Choice Compression™: The New Constraint
AI made the build fast. Now every decision is twenty minutes away, and a person still has to make it.
A note on evidence, before we start
This paper names something I have not yet seen in an analyst report. The scarce resource in AI-assisted software delivery is the human decision that has to follow every piece of generated code, and those decisions now arrive faster than most organizations can make them. We call that Choice Compression™.
Some of the support for that claim is public record: peer-reviewed psychology, Gartner research, a Department of Defense strategy, and an Inspector General report. Those items carry an author and a year and appear in the references. The rest comes from what we have watched inside ACC3’s own product and delivery work. Those items are marked [A] for ACC3 observation. They are real, they are ours, and I won’t dress them up as third-party statistics.
1. Every decision is 20 minutes away
Think about how a scrum team used to work a story. The lead accepts the stories and sizes them as fives and eights. Then the work begins. A day later, a developer comes back and says, “I was thinking about this one. What do you think?” The lead suggests something to try. The next morning, there is a different idea. Around day four of the sprint, the team has finally worked out what it is going to build.
Those four days were “think time.” Nobody budgeted for them, and everybody used them.
In AI-assisted delivery, and certainly inside the Alchemy SDLC™, every decision is about 20 minutes away. You answer a choice, the machine executes on it, and you are standing in front of the next choice. Our own team will tell you it is exhausting, and what wears them down is the choosing. [A]
I came back from Gartner’s IT Symposium in October with that on my mind and a name to describe it. Choice Compression™ is the flood of decisions that AI-accelerated delivery pushes onto people, with more choices arriving and less cushion between them. The industry spent two years making generation cheap. Judgment did not get any cheaper, so it became the constraint.
2. What Choice Compression™ is
A developer I know got a new boss a while back. When he asked what was expected of him, the answer was simple: if you make one or two good choices a day, you’ve done your job. Nobody mentioned his code. The manager cared about the decisions. [A]
That conversation stuck with me because it describes the unit of work that matters now. Two things separate Choice Compression™ from ordinary busyness. The first is density. Decisions that used to show up a few times a week now arrive all day: accept or reshape a generated change, decide whether a failed test is a real exception, ship or kill a prototype, remediate a vulnerability or defer it. The second is the missing cushion. The gaps that used to buy think time have shrunk to almost nothing.
Neither of those goes away. The tools producing the speed are here to stay, so Choice Compression™ is a permanent operating condition, and it has to be run like one.
That changes the scorecard. An organization still counting hours in a chair and story points closed will miss the constraint completely. The number that matters is how many good choices a team absorbs without creating the next round of debt or shelfware.
3. Where the cushion went
The old delivery system was slow, and all of us complained about it. It was also carrying weight nobody wrote down. End-of-sprint test cycles forced a pause before anyone could call something done. Change-control boards gave hard calls a weekly appointment. Security findings piled into a backlog that could wait for a quiet quarter, and big migrations ran old and new systems side by side for months. That friction gave human judgment a place to stand.
AI-assisted delivery removed the friction and put nothing in its place. Generated tests land with the code and scanners fire on every commit. A working prototype forces a ship-or-kill decision while the requirements argument is still going on in another meeting.
Google’s DevOps Research and Assessment (DORA) picked up the system-level effect in 2024. As AI adoption rose, more than one-third of respondents reported moderate to extreme productivity gains from AI, while delivery throughput fell an estimated 1.5 percent and delivery stability fell an estimated 7.2 percent (DORA, 2024). Individuals felt faster while the delivery system got slower because somebody still had to decide what each of those faster outputs meant.
4. A cousin of the jam aisle
Choice Compression™ borrows from 25 years of choice-overload research, so I want to be precise about where it departs.
Iyengar and Lepper (2000) ran the study everyone remembers. People offered a limited set of about six gourmet jams or chocolates were more likely to buy than when people offered 24 or 30, and students offered six optional essay topics were more likely to write the essay than students offered 30. People choosing from the smaller set also reported more satisfaction. Schwartz (2004) took the idea wider in The Paradox of Choice: as options multiply, the fear of making the wrong choice grows, and deciding turns into a source of stress instead of freedom.
Then the field checked itself. Scheibehenne, Greifeneder, and Todd (2010) pooled 50 experiments and found a mean choice-overload effect of virtually zero, with large variation between studies. More options do not freeze everyone. The effect depends on conditions.
That caution sharpens the claim. Choice Compression™ describes enterprise software delivery at AI speed, where decisions per hour have climbed inside workflows that still need a named human to own the outcome. Those are the conditions that matter. A shopper can walk away from the jam. An engineer who owns a release cannot walk away from the queue.
5. Outside observers are already seeing it
Nobody else calls it Choice Compression™, but serious institutions are describing the symptoms.
Gartner. Gartner’s October 2026 research on AI return on investment reports that AI use cases grew 400 percent from 2024 to 2025 while the likelihood of success stayed stuck at about one in five (Gartner, 2026a). Only 13 percent of CIOs surveyed said their enterprise has realized significant or transformational value from AI (Gartner, 2026b; Monsanto, 2026). Gartner identifies a Vanguard of about 11 percent of organizations that report both strong business performance and high realized value from AI (Gartner, 2026c), and only 4 percent of organizations report plans to redesign roles for a human-agent workforce (Gartner, 2026a).
I read those numbers as an absorption problem. Use cases went up 400 percent, but the success rate did not move. Gartner does not use the term Choice Compression™ and does not endorse ACC3 or this framing. The interpretation is mine.
The Department of Defense. The DoD’s 2023 Data, Analytics, and Artificial Intelligence Adoption Strategy ties decision advantage to agile, user-focused development and says that practicing agility and learning by doing will accelerate deployment speed, “measured in hours or days, not months or years” (Department of Defense, 2023). The same strategy says humans and machines will work together, and it asks personnel to embrace risk and learn by doing. I agree with the ambition. A deployment clock measured in hours also produces decisions measured in hours, and those land on people.
Jinyoung Englund, then chief strategy officer for algorithmic warfare at the DoD’s Chief Digital and Artificial Intelligence Office (CDAO), described the human side plainly in 2023: “the way that we’re trying to overcome paralysis is by analysis of the risk — and by being an organization that is setting the culture to experiment” (Vincent, 2023). She added that risk can and should be measured and managed. That is the right instinct.
A year later the DoD Office of Inspector General showed what happens when decision rights stay unclear. Its November 2024 evaluation found that delays in foundational AI policy, with an implementation plan still in draft, left the CDAO’s roles and responsibilities unclear to DoD stakeholders, and both CDAO and CIO officials told evaluators the lack of policy had caused confusion (Department of Defense Office of Inspector General, 2024). A culture of experiment without named owners still stalls.
An analogy from the Army. Writing for West Point’s Modern War Institute, an Army officer warned that the growth of data and analytic tools could produce analysis paralysis, and he cited Schwartz on the mechanism: more options heighten the fear of the wrong choice (Forester, 2023). The essay is the author’s own view rather than DoD policy, and I use it only as an adjacent analogy.
6. What the humans still own
Automation drafts. Ownership stays with people, and in our work, it lands in four places.
Someone has to decide what the definition of done is, which is the difference between a demo that impressed the room and a capability the organization can actually run. Exceptions need a human call too, because generated tests and scanner findings produce volume, and volume is not signal. Then there is evidence. What an auditor will trust is a requirements problem, and we treat it as one rather than as a pep talk. The fourth is the one people forget: every AI-accelerated path needs a person with the authority to stop it.
Faster code makes every one of those calls more urgent.
This is the oldest lesson of my career in new clothes. I spent years as the outside fixer on troubled Pega implementations, and every one of those disasters had the same root cause: a gap between what was sold and what the customer expected. AI can widen that gap faster than anything I have seen, because a capable model pointed at a fuzzy decision produces confident, expensive rework at scale. Good requirements with speed, but not without guardrails.
7. The fear under the delay
Choice Compression™ is a queue problem, and it is also a career-risk problem.
Nobody gets fired for leaving a legacy system alone one more quarter. People do get blamed when a bad technology call lands in production with their name on it. When AI multiplies the number of calls that need a name, that asymmetry gets louder, and delay starts to look like prudence. In our field, conversations around this fear come up more than any technical objection does. [A]
Schwartz explains the mechanism: more options raise the felt cost of being wrong. Englund points at the way out, which is to measure and manage risk instead of waiting it out. Managing Choice Compression™ means making a good choice safer to make than an endless deferral.
8. What we have seen at ACC3 [A]
Everything in this section is an ACC3 observation from our own product and delivery work. None of it is a third-party statistic.
The QA queue became the bottleneck. On one AI-accelerated application, a review produced a little over 100 individual observations in a short window. More than 20 were still waiting on triage while new ones kept arriving, some of them logged that same evening. The constraint was the human exception call, and we added people to QA for exactly that reason.
Prototypes arrive before the debate ends. When we replaced our own CRM, Alchemist AI Pro™ produced the requirements package in under 24 hours, and the first working build existed by Wednesday afternoon. We rebuilt the same application four times as the Alchemy Crew matured. On the production build, after the Crew’s automated work, the human final mile came to 61.8 tracked hours against an estimate of 11,912 hours of manual development effort, and that system runs our sales pipeline today (Campbell, 2026). Another ACC3 paper puts the pace bluntly: development teams can now complete their entire sprint before lunch (Brooks, 2026). The time between idea and go/no-go has collapsed into days.
Security never gets a quiet quarter. Our product team has watched vulnerability findings arrive throughout the week and then more over the weekend. Remediate-or-defer has turned into a running stream of choices instead of a backlog we burn down on a schedule.
Capable teams can stall. I have watched a prospect bring its brightest people into a workshop and get overwhelmed because the choices came faster than they could make them. Smart people beaten by the pace of decisions. An organization that cannot absorb AI speed does not receive a productivity gift. It overwhelms itself.
All of this has pushed me to a postulate, and I label it as mine. When AI takes over the busywork, a person’s value moves from hours worked to the quantity and quality of the choices they can own. The task-coder who waits for low-choice stories looks comfortable right now, but that comfort is fragile. The engineers I would hire first are the ones willing to put their name on a decision, even though it leaves them more exposed. Staffing and pay will follow choice quality.
9. A diagnosis you can run
You can run this section without buying anything from us.
What to watch for
-
The build finishes before the debate does. A prototype is “done” while requirements or ownership are still open.
-
Review and exception queues grow faster than triage can clear them, and security accept-or-defer calls stack up every week.
-
Teams stall because a wrong call feels more punishable than standing still.
-
Leaders say the new tools will move the organization so fast it will overwhelm itself, and nobody has been named to own the calls.
This Monday
-
Name the bottleneck out loud. Is it coding speed, or judgment under a firehose? Write one sentence your leadership can repeat.
-
Count open choices next to open tickets. List every decision waiting on a human: accept or reshape, real exception or noise, ship or kill, remediate or defer. If that list grew while the velocity charts stayed green, you have Choice Compression™.
In 30 days
-
Put high-stakes judgment on a cadence. An AI output does not need to become an executive decision the minute it appears.
-
Write down who owns the four calls from section 6 (done, real exceptions, trusted evidence, and the authority to stop) for every active AI-accelerated effort.
-
Map absorption capacity. Ask which teams would overwhelm themselves if build speed doubled next quarter, and sequence the work for that answer.
-
Start reporting choices absorbed next to story points closed.
In 90 days
-
Treat Choice Compression™ as a requirements and governance problem because training alone will not fix it. Require a decision record on every AI-accelerated path before it touches production data or users. The record says who chose what and why, and who can stop it.
-
Find the people who already make good choices at speed and put them where the firehose hits. Gartner defines its Vanguard at the organization level (Gartner, 2026c). Define yours at the person level: the people who handle a heavy flow of decisions without freezing.
10. The Absorption Layer
Telling people to decide faster just moves the firehose. Choice Compression™ has to be absorbed.
Our answer is what we call the Absorption Layer, delivered through the Alchemy SDLC™. We use the word precisely. Work that is absorbed is gone. Work that moved into a bigger human review pile was only relocated, and relocation makes the firehose worse. We also expect the residue of human judgment to shrink over time, with every decision a person keeps traceable to the requirement behind it.
In practice, the person facing the firehose gets a prepared choice and a filtered queue, a cadence for the calls that matter, a translator in place of a bypass, and guardrails on the person as well as the machine. The machine adapts to how people decide, and it should never wire a human into an endless approval treadmill.
Alchemist AI Pro™ sits at the front of that lifecycle as an intelligent preprocessor of AI agents. Its job is to get requirements hyper-defined before anything executes, so fewer half-formed choices reach a person at all. That is how speed turns into delivery on time, the first time.
Absorption takes the work. The human stays in the conversation.
11. Where this leaves us
AI made the build cheap and made judgment the expensive part, and that shift is permanent. The organizations that come out ahead will treat choices as the unit of work and build a way to absorb them. Everyone else will keep buying faster generators and keep landing at one in five.
The delivery evidence and the full ACC3 whitepaper library are public. Read the proof. If Choice Compression™ sounds like your Tuesday afternoon, reach out at jamie@acc3int.com.
References
Public sources
Department of Defense. (2023). 2023 data, analytics, and artificial intelligence adoption strategy: Accelerating decision advantage. https://media.defense.gov/2023/nov/02/2003333300/-1/-1/1/dod_data_analytics_ai_adoption_strategy.pdf
Department of Defense Office of Inspector General. (2024). Evaluation of the effectiveness of the Chief Digital and Artificial Intelligence Office’s artificial intelligence governance and acquisition process (DODIG-2025-039). https://www.oversight.gov/sites/default/files/documents/reports/2024-11/DODIG-2025-039_Redacted%20SECURE.pdf
DORA. (2024). Accelerate State of DevOps report 2024. Google Cloud. https://dora.dev/research/2024/dora-report/
Forester, B. (2023, May 17). Toward the data-driven Army of 2040: Avoiding analysis paralysis and harnessing the power of analytics. Modern War Institute. https://mwi.westpoint.edu/toward-the-data-driven-army-of-2040-avoiding-analysis-paralysis-and-harnessing-the-power-of-analytics/
Gartner. (2026a, October 19). More AI investment is not creating enough AI ROI. https://www.gartner.com/en/articles/ai-roi
Gartner. (2026b, October 19). Gartner survey finds only 13% of CIOs say their enterprise has realized significant or transformational value from AI [Press release]. https://www.gartner.com/en/newsroom/press-releases/2026-10-19-gartner-survey-finds-only-13-percent-of-cios-say-their-enterprise-has-realized-significant-or-transformational-value-from-ai
Gartner. (2026c). Bridging the gap between AI ambition and business outcomes [2027 CIO Agenda infographic]. https://emt.gartnerweb.com/ngw/globalassets/en/chief-information-officer/documents/bridging-the-gap-between-ai-ambition-and-business-outcomes.pdf
Iyengar, S. S., & Lepper, M. R. (2000). When choice is demotivating: Can one desire too much of a good thing? Journal of Personality and Social Psychology, 79(6), 995-1006. https://doi.org/10.1037/0022-3514.79.6.995
Monsanto, B. (2026, October 19). Few CIOs seeing transformational value from AI investments: Gartner. IT Brew. https://www.itbrew.com/stories/few-cios-seeing-transformational-value-from-ai-investments-gartner
Scheibehenne, B., Greifeneder, R., & Todd, P. M. (2010). Can there ever be too many options? A meta-analytic review of choice overload. Journal of Consumer Research, 37(3), 409-425. https://doi.org/10.1086/651235
Schwartz, B. (2004). The paradox of choice: Why more is less. Ecco.
Vincent, B. (2023, October 18). Pentagon’s CDAO promotes “culture to experiment” while confronting AI risks. DefenseScoop. https://defensescoop.com/2023/10/18/pentagons-cdao-promotes-culture-to-experiment-while-confronting-ai-risks/
ACC3 primary sources
Brooks, K. (2026, July 8). AI coding is moving fast. Requirements are still moving slow [Whitepaper]. ACC3 International. https://acc3int.com/whitepapers/ai-moving-fast-requirements-slow/
Campbell, J. (2026, July 14). [Case study of the AI Pro CRM rebuild on the Alchemy SDLC]. ACC3 International. https://acc3int.com/blog/replacing-our-crm-using-alchemist-sdlc
ACC3 observations [A]
Items marked [A] in the text are ACC3 field observations from our own product, delivery, and sales work. They are offered as practitioner observation for the reader to test against their own experience, not as settled research. Client and individual names are withheld.
© 2026 ACC3 International / AI Pro Holdings. Choice Compression™, Alchemy SDLC™, and Alchemist AI Pro™ are trademarks of AI Pro Holdings. Gartner, DORA, and the Department of Defense are cited as public sources only and do not endorse ACC3 or this framing.
