Development economics has spent three decades building an impressive arsenal of proven interventions. We know conditional cash transfers reduce poverty. We know deworming improves school attendance. We know chlorine dispensers at water points increase treatment rates. The randomized controlled trial revolution delivered exactly what it promised: rigorous evidence about what works.

Yet here's the uncomfortable reality that keeps program designers awake at night. Interventions that demonstrated remarkable effects in carefully controlled trials routinely produce disappointing results when scaled through government systems or large NGO operations. The evidence-to-impact pipeline leaks badly, and we've been slow to diagnose why.

The problem isn't that the evidence was wrong. The problem is that knowing what works tells us remarkably little about how to make it work in messy, resource-constrained, politically complex real-world settings. Implementation science—a field that emerged primarily in healthcare—offers frameworks for understanding this gap. Adapted thoughtfully for development contexts, these frameworks reveal why our most promising interventions fail at scale and, more importantly, what we can do about it. The distance between a successful pilot and effective national policy isn't just a matter of money and political will. It's a technical challenge that demands its own rigorous analysis.

The literature documenting implementation failure in development has grown uncomfortably large. Consider the trajectory of microfinance. Early RCTs showed promising effects on business investment and household consumption. Scaled globally, the sector now serves hundreds of millions—yet systematic reviews find average impacts on poverty that hover around zero. The intervention didn't change. The implementation context did.

Three mechanisms drive most evidence-practice gaps. Fidelity erosion occurs when the core components that made an intervention effective get diluted or dropped during scale-up. A teacher training program might show strong effects when delivered by expert facilitators over twelve sessions. Implemented by district education officers with two days of training and compressed to four sessions, the program retains its name but loses its active ingredients.

Then there's context mismatch. Interventions are developed and tested in specific settings with particular organizational capacities, cultural norms, and complementary services. An agricultural extension program that worked brilliantly in a region with functioning input markets and reliable rainfall may transfer poorly to areas lacking these enabling conditions. We often treat context as noise to be controlled for rather than as a fundamental determinant of effectiveness.

Finally, capacity constraints at every level of implementation systems create bottlenecks that trial conditions rarely encounter. Government health workers managing sixty patients per day cannot deliver the counseling protocols that worked when community health workers had time for hour-long home visits. Procurement systems that take eighteen months to approve purchases cannot support programs requiring rapid stock replenishment.

The uncomfortable implication is that effect sizes from trials should be understood as upper bounds on what real-world implementation can achieve. Voltage drop—the systematic reduction in impact when moving from controlled conditions to routine delivery—appears to average 50-80% across development interventions. We've been implicitly assuming implementation is a solved problem. It isn't.

Takeaway

The effect size from a trial represents what's possible under ideal conditions, not what's likely under real ones. Assume substantial voltage drop and design accordingly.

Implementation science offers structured frameworks for understanding why interventions fail to translate. The most useful for development contexts organize barriers across multiple levels: the intervention itself, the individuals delivering it, the organizations housing it, and the broader system surrounding it.

At the intervention level, complexity matters enormously. Interventions requiring many sequential steps, coordination across multiple actors, or sophisticated judgment calls are inherently harder to implement with fidelity. A simple technology like oral rehydration salts spreads readily; integrated community case management requiring diagnosis across multiple conditions and treatment algorithms struggles even with extensive support.

Provider-level determinants include knowledge, skills, motivation, and behavioral patterns. Development programs routinely underinvest in understanding frontline worker constraints. A health worker may understand a protocol perfectly yet fail to follow it because doing so conflicts with patient expectations, takes time she doesn't have, or contradicts guidance from supervisors. Changing provider behavior requires understanding what drives current practice, not just what evidence supports.

Organizational determinants shape what's possible within implementation settings. Leadership engagement, resource allocation, supervision systems, and organizational culture all influence whether new practices can take root. Many development interventions attempt to work around weak organizations rather than through them—creating parallel systems that work temporarily but cannot sustain beyond project periods.

System-level factors include policies, financing mechanisms, coordination structures, and political dynamics that create the environment for implementation. An immunization program cannot succeed if vaccine supply chains are unreliable, regardless of how well-designed the community mobilization component might be. Implementation diagnosis must extend beyond the immediate program to examine the enabling conditions that allow effective delivery.

Takeaway

Before asking whether an intervention works, ask whether your implementation system has the capacity to deliver it. Diagnosis across multiple levels reveals where investments in implementation support will yield the highest returns.

A central tension in implementation science concerns how much programs can be modified to fit local contexts without losing their effectiveness. Strict fidelity approaches argue that interventions should be delivered exactly as tested—deviation risks eliminating the active ingredients that produced effects. Adaptation approaches counter that rigid protocols ignore important contextual variation and may reduce acceptability and sustainability.

The resolution lies in distinguishing core components from adaptable periphery. Core components are the essential elements theorized to produce effects—the mechanisms of action. Adaptable elements are delivery features that can be modified without compromising function. A conditional cash transfer's core component is linking payment to verifiable behavior. The specific conditions, payment amounts, verification systems, and delivery mechanisms can be adapted to local contexts.

This requires something development programs rarely do: articulating theory of change at the mechanistic level. Why does this intervention produce effects? What are the necessary conditions for those mechanisms to operate? Without clear mechanistic understanding, implementers cannot distinguish essential from incidental features, leading either to inappropriate rigidity or modifications that inadvertently remove active ingredients.

Adaptive implementation frameworks structure how modifications should be made. They specify who has authority to adapt, what evidence triggers adaptation decisions, how adaptations are documented and tracked, and when accumulated adaptations require re-evaluation of effectiveness. This isn't permission to change things ad hoc. It's a disciplined process for learning what the intervention requires in each context.

The endpoint is what some call dynamic sustainability—ongoing alignment between intervention design, implementation context, and evidence. Neither the intervention nor the context is static. Effective implementation requires continuous monitoring of both fit and fidelity, with structured processes for making and documenting changes. This is harder than one-shot program design, but it reflects the actual work of making evidence-based interventions work at scale.

Takeaway

Distinguish what an intervention must do from how it does it. Protect core mechanisms while adapting delivery features—but this requires understanding why your intervention works at a mechanistic level.

The randomized trial revolution gave development economics unprecedented confidence in identifying effective interventions. Implementation science offers the complementary toolkit for understanding how to deliver them effectively under real-world constraints.

This requires shifting how we think about evidence-based practice. Demonstrating efficacy under controlled conditions is necessary but insufficient. Understanding implementation determinants, building adaptive delivery systems, and monitoring fidelity and fit become core competencies—not administrative afterthoughts.

The returns to getting implementation right are substantial. If voltage drop typically reduces impacts by half or more, improving implementation may deliver better cost-effectiveness than discovering new interventions. The evidence we need already exists. The challenge is building systems capable of using it.