Essay · AI, abundance & meaningful work

AI Made Building Easy.
What Is Worth Maintaining?

Building is getting cheaper. Choosing what is worth maintaining is the harder question.

8–9 min read

Building takes an afternoon. Maintaining it takes ongoing work.
Building itAn afternoon
Maintaining itOngoing

An illustration of the idea, not measured data. Bar lengths are not to scale.

With AI, a new tool, workflow, or system can take shape in an afternoon. It might have a polished interface, a useful process behind it, or both. A demo makes everyone lean forward. Someone says that a few months ago, this would have taken weeks.

They are probably right.

October 2026 has made that feeling harder to ignore. Tools such as Reverse Engineer Anything let AI agents inspect existing software, explain how its features work, and help recreate them. Building something new is getting easier, and so is taking apart something that already exists.

At the same time, a fall in tech stocks towards the end of October brought the AI bubble question back into view. Investors were questioning the expectations built into AI valuations. The possibilities keep expanding, while confidence in what they will be worth is less settled.

A week's market movements cannot settle what AI will become. But they make the question of lasting value feel immediate. In my own work, that question comes down to whether something will keep helping people after the excitement of building it has passed.

I help build and run a platform used by teams across the organisation to create AI agents. A new agent can now take hours rather than weeks. Keeping one useful is a different job. Someone has to check that it gives accurate answers or takes reliable actions, watch what it costs to run, and revisit it when the tools underneath it change. The first part got faster. The second part is still a person's time.

The same pattern appears when a tool, workflow, or system enters ordinary life. Someone needs to check what it produces, keep its information current, and explain how it fits with the ways people already work.

Soon, there is another place to look, another password to remember, and another thing to keep running.

The building got easier. The work did not shrink to match.

This is one way to think about the paradox of abundance: having more of something useful does not automatically mean getting more of what we need.

AI gives this old question a new setting. More people can try ideas, write software, and tackle problems that once felt out of reach, and that is worth celebrating. It also leaves us with a question no demo can answer on its own.

Once we can build almost anything we can describe, how do we choose what deserves our time?

I think the answer starts with a different question. Not "what can we build?" but "what are we prepared to maintain?"

Having more is only the beginning

There is an idea in economics called the resource curse, also known as the paradox of plenty.1 Some countries rich in oil or minerals have struggled to turn that wealth into better lives for their people. The reasons are debated, and much depends on policy and institutions. A country and a software team are very different things, so take this as an analogy, not proof. But it carries a useful lesson: a valuable resource does not look after everything around it.

For AI, access to a capable tool is a starting point. Someone still needs to understand the problem, decide what matters, and take responsibility for the result. The opportunity is enormous, but using it well takes more than having it.

Work that continues after launch

Imagine a team creating a system to help with workplace policies. It could answer a question directly, or check a request in the background and route unusual cases to a specialist. Either approach could save people time.

This is an imaginary example. To keep the system useful, someone still needs to update policies, manage access, review results, and handle exceptions.

Whether a system produces an answer, a decision, or a completed task, its visible result is only part of the work.

What the studies actually tell us

Most of the careful research so far comes from software development, where AI tools have been used and studied the longest. It is not a perfect match for every workplace, but it shows what to watch for: a gain in one step does not always carry through to the whole job.

The puzzle is older than AI. In 1987, the economist Robert Solow put it this way:

"You can see the computer age everywhere but in the productivity statistics."

Robert Solow, 19872

New technology often looks transformative up close and modest in the totals. The same question now applies to AI.

Up close, the gains can be real. In three experiments with software developers, involving researchers from MIT, those given an AI coding assistant completed about 26% more tasks, and less experienced developers tended to gain most.3 A large study of customer-support agents, by researchers at Stanford and MIT, found a similar pattern: AI help raised productivity, most of all for newer workers.4

Then there is the opposite result. In a small, carefully controlled experiment in early 2025, the research group METR found that experienced developers working on projects they knew well took 19% longer with AI tools.5 A 2026 follow-up said newer tools probably help more, but it could not give a reliable figure, because many developers who wanted to keep using AI chose not to take part.6 So there is no honest single number for how much AI helps everyone.

Three tests, three different answers

Estimated change in task speed with AI

Point estimates from METR's task-completion experiments; confidence ranges are not plotted. The late-2025 estimates could also reflect a small slowdown, and METR says the follow-up is an unreliable signal of current effects.

Estimated change in developer task speed when AI was allowed. Positive means faster, negative means slower.
StudyParticipantsResult
Early 2025Experienced developers on familiar projects19% slower
Late 2025Returning developers18% faster, could go either way
Late 2025Newly recruited developers4% faster, could go either way

Best estimates only. For the last two, the data is also consistent with a small slowdown, and METR says they probably understate AI's benefit, because developers who most wanted AI chose not to take part.

Zoom out to whole organisations, and the picture gets more mixed. Google's 2025 DORA report, based on a survey of nearly 5,000 technology professionals, described AI as an amplifier. It magnifies the strengths of teams that already work well and the weaknesses of those that don't. Delivery is now faster with AI, but it is also less stable.7

The most useful evidence for this essay comes from a 2026 study by the economists Mert Demirer, Leon Musolff, and Liyuan Yang.8 They followed more than 100,000 software developers through three generations of AI tools. The number of saved changes to code ("commits") roughly tripled. But the gains shrank at every later step: about 50% more projects, and only about 30% more versions actually released to users. Reviewing, testing, and fitting changes together still limited what reached people.

The wider the view, the harder the gain is to find

Four ways of looking at the same question.

Lens Source Verdict Finding
Up close Controlled experiments Mixed More got done in some studies, less in one.
Whole organisations Google's DORA survey Mixed Faster delivery, but less stable.
From code to release Demirer, Musolff and Yang Shrinks Almost 3× the code, but only 30% more releases.
The market App stores, same study No gain yet More new apps, but no more use.

The studies measure different things, so this is a map of the evidence, not a ranking.

The temptation to make another version

AI can help people understand unfamiliar software, explore how something works, or recreate a tool for a different need. It can help recover knowledge that was never written down properly. Those are worthwhile possibilities.

It can also make starting again feel easier than understanding what is already there.

Imagine replacing an old reporting tool. The new version looks better and handles the usual examples. Then someone discovers that the old tool also dealt with awkward records, unusual permissions, and exceptions that had built up over years.

The team now has to work out why those decisions existed. The easy rebuild has become a longer job of rediscovering what the original system knew.

Give experiments room to finish

There is value in building something just to learn. A prototype might answer a question, test a hunch, or show that an idea is less useful than it first seemed. Learning that early can save a great deal of effort later.

An experiment can have a clear question and a point where it stops. A service that people rely on needs an owner and a plan for keeping it useful. The trouble comes when the first quietly turns into the second.

A demo gets shared. Someone starts using it. More people follow. Eventually, an afternoon project has become a responsibility, even though nobody decided who would carry it.

A quick experiment can grow into something people rely on.

  1. A demo gets shared
  2. A few people use it
  3. More people rely on it
  4. It needs an owner

People may start depending on it before anyone agrees to own it.

Calling something an experiment should make it easier to be honest about the result. Sometimes the result is worth developing. Sometimes it tells us enough to leave the idea there.

A useful lesson is a perfectly good outcome.

What do we get back for the effort?

Return on investment can sound like a demand for immediate profit, but the value of a tool can be much broader. It might prevent mistakes, help someone access information, make a difficult task possible, or reduce a repetitive burden. The value of an experiment might be knowing what to try next.

Whatever the goal, say it clearly: what is difficult today, what should become easier, and how will we know?

Consider a simple example. A tool saves a team ten hours each week. Maintaining it takes eight hours. The time returned is two hours. That can still be a good result; it just isn't ten hours saved.

How much time is left after maintaining the tool?

A tool saves 10 hours each week

Compare those 10 hours with the time spent maintaining the tool. Whatever remains is time returned.

Each bar compares time spent maintaining the tool with 10 hours saved. In the last example, all 10 saved hours are used and maintaining it takes 2 more hours. This is simple arithmetic, not a financial return.

Weekly time returned after a tool saves 10 hours and the team spends time maintaining it.
Time spent maintaining the toolTime left from the 10-hour saving
2 hours8 hours left
8 hours2 hours left
12 hours2 hours more work than time saved

If maintaining the tool takes twelve hours, it takes two more hours than the tool saves. Another benefit might make that worthwhile, such as preventing important mistakes or making a service available to more people. Whatever the benefit is, name it.

There is another possibility too. The tool helps people produce more, but their workload stays the same. If the aim was to clear a backlog, that might be success. If the promise was a calmer working day, there is more to understand.

The measure should match the promise.

What would enough look like?

I would put it this way: a tool is enough when the value it gives back is worth the work of maintaining it, and someone has agreed to do that work.

Return to the policy-support system. Imagine it links each answer or action to the current policy source, so a person can check it quickly. It shows when each source was last reviewed. It has a named owner and a quarterly review where the team asks whether it is still worth keeping. It will not be the most exciting thing in the next presentation. It simply makes an ordinary Tuesday better, and it will still be trustworthy next year.

Before building something new, it helps to ask:

  • Who needs this, and what gets easier for them?
  • Who will own it six months from now?
  • What does it replace, and what did the old thing quietly handle?
  • How will we know it worked, and when will we stop?

Enough can be small: a change that removes one confusing step, or a tool that does a dull job reliably and then stays out of the way.

This is a hopeful way to use abundance. More people can try ideas. More problems are now cheap enough to tackle. Useful experiments get a chance.

It also means being careful about what we ask others to depend on. We can improve something that already works. We can close an experiment once it has answered its question. And we can spend the time we save on checking quality, teaching someone, or finishing work at a reasonable hour.

There will always be another thing to build. Enough means picking the one that helps, understanding what it will take to maintain, and then doing that work.

We can build more than before.

The opportunity is to make more of it worth having.

Sources, study limitations & further reading

Sources and scope

Research links checked on 30 October 2026. The Reverse Engineer Anything link points to the project’s GitHub repository, checked on 31 October 2026. Market movements provide context for the question of lasting value; they do not measure the usefulness of individual AI tools. The resource curse is used as an analogy.

  1. IMF, background paper on resource-rich developing countries. For a short introduction, see Wikipedia: Resource curse.
  2. Robert Solow, "We'd Better Watch Out," New York Times Book Review, 12 July 1987. Quoted in the VoxEU column in note 8.
  3. Cui, Demirer, Jaffe, Musolff, Peng, and Salz, The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers, Management Science (2026), published online in February 2026. The earlier draft covered 4,867 developers. Two of its authors also wrote the study in note 8, so the two are related rather than independent. The task-completion result describes the work measured, not general business returns.
  4. Brynjolfsson, Li, and Raymond, "Generative AI at Work", Quarterly Journal of Economics 140(2), 2025.
  5. METR, The early 2025 developer productivity experiment, July 2025. A randomized study of 16 experienced developers across 246 tasks.
  6. METR, The February 2026 update. The chart draws METR's reported estimates on one scale. The ranges METR reports around them are 2% to 39% longer (early 2025), 38% faster to 9% longer (late 2025, returning developers) and 15% faster to 9% longer (late 2025, new developers). METR describes the late-2025 data as an unreliable signal of the current effect.
  7. Google Cloud, 2025 DORA Report: State of AI-assisted Software Development, September 2025. A survey-based study, so it shows associations rather than proven causes.
  8. Demirer, Musolff, and Yang, Writing Code vs. Shipping Code, NBER Working Paper 35275 (2026); summarised in a VoxEU column, 21 June 2026. Observational evidence; app usage is measured through proxies such as ratings.