Firms in the same survey forecast a 1.4% productivity increase and a 0.7% employment reduction over the next three years: meaningful, and far short of the wholesale transformation the hype suggests. The evidence points at one conclusion. The models work; the organizations are not built to convert what they produce.
The binding constraint on enterprise AI value realization is knowledge transfer bandwidth: the conduit between organizational knowledge and AI execution. This holds in IT, and it holds equally in finance, legal, HR, operations, and sales. The pattern is consistent. AI excels at well-specified, legible tasks; everything that depends on tacit knowledge, organizational context, and judgment under genuine uncertainty still requires human mediation. The constraint persists across model generations, because it is an organizational design challenge, and organizational design challenges are closed by deliberate investment.
AI does not create enterprise value through generalized productivity gains alone. Value materializes when AI is connected to specific economic levers in the business model. In an insurance and financial services enterprise, the pathways typically concentrate in four places:
Automation and augmentation in claims processing, policy servicing, and internal operations.
Faster claims resolution, underwriting decisions, and customer response times.
Increased conversion, cross-sell effectiveness, and personalization in customer engagement.
Improved anomaly detection, fraud identification, and underwriting signal quality.
The implication is direct: every prioritized AI use case carries a clearly defined primary value pathway and an accountable owner responsible for realizing it. Productivity improvements that do not connect to an economic lever will not appear in firm-level results, which is precisely what the survey evidence shows at scale.
| Function | AI opportunity (near term) | Irreducible human gap |
|---|---|---|
| Finance | Transaction processing, anomaly detection, forecasting, close automation | Materiality judgment, audit accountability, board communication, novel standard interpretation |
| Legal / Compliance | Contract review, regulatory monitoring, policy drafting | Adversarial negotiation, examiner interfaces, novel regulatory interpretation, fiduciary accountability |
| HR / Talent | Screening, scheduling, onboarding, benefits administration | Performance conversations, org design, change management, leadership pipeline development |
| Sales / Distribution | Content generation, lead scoring, personalization, routing | High-stakes relationship management, brand judgment, competitive insight, customer trust recovery |
| Operations | Demand forecasting, workflow optimization, process automation | Novel disruption response, supplier relationship management, safety-critical decisions |
| IT / Engineering | Code generation, QA automation, infrastructure provisioning | Architecture judgment, incident command, accountability ownership, integration knowledge |
| Strategy | Scenario modeling, pattern synthesis, information aggregation | Conviction under uncertainty, organizational persuasion, competitive differentiation |
The gap map is a planning instrument. The left column is the automation frontier as it stands; the right column is where the conduit terminates in judgment that must be transferred to humans, developed in humans, and retained in humans. An AI roadmap that ignores the right column produces pilots; a roadmap that designs for it produces an operating model.
If the binding constraint is knowledge transfer bandwidth, then the response is an engineering program with instrumentation, and two companion bodies of work supply it. The Knowledge-Driven Design framework provides the measurement spine: knowledge transfer treated as a flow system, with a bounded sustainability index and a time derivative, so conduit width becomes a quantity that can be trended and invested against. The four-tier agentic architecture provides the sequenced implementation: each tier widens the conduit by acting on a distinct term of the same flow model, from complexity reduction through variance collapse to the removal of the expert-hour ceiling. Together they convert the argument of this paper into an operating discipline: name the constraint, instrument it, and elevate it deliberately. Both companion frameworks are documented in full on my private research site, available by invitation.
Yotzov, I., et al. (2026). NBER Working Paper 34836 (February 2026). Survey of approximately 6,000 executives across four countries.
EY US AI Pulse Survey, Wave 4 (October 2025). n=500 SVP+ decision-makers.
Fruits, E., & Stout, K. (2026). International Center for Law & Economics (February 2026).
Brynjolfsson, E., et al. (2025). "Canaries in the Coal Mine."