Leverage in the AI Era: Better Outcomes, Not More Output

I bought four pumpkin cake squares on clearance for $1.75. It's Fall, right? We'll give it a shot. The first went in the trash after two bites.

That's the whole incident. A clearance rack, a dessert that had sat out a day too long, and a texture that made me quit two bites in. I had three squares left and a choice: eat them as they were, throw them out, or bake something from scratch. Being an obsessive agency owner, I started the calculations... What if we just added a crust and more depth? I toasted a square in butter until the edges crisped, then hit it with cinnamon, cloves and a pinch of salt. The spice woke it up, and the salt covered for whatever the bakery had skipped. By my own standard, it went from "throw this away" to "I'd serve this to someone."

I wasn't trying to make the best pumpkin dessert possible. I was trying to make this one substantially better without turning dessert into a project. 5 minutes, some creative thinking, and I was biting into a whole different league of baked good.

What matters in that story is what I didn't do, and I make the same call on a product and on biz ops. I didn't start over. I found the part that was letting the whole thing down, texture and seasoning, a step or a handoff in the work, and I changed that part. The rest was fine, so I left it alone.

No AI was involved in that pan. The habit the rest of this article is about, deciding what to build, what to change, and what to leave, existed long before I had a model to talk to. AI changes the economics of acting on it. It doesn't supply the judgment.

TL;DR

• Leverage starts before you pick a tool. It starts with noticing a gap between what something delivers and what a sensible change could make it deliver.

• AI lowers the cost of investigating, testing and repeating an improvement, which changes which improvements are worth attempting. It doesn't decide which ones matter.

• Faster output and more valuable output are different things. Count the full cost, including review and maintenance, and decide in advance what would make you stop.

• The enterprise surveys say the same thing at scale. Most corporate AI spending shows no measurable return, and the money tends to go where results are easy to present rather than where the leverage is.

Why this is leverage and not perfectionism

Leverage, in plain terms, means getting an outsized result from what you put in.

It has two neighbors it gets confused with. Efficiency is about how economically a task gets done once you've decided to do it. Perfectionism keeps asking for one more improvement regardless of what the improvement costs. Leverage asks a question that comes before either: is this the right intervention at all, and does the benefit justify what I'd spend to get it?

Noticing the gap is the skill. The cake had a gap between what it was and what a small change could make it. Software has the same gaps, usually wider ones.

The part people leave out is the stopping rule. I pause work on Sonor features all the time, sometimes mid-thought, because another task has a better claim on the afternoon. That isn't a lapse in standards. Knowing when further effort stops earning its keep is part of the method, and I'd argue it's the harder half.

It doesn't take an unusual temperament, either. It's a habit, and habits transfer. Look for the gap, size the intervention against it, and stop when the two stop matching.

What AI changes about the math

AI doesn't change the question. It changes the price of answering it, in three places.

Investigating the opportunity. Before I commit to a change, I want to know what the alternatives are, which of my assumptions are shaky, and what I'd need to check before I trust the plan. A model is good at that kind of conversation. It'll surface options I hadn't considered and poke at the ones I had. The useful output of that step is a better-informed decision. A longer list of ideas is only a longer list.

Testing the intervention. This is where the economics move most. A prototype, a draft of a new workflow, a script that reshapes a messy export into something I can look at: these used to cost enough that I'd skip the test and go on instinct. With AI-assisted work, an experiment that would've taken a day can sometimes fit in an hour. I say sometimes on purpose. Some experiments get cheaper and some don't, and you find out which by running them.

Implementing and repeating what works. Once a change has proven itself, parts of it can harden into reusable software, a documented process, or a workflow someone else can run. AI helps with the typing, the refactoring, the documentation. The repeat benefit comes from making the thing reusable in the first place, and that's a design decision. The model speeds up the keystrokes.

Together, those three change the shape of the decision. AI's value isn't only that it makes existing work faster. It changes which improvements are worth attempting at all. The experiment I would've skipped last year gets run. The small fix that wasn't worth the context switch gets made.

There's decent evidence that the gains are real, and also that they're uneven. Generative AI at Work, published in the Quarterly Journal of Economics in 2025, followed 5,172 customer-support agents through the staggered rollout of an AI assistant. Agents with access resolved about 15% more issues per hour on average. That's a meaningful gain in one specific setting, and I'd read it as exactly that rather than as a multiplier you can apply to your own business.

The same study undercuts a story I hear a lot, which is that AI mostly rewards the people who were already the best. The researchers found the opposite. Less experienced and lower-skilled agents improved both speed and quality, while the most experienced saw small gains in speed and small declines in quality. If you're assuming AI will widen your lead because you're good, the evidence says check that.

From a better task to a better operating system

Everything so far has been about single tasks- the business version of the habit is way, way bigger.

There's a difference between doing a recurring task faster and changing the arrangement so the task doesn't recur. The first is efficiency, and it's worth having. The second is where leverage compounds, because every future repetition gets the benefit without anyone doing the work again.

I should disclose that I founded Sonor and I run Upforge, so take this example as firsthand rather than neutral. For years, every website we built needed the same infrastructure: forms that reach a CRM, analytics, metadata that search engines and AI crawlers can read, a way to publish articles, sometimes a store. We rebuilt variations of it per project. The intervention was to stop rebuilding. Sonor became the shared layer: a Next.js site connects with one API key and gets forms, analytics, SEO, chat and agent tools from a shared package, with publications and commerce as add-on kits. The developer documentation lists what the platform provides. It doesn't claim how many hours it saves for users, and I won't claim that here either, because I don't have a clean before-and-after record to point at. For Upforge, though? We're cutting development time in half, and passing those savings directly to our clients.

The distinction worth holding onto: reusable infrastructure supplies the repeat benefit. AI helps build it, adapt it and maintain it. Those are related advantages and they're easy to conflate, but one is about the shape of the system and the other is about the cost of working on it.

Reuse has its own bill. Shared infrastructure takes real upfront investment. It creates a maintenance obligation that doesn't go away, and a mistake in a shared layer ships to every site that depends on it at once. I've paid that bill more than once. Reuse earns its keep when the repeated benefit clears those costs with room to spare, and not before.

The rule I use: build custom where the difference creates value. Reuse where rebuilding creates none.

And the useful business outcome at the end of this doesn't have to look like more content or fewer people. It can be a quote that goes out the same day, an invoice that doesn't need a second pass, capacity to take a client you'd have turned away, or a service that was too expensive to offer at all until now.

Why most corporate AI spending shows no return

The expensive version of skipping that habit played out across 2025 at enterprise scale, and the numbers are worth sitting with, because they come from the firms whose business is advising the companies that spent the money.

McKinsey's State of AI survey, published in August 2026, found that eight in ten respondents say AI has improved their own productivity, while 37% attribute any effect at all to their organization's earnings, a share that hasn't moved in a year. The high performers, who credit AI with at least 5% of EBIT, are 6% of the sample. BCG's Widening AI Value Gap, published in September 2025, put 5% of companies in the group generating value at scale and 60% in the group "reaping hardly any material value, reporting minimal revenue and cost gains despite substantial investment." S&P Global's survey of 1,006 IT and business leaders found the share of companies abandoning most of their AI initiatives before production had jumped from 17% to 42% in a year, with the average organization scrapping 46% of its proofs of concept. The number that did the rounds in boardrooms came from an MIT NANDA report in August 2025, covered by Fortune: 95% of organizations with no measurable profit-and-loss return from their generative AI pilots, against an estimated $30 to $40 billion in enterprise spending. That report's method drew criticism and MIT no longer hosts it, so take the 95% as the loud version of what the larger surveys say more carefully.

What interests me isn't the failure rate. It's where the money went. The MIT researchers found more than half of generative AI budgets going to sales and marketing tools, where demo counts and response times fit neatly in a board deck, while the clearest returns showed up in back-office automation that fewer companies were funding. The report's lead author told Fortune that the ones getting results "pick one pain point, execute well." Nearly three-quarters of McKinsey's high performers say they've fundamentally redesigned workflows around AI rather than inserting it into the old ones. A quarter of everyone else has. S&P found the companies with the lowest abandonment rates were the ones with a more deliberate way of choosing which projects to run at all.

Read those together and it's the pumpkin cake at a billion-dollar scale. The companies that got nothing back weren't short on models or budget. They skipped the part where you find the gap before you buy the tool, and they measured activity instead of outcomes. The five questions that follow are the version of that step you can run on a Tuesday afternoon, before the spend. They're also where every AI and automation engagement at Upforge starts.

Disproportionate outcome analysis: five questions before you automate anything

Disproportionate outcome analysis is my working phrase for this, and I'm not pretending it's a discipline. It's five questions, asked in order, because most bad automation decisions skip the first two.

1. What outcome actually matters? Name the result, not the activity. A customer who got a usable answer is a different thing from a response that got generated.

2. What's holding that outcome back? Find the constraint before improving whatever happens to be easiest to improve.

3. What's the smallest credible intervention? Compare improving what you have, buying something, building something, delegating it, or leaving it alone.

4. What does it really cost? Setup, review time, maintenance, money, attention, and what happens if it's wrong.

5. What would make me continue, standardize, or stop? Decide in advance what counts as success and when further work loses its priority.

A worked example, and I'll flag that it's hypothetical. Say an automation takes two hours to set up and saves 15 minutes each time you use it, after you've counted the time to review its output. You need eight uses to get your two hours back, before any maintenance, subscription cost, or the occasional afternoon it breaks. For a task you'll do twice, that's a bad trade and the honest answer is to do it by hand. For a task you'll do weekly for the next two years, the math flips hard.

The lesson is to estimate the payoff before you start. "Automate recurring work" is a slogan. The eight-use breakeven is a decision.

Don't turn the questions into a scoring system with weights and a total, though. The questions are there to sharpen judgment. Uncertainty doesn't go away because you put it in a spreadsheet.

The speed-vs-value trap: cheap output can still be expensive

There's a difference between doing the same valuable work more economically and doing additional low-value work because it got easy. A business doesn't automatically need more pages, more dashboards, more proposals or more features. It needs the specific work that moves its outcomes, and that set doesn't grow because production got cheaper.

METR looked at this directly in a survey it published in May 2026. The researchers asked 349 technical workers about their gains from AI and deliberately separated two measures: speed, meaning how much faster tasks got done, and value, meaning how much more worthwhile output they produced. The median self-reported speed gain was 3x. The median self-reported value gain was between 1.4x and 2x. METR's explanation for the gap is the one I'd give. Speed numbers get inflated when people take on extra tasks that AI does quickly but that wouldn't otherwise have been worth prioritizing. Their example was a researcher who builds a polished interactive dashboard the project didn't need. These are self-reports, and METR says plainly that they aren't measured productivity. The shape of the gap is still the point. By the respondents' own accounting, a large share of the speedup wasn't showing up as value.

The other half of the trap is quality. AI help is uneven in a way that's hard to see from inside a task. Navigating the Jagged Technological Frontier, a preregistered field experiment with 758 consultants at Boston Consulting Group, published in Organization Science, found that on 18 realistic tasks the people using GPT-4 completed about 12% more work, about 25% faster, at higher quality. On one task the researchers chose because it sat outside what the model could do well, the AI-assisted group was 19% less likely to get the right answer. Models have moved a long way since that experiment ran, so don't read it as a verdict on today's tools. Read it as a reason to test and verify rather than assume, because the line between where AI helps and where it quietly hurts keeps moving, and you won't feel it move.

Which brings me to the sentence I'd put on the wall: an improvement isn't leverage if its savings disappear into checking, fixing or maintaining it.

Selective ambition

Back to the cake. Making it from scratch might well have produced a better dessert. That doesn't make it the better decision, because the decision has to account for the evening it would've consumed and what else that evening was for. Some problems run the other way. They deserve a substantial custom investment because their stakes, or the number of times they'll pay you back, justify it. Sonor was one of those for me. Most things aren't.

High standards and selective effort aren't in tension. The standard is about the outcome. The selectivity is about where the effort goes to get it.

I don't ask how much work AI lets me produce. I ask which outcomes became worth pursuing now that some of the work costs less.

Sometimes that means building a platform. Sometimes it means fixing one step in a process.

Sometimes it means putting butter in a pan and leaving the rest alone.

If you want something to do with this today, pick one recurring frustration in your business. Write down the outcome you want from it, in a sentence. Then find the smallest change that would credibly move that outcome, and try it before you buy or build anything bigger. If the change turns out to be software and you're about to put real customers in front of it, I wrote about how to check AI-built work before it goes live.

And if you'd rather run the five questions on your operation with someone who does it for a living, that's the work we do at Upforge: how we apply leverage in the AI era to a business.