# Constraint Whack-a-Mole: What AI Adoption Actually Feels Like

Published: 2026-07-20T13:00:00.000Z · Updated: 2026-08-01T13:25:16.000Z · Author: Rod Amora · Canonical URL: https://rodamora.com/blog/constraint-whack-a-mole-what-ai-adoption-actually-feels-like/

> Your firm was built for a narrow band of output per person, and AI breaks the ceiling. Google's DORA data shows individuals getting faster while delivery gets slower. The fix is a constraint hunt that never ends.

Google's DORA program surveyed [about 39,000 technology professionals in 2024](https://dora.dev/research/2024/dora-report/?ref=g.rodamora.com) and found something strange. When a team's AI adoption rises by 25%, individual productivity goes up about 2.1%. Delivery throughput for the same teams falls 1.5%, and delivery stability falls 7.2%. The people got faster. The firm got slower. Same tools, same quarter, same dataset.

That gap is not an AI problem. Your firm was designed around a narrow expected range of output per person, and AI pushes people out the top of that range into approval chains, staffing models, and coordination systems sized for the old number.

The fix is a loop: find the step where finished work is piling up, remove that step or speed it up, then expect the next pile one step up the chain. That loop is why AI adoption feels like whack-a-mole, and it does not end.

## Your firm is sized for a band, not a floor

Every operator already manages the low side of output. Minimums, utilization targets, performance plans, the awkward conversation when someone's numbers slip. A century of management practice went into handling people who produce too little.

Almost nothing went into handling people who produce too much, because until now the ceiling took care of itself. A hardworking analyst might outproduce a peer by 30%. Nobody builds systems for the case where she outproduces the peer by 300%.

The band is built into the structure, whether you chose it or not. Managers carry a handful of direct reports, [a median of six in Gallup's 2024 data](https://www.gallup.com/workplace/700718/span-control-optimal-team-size-managers.aspx?ref=g.rodamora.com), because that is how much reviewing and deciding one person can do. Professional-service firms run on leverage pyramids, and [David Maister's classic model](https://davidmaister.com/articles/16/2/?ref=g.rodamora.com) treats "managing the leverage structure" as the thing that keeps a firm in balance: each level is assumed to produce within a known range, and staffing, pricing, and promotions are all priced off that assumption. Client sign-off cycles and internal approval steps are sized to expected volume too.

![Isometric sketch of a stepped pyramid structure with small figures working on each tier and its joints marked in blue](https://rodamora-uploads-347628392068-sa-east-1-an.s3.sa-east-1.amazonaws.com/illustrations/2026-07-31-productivity-absorption-limits/fig-illus-1.jpg "The output band your org chart assumes")

Too little output and the firm loses money. Too much and the firm jams. The interesting question is what jamming looks like, because it rarely gets diagnosed as jamming.

## What does breaking the ceiling look like?

Software teams hit this first, because their output is measured better than anyone else's. [Benchmarks from Faros AI](https://www.faros.ai/blog/ai-software-engineering?ref=g.rodamora.com) across engineering organizations found that teams using AI merge 98% more pull requests. Review time on those pull requests goes up 91%. The changes are 154% larger, and 31% more of them merge with no review at all. The producing side of the pipeline doubled while the checking side stood still, so the checking side is now the pipeline.

An executive director at Morgan Stanley [put the ceiling in one sentence](https://moderne.ai/blog/ai-didnt-break-coding-it-broke-code-review?ref=g.rodamora.com): in the time a short conversation takes, "you probably could have produced a thousand-file, 10,000-line PR on the back of just a simple prompt. No human here is going to review that."

The same pattern shows up far from code. Researchers at BetterUp Labs and Stanford [surveyed 1,150 US desk workers in September 2025](https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity?ref=g.rodamora.com) and found that about 40% had received AI-generated work in the past month that looked polished but needed redoing. Each instance cost close to two hours of rework. Producing the document got cheap. Judging it and fixing it stayed expensive, and that cost moved downstream to whoever received it. That is where the AI productivity gain goes: absorbed by the tasks around it that never adapted.

There is old math behind why this feels so sudden. Queueing theory says that when work arrives faster than a reviewer can clear it, wait times don't grow gradually. They blow up. A review queue that ran fine for years can triple in a month once arrivals cross the line, which is why the jam surprises firms that thought they had slack.

So the jam is predictable. What puzzled me for a while is why most firms don't read it as a jam.

## The lackluster first result is the constraint announcing itself

I've watched hundreds of service firms adopt AI through a franchise network's delivery data, and the failed implementations mostly die the same way. The tool goes in, the surrounding workflow stays untouched, early results look lackluster, and AI gets dismissed. The firms that got it right approached the same lackluster result with curiosity and started [updating the functions and workflows around the tool](https://rodamora.com/blog/your-ai-is-just-a-better-google-search/?ref=g.rodamora.com).

The misread is understandable. When reports come back faster but the client still waits three weeks, it looks like the AI changed nothing. What the result is showing you is a location: the report generator was never your constraint. The partner who reviews reports is, and the faster generator just made that visible.

The gap between those two readings is now measurable at population scale. Gallup found in 2026 that [65% of employees say AI improved their own productivity, while only 12% strongly agree it has transformed how work gets done in their organization](https://www.gallup.com/workplace/713063/ai-workplace-productivity.aspx?ref=g.rodamora.com). In the same research, only 25% of US employees said their organization has communicated a clear AI strategy. Most firms are sitting in that gap right now, holding [a faster tool and an unchanged structure](https://rodamora.com/delivery-model-ladder/?ref=g.rodamora.com), deciding whether to get curious or quit. The curious ones follow a method much older than AI.

## Hunt the constraint up the chain

Eliyahu Goldratt built the Theory of Constraints around a blunt observation from factory floors: an hour saved at a non-bottleneck is a mirage. Speed up a step that was never the constraint and the work just piles up in front of the step that is. AI almost always lands on non-constraint steps first, because producing drafts, code, and reports is what it is best at. The constraint in a service firm is usually review, approval, or a decision, and often it is the owner.

So the work of AI adoption is a loop. AI exposes the next constraint up the chain. You evaluate the risk at that step, then either remove the step or apply AI and speed to it. Then you hunt the next one.

The loop looks like this in practice:

- Reports come back faster, and the reviewing partner becomes the pile-up point. Add an automated first-pass review, so the human pass starts from a checked draft instead of a raw one.
- Support answers routine tickets faster, and complex escalations now swamp your senior people. Give the AI more tools and context so it can propose solutions on the hard tickets too, and the senior person starts from a proposal instead of a blank screen.
- An AI SDR fills the pipeline, and your closers drown in leads they would never have chased. Use AI to filter and prioritize the pipeline with the tribal knowledge of your best salespeople, encoded from the way they qualify.

Three different departments, one identical move. The win relocated the constraint one step up the chain, and the fix was applied at the new constraint, not the old one. Firms that keep making that move compound. Firms that stop after the first install collect faster tools and slower throughput, which is exactly the DORA picture.

## Sometimes the right fix is deleting the step

Speeding up a step is the second-best outcome of a constraint hunt. Before you automate a step, ask a first-principles question about it: does this add value to the customer, or does it exist because humans were slow or error-prone? If the answer is the second, the step may not deserve acceleration. It may deserve deletion.

SpaceX's Starship booster has no landing legs. Instead of engineering better legs, they built a tower that catches the booster, and the legs, their weight, and their failure modes left the vehicle entirely. The best version of a component you can delete is no component. Service firms carry plenty of landing legs: a second approval that exists to catch fatigue errors a machine doesn't make, a weekly status meeting that exists to move information a system now moves on its own, a QA step that re-checks what an automated check already caught.

![Two rocket boosters drawn side by side: one standing on landing legs, one legless held by a tower arm, the legs in blue outline](https://rodamora-uploads-347628392068-sa-east-1-an.s3.sa-east-1.amazonaws.com/illustrations/2026-07-31-productivity-absorption-limits/fig-illus-2.jpg "Delete the step instead of improving it")

[Evaluate the risk first](https://rodamora.com/blog/delegate-the-inputs-own-the-outputs/?ref=g.rodamora.com). Some steps manage real exposure, like client sign-off or compliance review, and those get sped up, never removed. But a step that survives only because "we've always done it" is a strong deletion candidate, and deleting it costs less than automating it.

## This never ends, and that's the point

The constraint never disappears when you fix it. It relocates, one step up the chain, and it will keep relocating for as long as the underlying tools keep improving, which is to say for years. AI adoption becomes a game of whack-a-mole, and I mean that as a description of the job, since firms that stop playing drift back toward worse results. The DORA stability numbers are what standing still costs.

I've written before about the other wall of this same band: when AI frees up time and nobody directs it anywhere, [the gain quietly evaporates into slack](https://rodamora.com/blog/the-extra-time-ai-buys-you-is-already-gone/?ref=g.rodamora.com). Too little absorbed, too much absorbed, same root cause. The firm had no plan for output leaving its expected range.

The practical version fits in one week. Walk your delivery chain and find where finished-but-waiting work is piling up. A folder of drafted reports nobody has reviewed, a queue of tickets waiting on one senior person, a pipeline of leads nobody has qualified. That pile is your constraint announcing itself. Evaluate the risk, then remove the step or speed it up. Then expect the next pile, one step up the chain, and go looking for it before it finds you.

The question worth dropping is ["does AI work?"](https://rodamora.com/four-numbers/?ref=g.rodamora.com) The question that pays is "where is the pile right now?"