AI Didn't Make Your Company Faster. It Made Your Failure Scalable.
AI genuinely speeds up individual work. But an organization is not the sum of its productive individuals. When decision-making, ownership, review, integration, and trust stay slow, AI doesn't remove your bottlenecks. It makes them visible, and then scalable.
TL;DR. AI genuinely improves individual productivity: +15% issues resolved per hour in customer support, documents drafted 12% faster with Copilot, roughly two fewer hours of email per week in a randomized experiment across 66 companies. But that same experiment shows what doesn’t change: neither the quantity nor the composition of the work. And at the enterprise level, McKinsey measures EBIT impact in only 39% of organizations. The gap between the individual gain and the collective result is the whole story. AI didn’t create your bottlenecks: decision, ownership, review, integration, trust, adoption. It made them visible. And because tooling now scales almost linearly with effort, it has made your dysfunction scalable.
A scene that keeps repeating
There’s a scene that has been repeating in companies for two years now. Meeting notes arrive faster. Documents get written faster. User stories, analyses, emails, prototypes, even code: all of it appears faster. In a randomized experiment run by Microsoft researchers on more than 6,000 employees across 56 companies, M365 Copilot users completed documents 12% faster and spent half an hour less per week reading email.
And yet trade-offs still take weeks. Dependencies are still there. Pull requests still wait. The structural decisions move at exactly the speed they did before.
The usual reflex is to say AI “isn’t mature yet.” Sometimes that’s true. But it’s also a convenient way to avoid a more uncomfortable conclusion: in many organizations, the real bottleneck was never the ability to produce text, code, or slides. It was already in the structure.
Microsoft’s data (a joint Microsoft–LinkedIn study, 31,000 respondents) puts it bluntly: 79% of leaders consider AI adoption necessary to stay competitive, but 60% worry their organization lacks a plan and vision to implement it. Keep in mind, reading those numbers, that they come from Microsoft and concern Microsoft tools. The direction they point in is nonetheless consistent with independent sources.
My previous article called this gap assimilation debt: the distance between what an organization has adopted and what it has truly absorbed, in the lineage of Cohen and Levinthal’s absorptive capacity. This one looks at the same phenomenon from another angle. AI lowers the cost of individual production. It does not lower the cost of coordination, arbitration, or validation.
The gain that doesn’t scale
Too many AI critics skip past the obvious: yes, AI genuinely increases certain individual performances.
In the study by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, published in the Quarterly Journal of Economics in 2025, access to a conversational assistant raised the number of issues resolved per hour by 15% for 5,172 customer-support agents. But the detail matters as much as the average: the gains concentrate among the least experienced agents; the best ones gain almost nothing, and even show a slight quality decline on some conversations. In other words, the aggregate effect depends on the distribution of skill, not on a uniform boost. That heterogeneity is more than a technical footnote. It’s already a preview of the organizational problem.
The trouble starts when we confuse these individual gains with an acceleration of the whole company. An organization is not a sum of productive individuals. It’s a system of dependencies, standards, validations, roles, risks, and trust.
The cleanest demonstration comes from a randomized field experiment covering 66 companies and 7,137 knowledge workers (Dillon, Jaffe, Immorlica, and Stanton, NBER, 2025). The result fits in one sentence: individual access to generative AI sharply reduces time spent on email and after-hours work, about two fewer hours of email per week for those who use it, but does not change the quantity or the composition of tasks. You save time locally without changing the structure of the work. The freed-up time doesn’t turn into something else on its own.
McKinsey finds exactly this tension at enterprise scale. In its State of AI 2025 (nearly 2,000 respondents, 105 countries), 88% of organizations report using AI in at least one function, but only about a third have begun to scale it, and only 39% report a measurable EBIT impact, usually below 5%. The factor most strongly correlated with that impact isn’t the number of copilots handed out. It’s workflow redesign. That’s the empirical bridge between the individual gain and the collective result, and most organizations haven’t crossed it yet.
This is where the title stops being mere provocation. AI makes failure scalable because it lets you generate more work than you can integrate, govern, or convert into outcomes. More plans than the company can decide on. More code than it can review. More documents than it can use. More prototypes than it can productionize. And as we move from assistance to autonomy, the cost of that inability climbs. In June 2025, Gartner predicted that over 40% of agentic AI projects would be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate controls. It isn’t the first time the firm has called this wall: in a July 29, 2024 release, it estimated that at least 30% of generative AI projects would be abandoned after the proof of concept before the end of 2025.
One honest caveat: an analyst forecast is not a measurement. Gartner anticipates; it doesn’t observe. But read alongside McKinsey’s 39% EBIT figure, which is measured, the trajectory is coherent.
Why local speed doesn’t make system speed
There’s an engineering intuition that illuminates everything else. A system’s throughput isn’t set by its fastest steps, but by its slowest constraint. Speed up everything except the bottleneck, and the bottleneck becomes the entire system.
AI sharply accelerates one step: producing a first draft. If that step wasn’t your constraint, you’ve made something faster that didn’t need to be, and piled pressure onto everything downstream: decision, review, integration. It’s the organizational version of an old law of computing. Total gain is capped by the share of the work you don’t accelerate.
That’s the mechanism.
The bottlenecks, one by one
Cheap production meets expensive coordination at six predictable points. None of them got faster.
Decision
AI lowers the cost of the draft. It doesn’t grant decision rights. More summaries, more memos, more business cases don’t mean more arbitration. Microsoft’s data is almost cruel here: employees rush to AI (75% of knowledge workers say they already use it, and 90% of users say it saves them time) while leaders move slowly, unable to translate individual productivity into priorities and accountability for execution.
Ownership
When anyone can produce a proposal, an analysis, or a prototype, the question “who owns the final outcome?” becomes more important, not less. McKinsey’s State of AI 2025 shows that the organizations capturing the most value are markedly more likely to have leaders who explicitly own and drive their AI initiatives. Its State of AI Trust 2026 confirms it from another angle: organizations with explicit AI accountability reach an average maturity of 2.6, versus 1.8 for those without a clearly responsible function. Ownership isn’t a management word; it’s a measurable maturity gap.
Review
This is probably where the promise of speed hits reality fastest, and the place to cite most carefully. In a METR randomized trial (July 2025), 16 experienced open-source developers, working on mature repositories they had known for years, took on average 19% longer to complete their tasks with early-2025 AI tools (Cursor Pro + Claude 3.5/3.7 Sonnet). The unsettling detail: they expected a 24% speedup, and still believed, afterward, that they had been 20% faster.
This result is valuable, but it has to be handled honestly. The sample is small (16 people), the population very specific (experts on their own mature code), and the tool generation already dated. The authors themselves refuse to turn it into a general law about AI. More telling: in a follow-up post on February 24, 2026, METR explains that its new late-2025 round gives an unreliable signal, because 30–50% of developers declined to submit tasks they didn’t want to do without AI, biasing the estimate. So the takeaway isn’t that AI slows developers down. It’s narrower: in some expert contexts, perceived gain and real gain diverge sharply.
An observational analysis of open-source projects (Xu et al., 2025) points the same way without claiming RCT-level causality: core developers review about 6.5% more code and see their original-code productivity drop by roughly 19% after assistance tools are introduced. The logic is constant: AI can raise apparent throughput while shifting the load onto the experts who verify and repair.
Integration
Many companies still treat AI as an assistance layer sitting on top of the information system. You get better drafts, not better flow of value. McKinsey shows that the organizations extracting the most value are the ones that have redesigned their workflows and defined the precise moments where a model output needs human validation. Without that, AI produces lateral artifacts: individually useful, organizationally peripheral.
Trust
The more AI leaves suggestion behind and moves into action, the costlier the gap between technical capability and organizational capability becomes. McKinsey writes plainly that as systems gain autonomy, the consequences of a failure grow. Average AI trust maturity is rising (2.3 in 2026, up from 2.0 a year earlier), but only about a third of organizations reach a high level in strategy, governance, and agentic controls; nearly two-thirds cite security and risk as the main barrier to scaling agents; inaccuracy (74%) and cybersecurity (72%) remain the most-cited risks; and around 8% of organizations report AI incidents, with declining confidence in their own response capability. A 2026 qualitative study (16 practitioners, 12 companies) names it precisely: a capability-deployment verification gap, where teams can demonstrate experimental agentic capabilities more advanced than anything they can put into production, for lack of verification.
Adoption
AI reveals what an organization actually rewards. Microsoft observes that 78% of users bring their own AI tools to work, 52% are reluctant to admit using AI for their most important tasks, and 53% fear it makes them replaceable. A culture that demands AI adoption while punishing error, keeping legitimate uses ambiguous, and rewarding mostly the appearance of control will produce shadow AI and badly integrated usage, every time. The pattern reaches well beyond office work: in CNCF’s 2025 annual survey, “cultural changes within the development team” became the top barrier to cloud-native adoption (47%), ahead of technical obstacles. The hardest constraint is increasingly a human one.
What happens to that time in the system?
At this point, the debate “does AI really boost productivity?” becomes almost too simple. The better question isn’t how much time an individual saves. It’s what happens to that time in the system. Is it converted into faster decisions, better quality, avoided incidents, customer satisfaction, EBIT? Or recycled into more output, more coordination, more review, more rework? McKinsey notes that in 2025, 51% of organizations using AI had already experienced at least one negative consequence, with inaccuracy the most cited. Individual productivity can be quite real and still dissolve before it reaches the bottom line.
The recent history of data should make us modest, too. A systematic review presented at ACIS 2024 cites estimates that 80–87% of big data projects fail to produce sustainable solutions. Be careful with those numbers: they’re estimates from consultancies and industry surveys, not field-measured rates. The range has become a kind of folklore. What is solid is the cause: these reviews consistently show that failures rarely stem from a single factor. Technical problems (data quality and integration) sit alongside skills shortages, cultural resistance, ethical and legal constraints, financial limits, and methodological weaknesses. The stack wasn’t the whole problem; the organization around the stack was.
The counterargument worth taking seriously
The strongest counterargument deserves a fair hearing. In some contexts, AI doesn’t only accelerate individuals; it genuinely improves the system.
Customer support is one example, because the flow is well instrumented, the feedback loop is short, and quality criteria are observable, which is exactly what Brynjolfsson, Li, and Raymond show. Systems administration is another: a randomized trial on Microsoft Security Copilot reports +34.5% accuracy and −29.8% task time for IT administrators (again, Microsoft tooling, to be weighted accordingly). And in a far more demanding environment, Meta/WhatsApp engineers describe an internal system, WhatsCode, that multiplied privacy-verification coverage by 3.5 (from 15% to 53% over roughly 25 months) and generated more than 3,000 accepted code changes.
These counterexamples don’t contradict the thesis. They sharpen it: when an organization bounds the workflow, makes ownership explicit, decides where validation lives, and treats trust as a capability, AI can create useful speed. It’s no accident that McKinsey’s ~6% of “AI high performers” (those attributing more than 5% of their EBIT to AI) are also the ones who redesigned their workflows and named owners. So the message isn’t “AI doesn’t work.” It’s more precise and more demanding: AI’s organizational return is conditional. The conditions have names, and they’re exactly the six bottlenecks above.
What organizations should do
The practical conclusion, then, isn’t “use less AI.” It’s to stop steering AI as a personal-productivity tool.
The dashboard that matters tracks more than the number of users, prompts, or copilots activated. It tracks decision lead time, review time, rework rate, backlog age, incident-response quality, and business impact. Assign explicit decision rights and real accountability to every significant AI workflow; McKinsey quantifies the maturity gap that accountability produces. Budget for validation instead of treating it as an externality: if AI creates more volume, someone has to verify that volume. And redesign the workflows themselves, integrate AI into systems of record, define when humans validate, and treat the trust layer as a business capability rather than a compliance afterthought.
What it comes down to
So AI didn’t make your company faster. Mostly, it removed an excuse.
For a long time, it was comfortable to believe the organization moved slowly because producing was expensive. Producing is already cheaper now. What remains is the real problem: it was never the typing. It was the structure.
The organizations that treat AI as a design diagnostic, not a productivity gadget, will pull ahead. The others won’t simply be slow faster: they’ll make their own dysfunction scalable.