Two thresholds

Individual productivity gains from generative AI are among the most consistently documented findings in recent empirical economics. Brynjolfsson and colleagues, studying the staggered rollout of an AI conversational assistant across more than 5,000 customer support agents, found a 14 percent average productivity gain, with novice workers gaining 34 percent. Noy and Zhang found that 453 professionals using AI for writing tasks completed them 37 percent faster and produced output rated significantly higher in quality by independent evaluators. Dell'Acqua and colleagues, studying consultants at Boston Consulting Group, found that within the range of tasks AI can competently assist, AI-assisted consultants completed 12.2 percent more tasks and produced output rated 40 percent higher. Against these findings, the logistics sector's position is striking: as of 2026, only 44 percent of shippers report any AI use in transportation planning, and fewer than half of carriers use AI for pricing and lane optimization.

This divergence has two distinct explanations that most analyses conflate. The first is why logistics firms have been slow to adopt generative AI at all, a sector-level question about the structural and institutional barriers that suppress entry into adoption. The second is why, among firms that have adopted, systemic disruption at the organizational level remains elusive. BCG's 2024 survey of over a thousand executives found that 74 percent of companies had not achieved scaled value from AI investments, with roughly 70 percent of implementation failures attributed to people and process factors rather than algorithm inadequacy. These are different problems requiring different explanations. Both are threshold phenomena governed by interdependence, but they operate through distinct mechanisms and respond to different interventions.

This essay addresses both, because strategies that solve only one will continue to disappoint. A TMS vendor that embeds generative AI into its platform reduces the cost of starting but does not guarantee that adoption will cascade into systemic operational change. A change management program that drives high adoption rates across a logistics firm does not produce disruption if that adoption is distributed across functions whose outputs don't depend on each other. Understanding both thresholds is a precondition for building adoption strategies that compound rather than plateau. The analysis draws on network diffusion research, institutional economics, and organizational change management because the two-level problem requires two explanatory frameworks to resolve.

Why firms don't start

Generative AI adoption in logistics is suppressed by a cluster of barriers that reinforce each other across multiple levels of analysis. At the firm level, the primary constraint is the absence of complementary assets. Teece's framework for profiting from technological innovation defines these as the capabilities and infrastructure that determine whether an adopter can actually realize what an innovation offers, and in the generative AI-in-TMS context they consist of clean, structured, accessible freight data; APIs connecting TMS to ERP, WMS, and carrier networks; and organizational capacity to redesign workflows around AI-generated outputs. Most TMS customers, particularly mid-size shippers and small carriers, are missing one or more of these assets. Adoption is therefore economically irrational without significant prior co-investment in infrastructure that has no standalone value, which means expected returns from adoption are systematically underestimated and the timeline to positive ROI is consistently longer than vendor projections suggest.

The industry's structure amplifies this problem. The U.S. trucking sector alone contains approximately 750,000 carriers, with the vast majority operating fewer than 20 trucks. Digital maturity across this population is sharply bifurcated: large carriers and enterprise shippers have ELD-mandated telematics, API-capable TMS platforms, and some data science capacity, while smaller operators maintain baseline digital infrastructure that is insufficient for generative AI. This bifurcation creates a weakest-link constraint on AI value across the supply chain, where a sophisticated shipper's AI-powered routing optimizer is only as good as the carrier tracking data it receives, and when carriers in that network contribute poor-quality or delayed data, the AI's recommendations degrade and the observed ROI falls below what isolated pilots predict. Research on shipping SMEs identifies a structural trap in which limited budgets, absent digital capabilities, and lack of ecosystem collaboration combine to prevent technology investment even when individual firm ROI would be positive given sufficient complementary assets.

Path dependency compounds both problems. Enterprise TMS architecture typically involves deeply integrated platforms such as SAP TM or Oracle TMS, with custom EDI connections to hundreds of carriers, proprietary data schemas, and organizational processes built around specific system workflows. The longer a firm has used a given platform, the more its integrations, data, and processes are aligned to that platform, and the higher the transition cost of introducing a new capability layer. TMS vendors of established platforms have limited incentive to cannibalize their installed base with disruptive AI features. Customers with significant switching costs lack the leverage to force vendor R&D through credible exit threats. The deadlock is self-reinforcing: vendors cite absent demand, customers cite technology immaturity, and neither side moves unilaterally.

Behavioral economics adds a further suppressive force at the individual decision-maker level. Prospect theory, developed by Kahneman and Tversky, predicts that losses weigh approximately twice as heavily as equivalent gains in subjective value assessment, and for operations and IT managers who bear concentrated career risk from failed technology projects while capturing only diffuse benefit from success, this asymmetry is acute. The status quo generates active preference rather than mere comfort, because transition costs are front-loaded and certain while productivity gains are back-loaded and uncertain. Normative and mimetic waiting reinforces this individual calculus. Under genuine uncertainty about generative AI ROI, watching for peer success before committing is a rational institutional strategy. Until Gartner Magic Quadrant evaluations, CSCMP professional standards, and major consulting firm benchmarks treat generative AI as a standard capability expectation rather than an optional enhancement, logistics managers lack the professional legitimation to champion internal adoption even when the economic case is positive.

What breaks the deadlock

The enabling conditions for sector-level adoption are interdependent preconditions, jointly necessary rather than individually sufficient, and this joint necessity has significant implications for intervention design. Platform-embedded AI is the first and most structurally significant condition. The shift from custom AI development to AI delivered as part of the TMS product changes the adoption calculus: implementation costs fall from multi-million-dollar builds to incremental subscription fees, the data integration burden shifts to the vendor, and explainability obligations fall under enterprise SLAs rather than bespoke contractual negotiation. The evidence for this transition is already observable, with 76 percent of enterprise AI use cases now purchased rather than built internally, up from 47 percent in prior periods. Platform-embedded AI resolves the complementary asset deficit without requiring customer-level co-investment, but it does not create urgency on its own: lower cost still loses to inertia when the pressure to change is absent.

That urgency is developing from two independent sources. Labor shortages in logistics are shifting the adoption calculation from efficiency optimization toward operational necessity. Driver vacancies, dispatcher recruitment pressure, and logistics coordinator scarcity driven by demographic shifts and the physical demands of logistics work are making AI automation existential for some operations rather than discretionary. Competitive disruption operates through the same mechanism from the demand side: AI-native freight brokers generate spot quotes in seconds rather than hours, match carriers from larger networks at higher precision, and handle routine exceptions without human intervention. When incumbents experience measurable volume loss in specific lanes attributable to this capability gap, the urgency condition is met independently of any industry-wide demonstration of ROI.

Data infrastructure maturation is the third necessary condition. ELD mandate compliance has substantially increased carrier location data availability and API-grade accessibility across covered fleets. Logistics data network platforms aggregate carrier tracking data across hundreds of carriers, creating data pools that individual shippers cannot assemble and that form the substrate on which AI reliability depends. API standardization initiatives are progressively reducing integration complexity. These developments do not eliminate the data quality barrier but progressively lower it, reducing the cost of the complementary asset acquisition that had suppressed adoption. The point at which generative AI tools in logistics become reliable enough to generate consistent social proof from early adopters is the inflection point at which mimetic adoption pressure begins to build, activating the cascade that Abrahamson and Rosenkopf's institutional bandwagon model formally predicts.

Why disruption stalls

Even firms that clear the sector-level adoption barriers typically fail to achieve systemic organizational disruption. The gap between documented individual-level productivity gains of 14 to 55 percent and limited aggregate firm-level transformation is not primarily a technology problem. The deepest explanation is organizational rather than technological, and it requires a different theoretical frame than the standard diffusion model. Research by Damon Centola distinguishes two fundamentally different ways that behaviors spread through social networks. Simple contagions spread through any connection, including weak ties between distant nodes, because a single exposure is sufficient for transmission. Complex contagions require multiple independent exposures before an individual adopts, because the barriers to adoption are social rather than informational.

Centola identifies four barriers that make adoption complex: credibility (does the behavior work, and have I seen peers confirm it?), coordination (does adoption require others to adopt for me to benefit?), legitimacy (is the behavior socially sanctioned?), and emotional investment (does adoption implicate my professional identity?). Generative AI adoption in organizations meets all four criteria. Workers need to observe peers successfully using AI tools before trusting them with their own output. Many AI tools generate their greatest value in interdependent workflows, where the return to adoption depends on adjacent roles also adopting. Managerial endorsement is required before employees feel safe adopting tools that implicate the automability of their roles. The adoption carries emotional weight around professional identity and skill obsolescence that a single informational exposure cannot resolve.

These properties produce a counterintuitive prediction about network structure. In simple contagion, the fastest-diffusing network is one rich in weak ties and long-range connections, through which information jumps quickly between distant clusters. In complex contagion, this structure impedes diffusion, because weak ties provide only one exposure rather than the repeated independent exposures needed to overcome adoption barriers. Dense, locally reinforcing clusters are the structures through which complex contagions spread. These are groups where many members have interdependent tasks and frequent interaction. For generative AI in TMS, diffusion proceeds fastest through tightly coupled workflow clusters where everyone's daily output depends on shared data and shared processes, and slowest through peripheral roles with sparse organizational ties, regardless of those individuals' personal receptiveness to new technology.

Rogers' diffusion of innovations framework, commonly invoked to set adoption targets, places the tipping point at approximately 16 percent of the adopter population, the boundary between early adopters and early majority. This estimate was derived from innovations where adoption is an individual decision and where value does not depend on what others do. For workflow-interdependent technologies, Rogers acknowledged that critical mass can be substantially higher, because the value of adoption at any point depends on the adoption decisions of functionally linked others. Centola's empirical research on complex contagions, Kotter's observation that organizational change requires a guiding coalition of approximately 20 to 30 percent of influential nodes, and the PROSCI methodology's 25 percent operational milestone converge on a cluster-level threshold of approximately 25 to 35 percent of a task-interdependent cluster's membership as the point at which adoption becomes self-sustaining within that cluster. This estimate is directional rather than precisely calibrated; it is consistently higher than the Rogers tipping point because generative AI adoption is interactive and coordinative rather than individual, and the relevant social unit is the workflow cluster rather than the organization as a whole.

The cluster threshold

The cluster-level threshold has a concrete interpretation in logistics. A freight broker organization where 80 percent of the quoting team uses AI, but where the carrier relations team that feeds the quoting team with capacity data has not adopted, will see productivity gains in quoting absorbed at the interface with carrier relations. Those gains do not compound into systemic change because the downstream cluster is not positioned to process AI-assisted output more fluidly than manual output. A shipper's transportation team where load planners use AI-assisted optimization but exception management remains manual will experience the same absorption: gains accumulate locally but do not propagate through the workflow chain. In both cases, the organizational-level AI adoption rate can be meaningfully positive while systemic disruption remains absent, because cluster-level adoption has not crossed the threshold in functionally linked roles.

This reframes the measurement problem for AI transformation programs in TMS. Aggregate organizational adoption rate, the metric most commonly tracked in enterprise AI programs, is a misleading indicator of proximity to systemic disruption. A firm reporting 40 percent AI adoption may be far from disruption if that adoption is concentrated in parallel workflow clusters with no interdependence. A firm reporting 20 percent adoption strategically concentrated across a single critical workflow chain, covering quoting, tendering, carrier scoring, and exception management in freight brokerage, may be near the disruption threshold. The relevant measure is effective use rate within task-interdependent clusters, operationalized as the proportion of role-relevant task completions that incorporate AI-generated output, not access rate or training completion rate across the organization.

Sequential threshold crossing is the mechanism by which disruption propagates through a logistics organization. When a cluster in a workflow chain crosses its local adoption threshold, productivity gains in that cluster begin to compound rather than accumulate marginally, and the interface cost with the next cluster falls, because AI-assisted output from one cluster is more processable by an AI-assisted adjacent cluster than by a manual one. This makes the adoption threshold in the next cluster easier to reach, and so on through the chain. The 74 percent of firms BCG found failing to achieve scaled AI value are, on this account, characterized by uneven adoption across their workflow chains: above threshold in some clusters, sub-threshold in functionally linked others, with productivity gains absorbed at the interface bottlenecks rather than compounding toward systemic transformation. The organizational-level adoption rate these firms report, and frequently publicize internally as evidence of progress, tells them almost nothing about how close or far they are from the actual disruption threshold.

Implications for vendors

For TMS product and strategy teams, the two-level framework generates distinct priorities at each stage. At the sector level, the primary vendor responsibility is reducing complementary asset acquisition costs, moving generative AI from a capability that customers must build around to one embedded in the product and supplied with the data it needs. Vendors with existing data networks have the clearest structural advantage: they can absorb the data infrastructure burden that most TMS customers cannot, and their scale in carrier data aggregation directly raises the accuracy floor of embedded AI features. Platform-embedded AI is the mechanism through which the sector-level adoption barrier falls. Vendors who invest in data onboarding support alongside AI feature development address the complementary asset deficit more completely than those who deliver AI features and leave the data preparation work to the customer.

At the intra-organizational level, TMS vendors who help customers think about deployment sequencing across workflow-interdependent clusters will produce better measured outcomes than those who optimize for broad-license adoption. Deploying generative AI into freight brokerage end-to-end, covering quoting, tendering, carrier scoring, and exception management before expanding to adjacent functions, produces cluster-level threshold crossing in the highest-ROI workflow. Deploying it broadly across the firm at low penetration within each cluster produces many individual productivity improvements that fail to compound. The customer success model for enterprise TMS AI should be built around measuring effective use within workflow clusters and sequencing cluster-by-cluster saturation rather than maximizing license penetration across the organization, which is the current standard. Practically, this means that the first question in any enterprise deployment conversation should be a workflow map, not a license count.

The regulatory dimension requires attention at the vendor level as well. When generative AI recommendations contribute to freight claims or compliance violations, the allocation of liability between vendor, customer, and AI system is unresolved under current legal frameworks, including those being shaped by the EU AI Act. TMS vendors who invest in explainability infrastructure, meaning the ability to reconstruct AI decision logic for insurance, legal, and accountability purposes, both reduce customer adoption risk and establish defensible positions under emerging regulatory requirements. In logistics operations where most decisions carry regulatory dimensions (Hours of Service compliance, customs and trade compliance, dangerous goods routing), explainability is a precondition for adoption at the core of the market. Vendors who design AI tools to generate auditable decision records, rather than black-box recommendations requiring human approval, solve the liability ambiguity for their customers and accelerate adoption in precisely the compliance-sensitive segments where potential value is highest.

When the cascade begins

The sector-level analysis predicts punctuated equilibrium rather than smooth S-curve diffusion: extended inertia, then a compressed adoption cascade. Abrahamson and Rosenkopf's institutional bandwagon model formally shows that when adoption returns are ambiguous and institutional pressure is high, organizations can remain in a non-adoption equilibrium even when most expect that adoption would be net positive, and that when the threshold is crossed, adoption accelerates discontinuously. The observational signature of this prediction in logistics AI would be flat or slow adoption growth through 2026 and 2027, followed by a sharp inflection of 15 to 25 percentage points in adoption rate within a single 12-month period. That inflection would be triggered by a sufficient number of enterprise shippers and large 3PLs publicly documenting generative AI ROI, which triggers normative legitimation from Gartner and professional bodies, which activates mimetic and normative isomorphic pressure across the industry, which drives TMS vendors to accelerate product investment, which lowers the adoption cost for the next tier. The cascade is self-reinforcing once initiated.

The intra-organizational analysis predicts something less visible but equally consequential. When cluster-level thresholds are crossed sequentially across a critical workflow chain, the productivity effects will not be marginal but compounding. The firms doing the quiet work of workflow-by-workflow deployment, achieving genuine adoption saturation in one interdependent cluster before moving to the next, will experience this compounding before firms that have achieved broad but shallow adoption. The lag between the two groups will not be explained by technology access or investment level but by the structure of how adoption was distributed across the workflow, which is a design decision made before deployment begins. This is, practically, a consequence of where cluster-level threshold dynamics sit in the organizational change process: not in the technology selection decision, not in the budget approval decision, but in the deployment sequencing decision that most enterprise AI programs treat as an afterthought.

Both thresholds are approaching. Whether the sector-level cascade initiates in 2027 or 2029 depends on variables that are genuinely uncertain: freight market conditions, vendor product timelines, regulatory developments, and the pace at which early adopters generate credible, comparable social proof. Whether a given firm crosses its intra-organizational disruption threshold depends on how its AI deployment is sequenced, which is within the control of the teams designing those programs now. The firms that understand both levels of the problem are positioned to act on the one within their control rather than waiting on the one that is not. The sector-level cascade, when it comes, will lift all boats; the intra-organizational threshold is earned cluster by cluster, and that work does not wait for the cascade.

References

Abrahamson, E., & Rosenkopf, L. (1993). Institutional and competitive bandwagons: Using mathematical modeling as a tool to explore innovation diffusion. Academy of Management Review, 18(3), 487–517. https://doi.org/10.2307/258906

Arthur, W. B. (1989). Competing technologies, increasing returns, and lock-in by historical events. The Economic Journal, 99(394), 116–131. https://doi.org/10.2307/2234208

BCG. (2024, October). AI adoption in 2024: 74% of companies struggle to achieve and scale value. Boston Consulting Group. https://www.bcg.com/press/24october2024-ai-adoption-in-2024-74-of-companies-struggle-to-achieve-and-scale-value

Brynjolfsson, E., Li, D., & Raymond, L. R. (2023). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889. https://doi.org/10.1093/qje/qjae045

Centola, D. (2010). The spread of behavior in an online social network experiment. Science, 329(5996), 1194–1197. https://doi.org/10.1126/science.1185231

Centola, D. (2018). How behavior spreads: The science of complex contagions. Princeton University Press.

Centola, D., & Macy, M. (2007). Complex contagions and the weakness of long ties. American Journal of Sociology, 113(3), 702–734. https://doi.org/10.1086/521848

Dell'Acqua, F., McFowland, E., III, Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality (HBS Working Paper No. 24-013). Harvard Business School. https://www.hbs.edu/faculty/Pages/item.aspx?num=64700

DiMaggio, P. J., & Powell, W. W. (1983). The iron cage revisited: Institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), 147–160. https://doi.org/10.2307/2095101

G2. (2025). AI adoption in 2025: 79% of businesses prioritize AI in software. https://learn.g2.com/ai-adoption

Hiatt, J. M. (2006). ADKAR: A model for change in business, government and our community. Prosci Learning Center Publications.

Hirshleifer, J. (1983). From weakest-link to best-shot: The voluntary provision of public goods. Public Choice, 41(3), 371–386. https://doi.org/10.1007/BF00141070

IBM Newsroom. (2024, January 10). Data suggests growth in enterprise adoption of AI is due to widespread deployment by early adopters. https://newsroom.ibm.com/2024-01-10-Data-Suggests-Growth-in-Enterprise-Adoption-of-AI-is-Due-to-Widespread-Deployment-by-Early-Adopters

Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 263–291. https://doi.org/10.2307/1914185

Kotter, J. P. (1996). Leading change. Harvard Business Review Press.

McKinsey & Company. (2024). *Digital logistics: Into the express lane?*https://www.mckinsey.com/capabilities/operations/our-insights/digital-logistics-into-the-express-lane

MDPI Sustainability. (2025). Factors leading to the digital transformation dead zone in shipping SMEs: A dynamic capability theory perspective. Sustainability, 17(12), 5553. https://www.mdpi.com/2071-1050/17/12/5553

Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. https://doi.org/10.1126/science.adh2586

Rogers, E. M. (2003). Diffusion of innovations (5th ed.). Free Press.

Teece, D. J. (1986). Profiting from technological innovation: Implications for integration, collaboration, licensing and public policy. Research Policy, 15(6), 285–305. https://doi.org/10.1016/0048-7333(86)90027-2

Trimble. (2026). Transportation pulse report 2026: Transportation industry at AI inflection point as adoption accelerates. https://news.trimble.com/Transportation-Pulse-Report-2026-Transportation-Industry-at-AI-Inflection-Point-as-Adoption-Accelerates