Vol. III · Essays on work after AI
Saturday, October 3, 2026
evolving.work
Strategy · Org Design · Essay № 026

The Iteration Gap

AI's biggest gift to physical businesses isn't automation — it's the ability to iterate a thousand times digitally before committing once in the real world. The ones who figure this out first will outcompete the ones who don't.

H Héctor Reinoso May 19, 2026 13 minute read Share · Save

Eighty-two percent of the smallest businesses — the ones with fewer than five employees — say AI is not applicable to what they do. Restaurants. Roofers. Manufacturers. The people who build, fix, cook, and install things with their hands.

I understand the instinct. A conversation about models and prompts can seem far removed from a day spent lifting, measuring, cooking, or installing. The connection becomes clearer when we look at the decisions surrounding that work.

There are uses for AI beyond automating the physical work itself. AI’s biggest gift to physical businesses is not automation. It is iteration: the ability to test a hundred ideas digitally before committing a single dollar of materials, crew time, or reputation. And because physical businesses pay the steepest price for wrong guesses — wasted materials, lost crew hours, spoiled inventory, unhappy customers — they have the most to gain from getting those guesses right first.

Andrej Karpathy, one of the minds behind the AI systems everyone is debating, recently made this tangible. His autoresearch script ran 700 autonomous experiments in two days, discovering 20 improvements that transferred to larger systems. The principle underneath is simple: any problem with a verifiable metric can be iterated at machine speed. And it does not stop at software. Menu margins, bid conversion rates, production throughput, waste percentages, cost per unit — all verifiable, all iteratable at speeds no human team could match. Physical businesses have more of these metrics than they realize. They just never had a cheap way to test against them.

The opportunity is to compare more options before spending materials, crew time, or reputation on a choice.

Three Hours, One Menu, Zero Wasted Plates

Justin Trainor needed to rework the menu at Hardwood Cafe. Anyone who has run a restaurant knows what that normally looks like: weeks of napkin math on food costs, trial-and-error with descriptions, and customers quietly absorbing the cost of bad guesses while you figure out which items land.

There is a limit to digital testing: you cannot taste a dish digitally, or simulate the warmth of a dining room or the way a plate catches the light under your fixtures. Those parts still need physical trials.

A menu rework involves hundreds of decisions around the food itself, rather than one decision: what to name the dish, how to describe it, where to place it, what to charge, which underperformers to cut, what the margin looks like at each price point. Those descriptions, placements, and prices are information, so you can compare them for practically nothing before asking the kitchen to change a dish.

Trainor used Toast IQ, an AI built into his POS system, to work through the entire menu digitally. It analyzed sales data, suggested descriptions, flagged underperformers, and recommended pricing — all before the kitchen changed a single dish. The rework took three hours.

Almost every decision has a concept layer and an execution layer. The concept layer — what to serve, how to describe it, what to charge — is pure information, iterable a thousand times for nearly zero cost. The execution layer — cooking, training staff, plating — still requires hands, heat, and a kitchen. The restaurant cannot taste-test a thousand dishes. But it can test a thousand menu configurations and commit kitchen time only to the top five. The iteration happens digitally; the kitchen executes the winners.

The Deloitte data shows where this is going: 73% of restaurant executives expect AI investment to increase next year, yet menu and new-product iteration sit under 50% adoption. The money is going to chatbots and inventory management — the operational plumbing. McKinsey’s restaurant work points to the same gap: the operators pulling ahead use transaction and footfall data to refine offers before rolling out changes chain-wide, not after a dish has bombed in three locations. The creative engine — deciding what to serve and how to position it — still runs on gut instinct. The iteration layer restaurants need most is the one they have not built yet.

If a restaurant can collapse weeks of experimentation into hours, what happens in an industry where a single wrong estimate costs thousands?

Forty Seconds to a Roof Quote

A roofing contractor producing one estimate the traditional way spends 30 to 90 minutes per quote. Site visits. Manual aerial photo analysis. Material takeoffs. Labor calculations. A proposal formatted to look professional enough to close the deal. With that cycle time, you pick which jobs to bid on. You have to. The bottleneck is not your crew’s skill or your installation quality. It is the paperwork around it.

Roofle’s AI platform changed that equation. Using aerial imagery to generate roof blueprints, live material pricing, and automated labor rates, it produces a customer-ready proposal in about 40 seconds. Contractors on the platform reported bidding on two to three times more jobs — not because they became better roofers, but because they got faster at everything around roofing.

In iteration terms: 135 quotes in the time it used to take to produce one. A 135-to-1 ratio.

And quoting is only the most obvious target. The same logic runs through marketing. A painter who once papered whole neighborhoods with flyers can now test which audience converts — homeowners who recently moved, or households that have just started browsing listings — before printing anything. Here, the concept layer is the choice of audience. Comparing audiences is cheap; printing and distributing flyers to the wrong households is expensive.

Field technicians spend 30% of their working hours on non-billable administrative tasks — more than the 28% they spend on actual service delivery. More time on paperwork than on the work itself. The iteration opportunity is not in the craft. It is in everything surrounding it — quoting, estimating, scheduling, follow-ups — all concept-layer work hiding behind a physical label.

It is not just a roofing problem. The 2026 Simply Business Growth Gap Report found 80% of small business owners still handle administrative and finance tasks themselves, and 54% name lack of time as the primary barrier to expansion. The bottleneck is not market demand or capital. It is the owner’s calendar.

The window is wide open. Fewer than 30% of roofing companies use AI in any capacity. The rest compete at 1:1 — one manual quote at a time — while early adopters work at 135:1. Not because the craft is different. Because the iteration speed around the craft is different.

And no, most roofers are not going to learn prompt engineering — but they did not learn to code an Instagram page either, and most use it for marketing. Vertical tools that speak roofing, not machine learning, are following the same accessibility curve.

Click Undo vs. Call the Crane

Lagor makes ferromagnetic cores for power transformers in Italy. Their products weigh up to eight tons. This is about as far from a tech company as a factory floor gets.

Their production line used to bottleneck unpredictably. Cores would pile up at transition points, blocking the entire sequence. The most practical way to clear a major blockage — and I love this detail — was to unload the cores by crane and reset the whole line. Every production run was, in effect, a live experiment with eight-ton stakes.

Their engineers built a digital twin of the production floor. Now they rearrange sequences in simulation — testing dozens of configurations, spotting bottlenecks, optimizing flow — before committing any physical resources. The loop shifted from “run the line, hit a bottleneck, call the crane” to “simulate the sequence, find the problem, fix it, then run the line.”

The cost of being wrong dropped from crane intervention to clicking undo.

Scale this up and you reach Unilever, which runs eight digital twins across factories on four continents. Daily alerts dropped from 3,000 to 300 — a 90% reduction — and the company reports $52 million saved annually as the algorithms grew accurate enough to control machine operations directly. McKinsey makes the principle explicit: the value of a digital twin is not the visual replica but the ability to run repeated experiments on layouts, demand shocks, and maintenance policies before anyone touches the floor. Early adopters in manufacturing see roughly 15% cost reduction and 25-plus percent efficiency gains in their first year.

I wrote about how cognitive phases collapse while physical reality remains — this is that principle at work. The thinking, the sequencing, the optimizing iterates in the digital layer. The physical layer executes the winners.

Making more time for the craft

AI cannot iterate the warmth of a family-owned dining room. It cannot replicate the judgment a veteran roofer brings when something about a particular roof just does not look right. It cannot simulate the trust a contractor builds by showing up on time, every time, for fifteen years.

Tacit craft quality — the accumulated knowledge that makes a customer come back and tell their friends — is not reducible to a metric. And the research confirms what operators already feel: a Harvard Business School study of Kenyan entrepreneurs using an AI assistant found that already-strong operators raised profits and revenues by 10-15%, while weaker operators saw theirs drop by roughly 8%. The same tool helped the stronger operators and hurt the weaker ones. Having suggestions to compare still leaves the operator responsible for choosing which fit the business.

AI does not iterate your craft. It iterates everything around your craft so you can spend more time on what only you can do. The roofer who saves 90 minutes per estimate does not become a worse roofer — they gain 90 minutes for the relationship-building and on-site judgment no algorithm touches. The restaurateur who collapses a three-week menu rework into three hours gains back three weeks for the hospitality that makes regulars feel at home.

The competitive advantage is how quickly you can test alternatives and free your hands for the work that sets you apart.

What happens when every competitor adopts the same tools? If the tools are the advantage and tools are commodities, the advantage is temporary. MIT Sloan made this case directly — once AI use becomes ubiquitous, it lifts entire markets but stops uniquely benefitting any single firm.

At the tool level, exactly right. But businesses that build the habit of relentless experimentation develop a muscle that persists long after the tools become universal. Every company got spreadsheets; the ones that built data-driven cultures still outrun the ones using Excel for invoices only. The iteration gap is not about having the tool — it is about building the practice of using it. And with 82% of small businesses not even in the game, that window is as wide as it has ever been.

Finding tasks to test digitally

The economics already shifted underneath this conversation. AI inference costs dropped 280-fold in 18 months, and startups now reach product-market fit with 60% less capital than last year’s cohort because the cost of experimentation collapsed. A demo that cost five dollars per user in 2023 runs under fifty cents today; 500 variations of your marketing copy cost less than printing one flyer. The decision quietly inverted: no longer “can we afford to use AI here?” but “can we afford not to iterate?”

That matters more for small operators than for funded startups. Simply Business found 45% of small business owners hold less than three months of operational cash reserve. When your buffer is that thin, every wrong guess in the real world is existential. Digital iteration is not fashionable — it is the only experimentation a thin-margin business can actually afford.

How to find your own iteration gap this week:

1. Map your iteration tax. List every decision from last month that involved guessing, testing, or adjusting. For each, estimate the cost of being wrong — in time, materials, lost customers, or rework. The highest-wrong-cost items are your biggest opportunities. Look especially at the decisions you have stopped questioning: the production sequence you always run the same way, the pricing you set once and never revisited.

2. Separate concept from execution. Split each item: what can be tested digitally (descriptions, pricing, sequences, schedules, copy) versus what requires physical commitment (cooking, building, installing). Operators assume the whole process must be physical because the product is. It rarely is.

3. Find your verifiable metric. Revenue per item. Quote conversion rate. Time to completion. Cost per unit. If it lives in your POS, CRM, or QuickBooks, the feedback loop already exists. You do not need a taste algorithm — you need to narrow 100 menu configurations to 5 on margin and sales velocity, then taste-test those 5.

4. Calculate your iteration ratio. How many options do you test before committing? If the answer is one or two, you are running 1:1. Roofle contractors operate at 135:1. The point is not a magic number — it is that nearly every physical business defaults to 1:1 when 50:1 or 100:1 is already possible.

5. Start the smallest loop. Pick one decision. Generate ten alternatives digitally. Evaluate them against your metric. Commit the best one physically. Measure. That is one complete loop. The second will be faster; the tenth will feel automatic. Unilever did not start with eight digital twins on four continents — they started with one factory and one process.

The gap is not between digital businesses and physical ones. It is between those who iterate digitally before committing physically and those who still test everything live.

If your work demands creating something — a menu, a bid, a production sequence, a marketing plan — digital tools can help you compare alternatives before committing resources. The 82% who believe AI does not apply to them are paying the highest iteration tax in their industries. Every wrong guess tested in the real world could have been tested for nearly nothing on a screen.

Begin with one decision and ten digital alternatives. Compare them, try the strongest option in the real world, and see what happens. Keep physical testing where the screen cannot tell you enough.

References

physical workiterationdigital twinssmall businessproduct developmentsimulationoperational practices
About the author
Héctor Reinoso writes about how organizations should be designed when intelligence becomes abundant. He works on enterprise AI adoption inside a large industrial organization, after 25 years as a strategy consultant and transformation executive across four continents. Harvard MBA, industrial engineer, and still writes the code he writes about. More →
Continue reading
Org Design · Oct 2, 2026

When Coding Gets Cheaper, the IT Handoff Gets Harder to Justify

If your supplier uses coding agents while you still specify and verify the work, it is time to reconsider what you are outsourcing.
Org Design · Sep 26, 2026

The Myth of the 50-Agent Founder

The fantasy of a solo founder running a company with 50 agents misunderstands what agents do to cognitive load.
Strategy · Aug 11, 2026

Compute Is Not Free Lunch

Don't ask what AI costs — ask what your people can do with it.
Strategy · Aug 3, 2026

The Fixed Pie Fallacy

When AI collapses execution cost by 10x, the rational response is not to do the same work with fewer people — it is to do dramatically more work.
Org Design · Jul 28, 2026

Enterprise AI Will Be Slower Than the Demos

Enterprise AI is not slow because the models can't do the work.
Org Design · Jun 18, 2026

The Dissipation Gap

The real AI advantage is not the technology — it is the distance between what is possible and what your organization can absorb