft. Does the customer receive the result sooner? Does quality hold? Does the reviewer spend longer correcting something that looks convincing? RoleShock lets you explore the tension between rapid productivity changes, cash pressure and workforce capability. Play the short version, then ask which skills people need to develop before a transition. The debrief matters more than the score. Will staff tell you what they have automated? ============================================= Suppose an employee finds a way to do a day's work in an hour. What happens when they tell you? If the obvious consequence is that their position disappears, keeping the discovery private may be a sensible response. You then lose the chance to improve the process, understand the risks or spread the learning. A useful management response is specific. Make time available to document the improvement, check its reliability and work on the next valuable task. Be clear about responsibility and about what decisions you have not made. Do not promise permanent job security if you cannot deliver it. Make a narrower promise you can keep: that reporting a useful improvement will lead to a fair conversation, that checking costs count, and that people will have time to learn. The incentives determine what information reaches management. One strategic choice. One reversible test. ========================================== Separate the strategic choice from the experiment. The strategic choice concerns the advantage you want to build over several years. The experiment concerns an uncertainty you can reduce soon. For example, you might decide that integration with a customer's operations matters more than producing another standalone feature. A short experiment could test whether a specific integration improves adoption or reduces the customer's work. The experiment should be small enough to reverse, but real enough to teach you something. A demonstration using ideal inputs may answer the wrong question. Write one choice and one test. If you have ten of each, select the pair most likely to affect the investment decision you brought into the course. Give the test an owner and a review date. State the result that changes the decision ========================================== Before you begin the test, record the current process. Otherwise, it is remarkably easy to compare the new system with a vague and unusually pessimistic memory of the old one. Measure the whole outcome. Include checking and rework, and follow the result far enough to see whether a customer or colleague benefited. Decide how you will handle an error before you expose anyone to it. Write the decision rule in advance. What result would justify expanding the trial? What would lead you to change the approach or stop? You can revise the rule when new information warrants it, but record why. Finally, give one person responsibility for reviewing the evidence on a specific date. A pilot without a decision can become a permanent expense with a cheerful name. What will you do differently on Monday? ======================================= Return to the decision you chose at the beginning. What has changed in your understanding of it? Perhaps you now think the scarce resource is somewhere else in the business. Perhaps you have found an assumption that deserves a forecast. Perhaps the next step is a small test before a larger commitment. Write down what you will do differently on Monday. Name the person who will act, the evidence they will collect and the date you will review it. The future will supply plenty of surprises. You do not need to predict every one of them. You need a way to notice the changes that matter to your business and enough room to respond before an avoidable commitment becomes an expensive lesson.