AI capability improves, productivity increases, and margins expand. The resulting cash flow is directed toward additional compute, larger training runs, and faster deployment. Each round of investment enhances performance and lowers costs, reinforcing the economic case for further adoption. Over time, the gains generated by AI finance its continued expansion, creating a cycle that sustains itself.

Today, AI capabilities are improving rapidly and measurably. Frontier models have advanced significantly. Anthropic's Claude 4.6 Opus and Sonnet lead independent rankings for agentic coding, complex reasoning, and daily workflow reliability, often handling hours-long autonomous tasks with low hallucination rates. Sonnet 4.6 beats prior-generation Opus on most coding benchmarks at roughly one-third the cost. OpenAI's GPT-5.2 delivers a perfect 100% on the AIME 2025 math benchmark, and hallucination rates down to ~6.2%, a ~30% improvement from GPT-5.1. Models now assist in designing their own successors.

These capability leaps are driving productivity increases. At the firm level, productivity gains are already double-digit. According to BCG, companies that have scaled AI in the tech function are seeing productivity improvements of more than 25% on average, with expectations of more than 45% once AI is fully deployed across workflows. According to Morgan Stanley, enterprises deploying AI for a year or more report average net productivity rises of 11.5%, with 14% of firms seeing over 20% jumps. Macro data confirms. According to the US Bureau of Labour Statistics, nonfarm business sector labour productivity rose 4.9% through Q3 2025, 1.9% in Q3 from a year earlier. However, unit labour costs declined 1.9%. Firms are generating more output per hour without proportional increases in labour cost.

As AI capabilities improve, companies will begin adjusting their labour needs, particularly in white-collar functions. Hiring slows, selective layoffs increase, and payroll growth weakens. Productivity rises, but income growth does not keep pace. Intelligence becomes abundant and inexpensive and in the short term, margins benefit. Over time, weaker household income constrains consumption. This is a “feedback loop with no natural brake”, as Citrini writes. Output remains visible in national accounts yet disconnected from wages and consumer demand. This divergence is the early form of what Citrini describes as “Ghost GDP.”

The structural distinction from prior technology waves lies in cognitive substitution. Previous automation created new categories of work that still required human execution. AI improves within those same categories. Displaced workers historically redeployed into adjacent cognitive roles. That buffer is narrowing. New roles appear, but the wage and volume absorption is asymmetrical. The legal analyst displaced by contract review software does not easily become the prompt engineer overseeing it. The transition assumes a pace of retraining that the current rate of model improvement does not allow.

In Citrini’s scenario, autonomous AI agents operate continuously, handling consumer decisions, automatically comparing prices, assembling itineraries, negotiating renewals, and executing transactions without human involvement. In such an environment, business models built on habitual intermediation where companies profit from consumer inertia, loyalty, or search friction lose their edge. A large share of what has traditionally been labelled “relationship value” is, in economic terms, friction embedded in the process.

AI agents remove that friction. They do not rely on habit, brand loyalty, or search shortcuts. They systematically route toward the lowest cost and most efficient execution path. Platforms that monetise consumer inertia, such as food delivery apps, travel booking sites, insurance renewal channels, tax preparation services, financial advisory layers, payroll processors, marketplace commissions, advertising exchanges, and payment networks, become exposed to compression.

Markets have already reacted to this possibility. Following the publication of Citrini Research report, shares of major payment networks fell sharply. Mastercard dropped about 5.7 %, Visa declined roughly 4.5 %, and American Express fell about 7.2 % in a single session, erasing more than $18bnin market value. If AI-driven commerce systematically routes transactions toward lower-cost rails, traditional card networks could see fee compression. Stablecoins, instant bank transfers, or alternative settlement layers become economically viable substitutes.

The structure of private credit is built to absorb shocks through mechanisms like closed-end funds, senior secured lending, diversified borrowers, and stable institutional capital. Losses are typically recognised gradually, as illustrated by episodes such as Zendesk’s $5bn ARR-backed loan, where capital structures absorbed pressure without systemic contagion. However, this stability depends on the health of underlying capital, including household savings and insurance liabilities. Regulatory downgrades of software debt and rising capital requirements could spread these losses to broader markets.

Despite a solid financial foundation, investor fears in private credit have been rising.

Since the beginning of the year, investors reflected the pressure on private credit: Ares Management is down -27%, Blue Owl -24%, KKR -24%, TPG -30%, and Apollo -20%, while the S&P 500 is broadly flat.

Moody’s and Fitch report that most direct lending portfolios maintain senior secured loans, low non-accruals, and conservative leverage. US private credit default rates remain low at 2.46% for Q4 2025, and first-lien protections safeguard recoveries. Fundraising remains strong; Preqin notes record private debt fundraising in 2025 reflecting continued confidence in the underlying assets.