An immersive study of a Kenyan financial firm finds that concrete organizational routines — spanning detection of AI opportunities, translating models into workflows, and embedding learning — form the backbone of successful Generative AI adaptation; deep, sustained internal processes, not ad-hoc pilot projects, drive durable adoption.

Main Finding

A 14-month ethnography of a Kenyan legacy financial firm (Kenya First Bank) identifies three concrete organizational routines that constitute the micro‑foundations of dynamic capabilities for Generative AI: (1) the AI Opportunity Radar (systematic sensing), (2) the Proof‑of‑Concept Sprint (disciplined seizing), and (3) the AI Integration Squad (continuous transforming). Together these interlinked routines create a repeatable Radar → Sprint → Squad pipeline that turns external signals into validated prototypes and then embeds solutions into operations, enabling reproducible adaptation to GenAI disruption.

Key Points

• Three routines and their roles:
  • AI Opportunity Radar: monthly 90‑minute, cross‑functional meeting (12–15 reps). Uses a standardized template (one external trend, one competitor action, one internal implication) and STEEP framing to convert dispersed observations into prioritized opportunities.
  • Proof‑of‑Concept Sprint: 30‑day, time‑boxed cycle (Week 1: problem definition; Weeks 2–3: solution build; Week 4: validation/presentation) with predefined go/no‑go criteria and mechanisms to capture learning.
  • AI Integration Squad: permanent, rotating 8–10 member team (6‑month rotations with two overlapping members) that embeds validated experiments into operations, documents integration patterns, and transfers capability to business units.
• Micro‑foundations highlighted: distributed sensing via diverse participants; standardized interpretation frames (STEEP); rapid, time‑bound validation to limit resource exposure and create progressive commitment; embedded, rotating teams to capture organizational memory and ensure diffusion.
• The three routines operate as an interconnected system: Radar feeds the Sprint pipeline; successful sprints are handed to the Integration Squad for operationalization—producing a self‑reinforcing dynamic capability.
• Practical orientation: routines are observable, repeatable, and codified (templates, SOPs, decision rules), offering a replicable “anatomy of adaptation” for legacy firms.
• Limitations noted by authors: single‑case study (external validity), snapshot in time (routines may evolve), and need to study contextual moderators (industry, size, leadership).

Data & Methods

• Research design: organizational ethnography (practice‑based perspective) conducted Jan 2023–Feb 2024 at a Kenyan bank undergoing active GenAI development.
• Access and immersion: first author ~120 hours of direct observation (45 hours formal meetings, 35 hours shadowing, 40 hours observing routine work and informal interactions).
• Interviews: 42 semi‑structured interviews using a “routine elicitation” technique to walk participants step‑by‑step through processes.
• Documents and artifacts: 86 internal documents analyzed (process manuals, SOPs, meeting minutes, project charters, templates, training materials) plus digital/physical artifacts and workflow diagrams.
• Analysis: grounded theory-style coding and mapping of routines (actors, actions, artifacts, patterns); triangulation across observations, interviews, and documents to identify micro‑foundations and routine interplay.

Implications for AI Economics

• Firm‑level productivity and adoption economics:
  • Reduces uncertainty and experimentation costs: time‑boxed sprints and standardized sensing lower search and evaluation costs for AI opportunities, improving the economics of early adoption in resource‑constrained firms.
  • Enhances returns to organizational capital: codified routines (templates, decision rules, integration patterns) convert tacit learning into reusable assets, increasing the marginal productivity of AI investments.
  • Enables scalable diffusion within firms: dedicated integration squads and embedded work reduce implementation frictions, accelerating realization of value from prototypes and affecting the pace of firm‑level productivity growth.
• Market structure and competition:
  • Creates persistent capabilities as competitive advantages: repeatable adaptation processes can generate first‑mover or second‑mover advantages that are harder for competitors to imitate quickly, potentially raising entry barriers.
  • Affects industry dynamics in emerging markets: demonstration that legacy firms in Africa can build advanced routines implies faster institutional diffusion of AI, with implications for competition between incumbents and digital natives.
• Labor and human‑capital effects:
  • Emphasizes complementarities between AI and organizational skills: returns to AI depend on managerial, process, and integration skills—policy and firm investments in training and routines matter for equitable gains.
  • Potential for task reallocation: routinized adoption can change job content systematically; research should measure effects on employment composition, wage premia for integration skills, and retraining needs.
• Policy and development implications:
  • Capacity building policies should go beyond hardware/software subsidies to support organizational capability formation (e.g., templates, playbooks, training for routine design).
  • Public programs supporting experimentation (sandbox funding, shared integration squads across SMEs) may lower collective adoption costs and generate positive spillovers.
• Suggested empirical economic research directions:
  • Quantify adoption economics: measure cost per prototype, time‑to‑decision, ROI of sprints, and productivity gains post‑integration.
  • Heterogeneity and causality: multi‑firm, cross‑industry studies to test generalizability and causal impact of routine systems on firm performance, survival, and market shares.
  • Labor market impacts: estimate how routine‑based GenAI integration affects employment, wages, and skill demand within and across firms in developing economies.
  • Spillovers and externalities: study whether codified integration patterns lead to sectoral knowledge spillovers and how intellectual property or confidentiality choices mediate diffusion.

Limitations to bear in mind when using this case for economic modeling: single‑firm context, qualitative measurement of routines, and potential selection bias (a firm willing to allow deep access may be atypically capable or motivated). Future quantitative work should operationalize the routines into measurable variables (e.g., frequency/attendance of Radar, % projects passing sprint, integration squad headcount and rotation rate) to test economic hypotheses.