If you took a parallel computing class, you know this formula by heart:
Speedup = 1 / (s + (1 - s) / N)
Gene Amdahl published it in 1967 (Amdahl, 1967). It says the maximum speedup you can get from throwing more workers at a problem is bounded by the fraction of the problem that has to run sequentially. The arithmetic is brutal. If 5% of the work is serial, the ceiling is 20x even with infinite parallelism. If 10% is serial, you cap at 10x. If 20% is serial, you cap at 5x.
The serial fraction is the bottleneck. Always.
Now substitute. The parallel fraction is the work an individual engineer does at a keyboard, with AI making it faster. The serial fraction is everything else the team does: decisions, handoffs, alignment, code review, judgment calls, escalation, planning, debugging the thing the agent confidently broke.
This is why your developers feel 30% faster and your shipping velocity feels flat. Amdahl warned you 59 years ago.
The Individual Numbers Are Real
Nobody is claiming AI does nothing for individual output. The data is solid where it is solid.
Peng et al. ran a randomized controlled trial of 95 developers using GitHub Copilot on a JavaScript HTTP server task and found a 55.8% speed-up on completion (Peng et al., 2023). That study is narrow (one controlled task, one language, one tool), but the effect size was clean.
Dell'Acqua's BCG / Harvard Business School "Jagged Frontier" study put 758 BCG consultants through 18 realistic business tasks with and without GPT-4. Inside the AI's frontier of capability, consultants completed 12.2% more tasks, worked 25.1% faster, and produced output rated 40% higher in quality. Outside the frontier, accuracy dropped by 19 percentage points compared to the no-AI control (Dell'Acqua et al., 2023).
GitHub and Accenture ran an enterprise rollout study that reported roughly 55% individual code velocity improvements (GitHub Research, 2024). Shopify's CEO Tobi Lütke went public in April 2025 with a memo making "reflexive AI usage" a baseline employment expectation; teams have to prove AI cannot do the work before requesting headcount (Lütke, April 2025).
The Firm-Level Numbers Are Not Real
NBER working paper w34836 (Atlanta Fed, Bank of England, Bundesbank, Macquarie, surveying roughly 6,000 executives between November 2025 and January 2026) found that 89% of firms report no measurable firm-level productivity impact from AI. More than 90% report no employment impact (NBER w34836, 2026).
NBER w34984 from March 2026 (about 750 corporate executives) put the 2025 labor productivity lift from AI at 0.4-0.8% and noted explicitly that "perceived gains exceed measured gains" (NBER w34984, 2026).
Daron Acemoglu's NBER w32487 estimated AI's ten-year contribution to total factor productivity at less than 0.66%, or roughly 0.05% per year (Acemoglu, 2024).
McKinsey's November 2025 State of AI found only 39% of organizations link any EBIT impact to AI, mostly under 5%. Only 5.5% qualify as "AI high performers" with more than 5% EBIT lift (McKinsey, Nov 2025).
MIT's Project NANDA, published in July 2025, estimated that 95% of enterprise GenAI pilots had produced no measurable P&L return despite $30-40 billion invested. The methodology has been contested, but the directional finding lines up with the others (MIT NANDA, July 2025).
And then the most uncomfortable one. METR ran an RCT in early 2025 with 16 experienced open-source developers working on 246 real issues in their own repositories. The developers were 19% slower with AI assistance. They believed they were 20% faster (METR, 2025). Small sample, real codebases, expert engineers. The gap between perception and measurement is the entire problem.
The Mapping
Amdahl explains the disconnect.
Take a software team of eight engineers. Roughly speaking, the work splits into two buckets. The parallel bucket is the actual coding, reviewing, and shipping the engineers can do in isolation. The serial bucket is the work that requires the team to act as one organism: deciding what to build, agreeing on an interface, handing off context across roles, reconciling conflicting designs, running the standup, getting unblocked on the database migration, deciding whether the agent's PR is safe to merge.
AI compresses the parallel bucket. A task that took an hour now takes 35 minutes. Great.
But the serial bucket does not parallelize. The architecture decision still has to be made by one or two humans with judgment. The handoff still depends on whether the other engineer is in a meeting. The code review still waits on the reviewer's cognitive bandwidth, which is finite and is now being eaten by AI oversight work (BCG/HBR's March 2026 study clocks a 14% jump in mental energy spent on AI itself for high-oversight workers; BCG/HBR, March 2026).
If your serial fraction was 30% before AI and stays 30% after AI, your maximum speedup is 3.3x no matter how many copilots you buy. If your serial fraction grows to 40% (because every engineer now also has to police AI output), your ceiling drops to 2.5x. The individual gains land in the bucket that was never the bottleneck.
This is Amdahl, but it is also Brooks. Fred Brooks observed in 1975 that adding manpower to a late software project makes it later, because communication overhead grows as n(n-1)/2 (Brooks's Law, 1975). Same family of math. The cost of coordinating workers grows faster than the output they add.
Where the Coordination Tax Lives
You can see the serial fraction if you know where to look.
Asana's Anatomy of Work 2023, surveying 9,615 knowledge workers across six countries, found employees spend roughly 58% of their day on "work about work": status updates, searching for information, switching between apps, hunting down decisions, clarifying who owns what (Asana, 2023). More than half the day is the serial fraction.
Microsoft's 2025 Work Trend Index, drawing on its own telemetry plus survey data, mapped the "infinite workday." An interrupt every 2 minutes. 60% of meetings called ad-hoc. 30% involve cross-time-zone scheduling. 48% of workers describe their day as chaotic (Microsoft WTI, 2025). Microsoft is its own customer. They sell the AI tooling. The coordination tax grew anyway.
And then BCG's "AI brain fry" finding from March 2026: workers who maintain high oversight of AI tools spend 14% more mental energy on AI itself, are 12% more mentally fatigued, and report 39% more major errors (BCG/HBR, March 2026). The cognitive surplus AI was supposed to free up is being consumed by AI oversight.
The serial fraction is not a fixed constant. It is growing.
The Counter-Arguments
Three reasonable objections exist. None of them dissolve the problem.
Gustafson's Law (1988). John Gustafson argued that Amdahl's framing is wrong for real workloads, because problem size scales with available compute. If you give the team more capacity, it takes on bigger problems rather than running the same problem faster, and speedup becomes nearly linear in N (Gustafson, 1988). True for some teams. But most enterprise teams operate under fixed budgets, fixed headcount, and fixed deadlines. The shape of their constraint is Amdahl, not Gustafson. They cannot expand the problem to absorb the surplus capacity. The surplus shows up as faster individuals stuck waiting on the same coordination loop.
The J-Curve. Erik Brynjolfsson's productivity J-Curve work argues that general-purpose technologies require complementary intangible investments (process redesign, new skills, new org structures) and productivity dips before it lifts. He estimated U.S. TFP was 15.9% higher than measured by end of 2017 once you adjusted for intangible complements (Brynjolfsson et al., NBER w25148). Reasonable framing. The McKinsey 2025 data is consistent with the dip phase. But the J-Curve does not promise every team reaches the upswing. McKinsey found the high-performer gap is widening, not closing. Some teams are climbing; most are still in the dip; and the variable that separates them is workflow redesign, not tool adoption.
Agentic AI will fix this. The argument: better agents will handle the coordination work itself. Maybe. But the METR result and the Microsoft "infinite workday" data point the other way. As agents proliferate, new handoffs (human-to-agent, agent-to-agent, agent-to-human-for-approval) and new judgment steps appear. The bottleneck moves. Engineers now spend their cognitive surplus verifying agent output (the -19 percentage point accuracy drop outside the frontier, per Dell'Acqua) rather than gaining surplus. The serial fraction reshapes; it does not vanish.
What Actually Moves the Team-Level Number
McKinsey's 5.5% high performers were not running different tools. They were running different workflows. McKinsey reported that high performers redesign workflows about three times more often than the rest, and workflow redesign is the single largest differentiator in EBIT impact from AI (McKinsey, Nov 2025).
Klarna is the cautionary tale. In February 2024, Klarna's OpenAI partnership replaced 700 customer-service roles, handled 2.3 million chats in the first month, cut resolution time from 11 minutes to 2 minutes, and the company said the AI assistant was doing the work of 700 FTEs (Klarna, Feb 2024). By 2024 and 2025, Klarna was rehiring humans because CX dropped on the edge cases (Fast Company, 2025). The lesson is on the nose. Klarna automated the parallel fraction (volume) and crashed on the serial fraction (judgment, empathy, escalation). The bottleneck always wins.
The teams that move the firm-level number are the ones that look at the serial fraction directly and ask: where is it, why is it there, can it be shrunk, can it be shared?
Three concrete moves work.
Shrink it. Audit the serial steps your team runs in a week. Decision rights left ambiguous on purpose? Code reviewers who are bottlenecks because only they know the system? Handoffs that go through three Slack channels and a Notion page? Each of those is a slice of the serial fraction. Killing one of them is worth more than another copilot license.
The way you reduce information silos is also the way you shrink the serial fraction. Same problem, different name.
Share it. The serial fraction is mostly mental. Decisions live in heads. Context lives in chat logs. The team can only act in parallel when its mental models are aligned, which is why a cross-functional team that shares context outperforms one that does not. Shared mental models are the cheap way to compress the serial fraction without firing anyone or building anything.
Don't automate it. The Klarna pattern repeats whenever a team tries to AI-automate the serial fraction directly. The judgment work, the escalation work, the empathy work, the "is this safe to merge" work, the "do we ship this on Friday" work. That is the part of teaming that AI is structurally bad at, and the part the team needs the human reps in. This is the teaming skill AI cannot replace, and the firms that try to replace it land back where Klarna did.
Where QuestWorks Fits
If the serial fraction is the ceiling, the work is to give the team better reps at it. That is what QuestWorks does. Twenty-five-minute, AI-facilitated quests of 2-5 players, running weekly on QuestWorks' own platform. Players run shared scenarios that force the same decisions, handoffs, and escalations a real workweek runs, but in a compressed environment where the team can practice the serial steps directly. HeroTypes (the nine public role archetypes) surface who handles what kind of decision. QuestDash gives the whole team a leaderboard with behavioral callouts. The Weekly Team Health Report gives leaders aggregate trends and individual strengths-based XP highlights, without exposing private coaching conversations.
The point is not that quests are productive in themselves. The point is that the serial fraction is the bottleneck, and the only way to shrink it is to run the team through enough shared reps that decisions, handoffs, and alignment get faster. That is the missing tier in the team intelligence stack, and it is the part Amdahl tells you matters most.
The teams that ship the AI dividend will be the ones with the smallest serial fraction. Copilot count is a distant second.
