If speed alone created value, the fastest typist would be the richest person in the room. AI has multiplied the volume and velocity of tasks, yet bottom-line productivity remains stubborn. The gap between eye-catching demos and financial performance is not a tech problem. It is a systems problem, and systems fracture when we add acceleration without redesign.
The Financial Times asks a timely question because this cycle rhymes with history. When electricity arrived, factories kept their steam layouts and saw little gain until they reconfigured lines and workflows. Early IT installed computers on top of paper processes and watched costs rise before efficiency followed. Economists have a name for this: the productivity paradox. General-purpose technologies demand complementary investments, organizational rewiring, and time. Until then, statistics understate benefits and overstate frustration. That is not a reason to be bullish or bearish. It is a reason to be precise about where value accrues and where fragility hides.
AI reduces the time to draft, code, summarize, and respond. That is activity, not output. Productivity is useful output per unit of total input, including error correction, oversight, energy, and risk. In probabilistic systems, multiplying throughput increases both signal and noise. An extra one percent error rate at scale is not one percent more cleanup. It can be a nonlinear spike in rework, customer churn, and liability. Queueing theory teaches that small increases in variability cause outsized congestion when systems run near capacity. Many enterprises have layered AI atop already tight workflows. They have accelerated both the best and the worst parts of their process. The visible speed is intoxicating; the hidden costs pile up in quality control, audit, and compliance.
Investors crave straight lines; transformation shows up as a J-curve. First, costs rise: cloud bills, model subscriptions, data labeling, prompt libraries, red-team testing, retraining, security hardening. Then you get reorg friction: job design, skills gaps, incentive rewrites, vendor sprawl. Only after these complements mature do we see measurable output gains. Until then, spreadsheets misattribute savings to AI that actually come from hiring freezes, contractor cuts, and deferred projects. Surveys of thousands of senior executives show the same pattern: high reported adoption, low depth of use. Many leaders engage with AI for an hour or two per week. That is a signal of symbolic adoption and exploration, not systems-level redesign. The J-curve is healthy if you know you are on it. It is dangerous if you mistake the learning phase for productivity.
There is an engineering constraint most narratives skip: power and capital intensity. Training and inference have real energy footprints. Data center build-outs face grid limits, transformer lead times, and permitting friction. Compute supply is lumpy. That makes the cost curve noisy and the ROI case fragile. If your AI margin story relies on cheap, abundant inference forever, you are running a single-point-of-failure model. Meanwhile, vendor economics remain unsettled. Model providers chase scale with subsidized pricing. Enterprises extrapolate from that subsidy. When pricing reverts to cost, unit economics can flip. The winners in past waves were those who controlled bottlenecks: electricity access in electrification, logistics and distribution in retail, and network effects in consumer internet. In AI, the controllable bottlenecks may be proprietary data quality, workflow integration, and access to stable power. Betting on what you do not control is not a strategy; it is a prayer.
Firms do not adopt technology in a vacuum. They play games with investors, employees, and competitors. Executives face herding pressure: show an AI roadmap to avoid a valuation penalty. Vendors mark growth to pilot programs. Managers rebrand routine automation as AI to win budget. Analysts search for leverage in headcount per revenue, often rewarding announcements more than verified output. This is a signaling equilibrium, not a productivity equilibrium. In game theory terms, we see a classic coordination problem: everyone moves because everyone else moves, not because the payoff matrix is proven. When corporate skepticism rises, it often reflects reality catching up with the payoff mismatch. The pendulum swings from fear of missing out to fear of wasting capital. The silent question across boardrooms is simple: who bears the risk when promised gains slip—vendor, buyer, or customer?
There is also a demand-side variable that supply-side boosters ignore: trust. Consumer surveys show only a minority confident in AI in retail settings, with a notable trust gap by gender. Retailers report minimal measurable shifts in headcount or productivity despite higher AI talk. Institutional sentiment is sliding too, with a marked jump in corporate skepticism as promised gains fail to appear on the P and L. This divergence between adoption headlines and lived experience matters. In past cycles, consumer behavior often led analyst models. Investors who listened to usage friction, not press releases, spotted when hype detached from unit-level economics. Today, the same divergence is widening: executives experimenting lightly, customers watching warily, and models still looking for durable fit.
Even when output improves, we may miss it in the data. Traditional productivity stats struggle to see quality, variety, and speed of iteration, especially in services. But measurement error cuts both ways. Firms also over-measure local velocity and under-measure systemic cost. If you cut drafting time by 50 percent but double compliance checks and customer support escalations, the net may be negative. The denominator problem is everywhere: counting tokens and tasks instead of revenue per hour of total system time. Risk-adjusted productivity matters. That means factoring in new failure modes: privacy leakage, content harms, biased outputs, vendor lock-in, and operational resilience when APIs go dark. Fragility hides in the denominator, where costs compound in silence.
History suggests general-purpose tech rewards the few who redesign the work, not the many who decorate it. Gains flow to firms that rebuild workflows end to end, rewrite incentives, and invest in proprietary data and distribution. They also show restraint. In engineering, control rods keep reactors stable; in markets, governance keeps compounding predictable. The antifragile posture is not maximal automation. It is optionality with redundancy: human-in-the-loop where errors are heavy-tailed, smaller models for local tasks, clear fallback paths, robust measurement of outcomes, and exposure to upside without catastrophic downside. That does not read as an earnings call soundbite. It reads as patient operational craft. The payoff is slower to market, faster to durable returns.
Invert the question in the headline. Ask not how much value AI is creating, but how much fragility it is quietly adding. Then ask who is short that fragility. If history holds, the market will briefly overprice generic AI winners and underprice the dull complements: grid capacity, resilient software plumbing, rights-managed data, and companies with the cultural skill to change how work is done. In probability terms, returns will be heavy-tailed and late. The base case is a long implementation lag, a mid-cycle backlash, and a smaller set of real winners than the indices imply. Speed without redesign is noise. Value shows up only after the hard, boring work of rebuilding the system to carry the load.
