Your AI Pilots Won't Move the P&L. Here's What Will with Adam Nahari, Apollo

Successful AI transformation that actually moves the P&L is driven far more by change management than by engineering - picking the right high-value problem,

Abstract

This synthesis examines why enterprise generative AI adoption, despite reaching the fastest technology diffusion rate on record, has produced measurable profit-and-loss (P&L) impact in only a small fraction of firms. Drawing on direct practitioner evidence from enterprise AI transformations, the analysis argues that the binding constraint is organizational rather than technical: successful transformations allocate roughly five units of change management for every unit of engineering. A three-layer taxonomy distinguishes general productivity tools, domain-specific copilots, and core operational AI embedded in primary business processes, with margin retention concentrated in the third layer. An airline crew-scheduling case study illustrates end-to-end workflow redesign, and a six-move operational framework - anchored by CEO commitment, co-design with business owners, and a three-gate kill-metric system - is presented as a governance model for capital allocation under uncertainty.

1. Introduction

Generative AI has diffused through the US population faster than any prior general-purpose technology. More than 55% of US adults report having used generative AI within the last three years, compared with approximately 20% for the personal computer and 30% for the internet at equivalent points in their adoption curves. Enterprise uptake mirrors this pattern: nearly 90% of businesses report using AI in at least one business function.

This breadth of adoption, however, has not translated into commensurate financial return. Only 6% of firms report significant impact on EBIT (earnings before interest and taxes), and aggregate reported time savings amount to roughly 2% of total US working hours - gains that largely fail to appear in corporate P&L statements. This divergence between usage and financial outcome is the central puzzle this analysis addresses: if the tools function and are widely used, why does value capture remain so rare?

The thesis advanced here, based on direct observation of multiple enterprise AI transformations, is that the determinant of P&L impact is not the sophistication of the AI system but the quality of the organizational work surrounding it - problem selection, workflow redesign, executive sponsorship, co-design with end users, and disciplined measurement. The analysis proceeds through a taxonomy of enterprise AI layers, a detailed case study in airline operations, and a six-move framework with explicit governance gates.

2. Background and Related Work

The relevant historical precedent is economist Robert Solow's 1987 productivity paradox, which observed a lag between information-technology investment and aggregate productivity growth:

You can see the computer age everywhere but in the productivity statistics.

That lag was later attributed in substantial part to the organizational complements required to realize technology's value - process redesign, training, and incentive alignment - rather than to any deficiency in the technology itself. The present evidence suggests generative AI is reproducing this pattern at greater scale and speed, with the same underlying mechanism: technical capability diffuses faster than organizational absorption capacity.

A related concept concerns competitive dynamics under commoditized inputs. When a capability is universally purchasable, it cannot by itself sustain differentiated financial returns, since competitors capture the same capability and gains are competed away into price rather than retained as margin. This reframes absorptive capacity - an organization's ability to redesign its own processes around a new capability - as the scarce resource, not model access itself.

3. Core Analysis

3.1 The Change Management to Engineering Ratio

Across transformations observed directly, engineering consistently represented the smaller share of total effort required to realize financial impact. The approximate ratio is five units of change management to one unit of engineering, and this ratio is reported to be widening further as underlying models improve, since model improvement reduces the engineering burden while leaving organizational work largely unchanged. The practical implication is that the majority of transformation risk and effort lies in problem selection, buy-in, adoption, and measurement - areas where most observed failures originate, not in model performance.

3.2 Three Layers of Enterprise AI

A three-layer taxonomy clarifies where financial returns are retained versus competed away:

• Layer 1 (general productivity): copilots and chatbots available to any employee or competitor. Because capability is undifferentiated, gains tend to be competed away into price rather than retained as margin.
• Layer 2 (domain-specific AI): business-function copilots and retrieval systems operating over proprietary enterprise documents, enabling faster and deeper use of internal knowledge.
• Layer 3 (core operational AI): vertical systems embedded directly in core business processes (e.g., crew scheduling). Gains here remain in margin because the required change management cannot be purchased off the shelf - it is specific to the firm's own workflows and data.

Most businesses invest only in Layer 1, which correlates with the low observed EBIT impact. Firms reporting material EBIT impact are three times as likely to have redesigned an entire workflow end-to-end rather than simply adopting a new tool, indicating that depth of organizational integration, not tool sophistication, differentiates financial outcomes.

3.3 Case Study: Airline Crew Scheduling

An airline crew-scheduling transformation illustrates Layer 3 in practice. Under the prior manual process, an aircraft failure required rescheduling crew within 15 minutes while satisfying rest rules, qualifications, and location constraints. Schedulers frequently found no legal solution within the time window, triggering premium pay as an operational release valve. Crew costs represent the second-largest cost driver for airlines after fuel, making this a P&L-relevant problem rather than a convenience improvement.

The redesign consolidated previously siloed company data - flights, crew, and maintenance records - into a unified system, built an optimization agent engine that evaluates all applicable constraints and ranks the full set of legal solutions by cost in under a minute, and layered a decision-support application on top for scheduler use. The outcome compressed decision latency from 15 minutes to under one minute, directly reducing premium-pay exposure. The redesign spanned data infrastructure, optimization engineering, and interface design - underscoring that the technical component, while necessary, was embedded within a substantially larger organizational and data-integration effort.

3.4 Six Moves for Successful AI Transformation

Six operational moves are presented as the mechanism by which change management is executed in practice:

1. Pick one or two problems, not forty. Selection should start from the P&L and work backward, applying criteria including: the problem moves a line item the CFO recognizes; it is measurable against an agreed-upon baseline set before work begins; technical and data infrastructure are ready; the company owns the underlying data; a named champion user exists; and the problem ranks among the CEO's top five priorities.

2. Reimagine the process end-to-end. This distinguishes the 10% play (incremental automation of an existing process) from the 10x play (full redesign), with evidence favoring the latter for material EBIT impact.

3. Start with the CEO. Initiatives should not receive funding absent explicit, public CEO commitment; this factor correlates strongly with transformation success. Without it, teams encounter scheduling conflicts, slow data access, and misaligned priorities.

4. Build with the business. The P&L owner co-sponsors the initiative, and AI champions (actual end users) are embedded in solution design from inception, meeting on a near-daily cadence, co-designing interfaces, and running A/B tests. Leadership should combine technical and domain expertise.

5. Put tools in everyone's hands. Stakeholders must personally use AI tools to build intuition for where value resides, since workflow redesign is not feasible without hands-on experimentation:

   It's impossible to redesign a workflow with AI if you've never used AI.

6. Set kill metrics first and let the CFO keep score, discussed in detail below.

4. Technical Insights

The kill-metrics framework operationalizes capital discipline through three sequential gates: a time gate limiting each stage to 90 days, a metric gate requiring A/B-tested performance against the pre-agreed baseline, and an adoption gate requiring that over 50% of target users actively adopt the system. A supplementary check tests whether cost at scale - compute, token consumption, and personnel - remains significantly lower than the projected financial impact, guarding against solutions that are technically functional but economically unsustainable once deployed broadly.

A key implementation consideration is sequencing: the baseline metric must be agreed upon before the pilot begins, not retrofitted afterward, since post-hoc baselines are vulnerable to selection bias and undermine the credibility of reported gains. The crew-scheduling case further suggests that data consolidation (unifying flights, crew, and maintenance records) is frequently a prerequisite engineering step distinct from the optimization logic itself, and should be budgeted and gated separately.

5. Discussion

The evidence presented reframes the AI productivity paradox as a capacity-allocation problem rather than a technology-maturity problem. Enterprises spending primarily on Layer 1 tooling are, in effect, purchasing commoditized capability that competitors can replicate, which explains why usage statistics and EBIT statistics diverge so sharply. The firms capturing margin are those directing effort toward Layer 3 integration, where the organizational redesign required is specific to the firm and cannot be purchased as a product.

This has implications for how AI investment is currently evaluated within organizations. Metrics centered on tool adoption or user satisfaction at Layer 1 may systematically overstate progress toward financial impact, while the six-move framework and kill-metric gates suggest that governance structures borrowed from capital-intensive engineering projects - staged funding, baseline-gated evaluation, explicit executive sponsorship - are underused in AI initiatives relative to their apparent necessity.

An open question concerns scalability of the 5:1 change-management ratio across firm size and industry; the airline case involves a large, capital-intensive, safety-regulated operation, and it remains unclear whether comparable ratios and gating structures hold in smaller organizations with less formalized data infrastructure.

6. Conclusion

The central contribution of this analysis is the empirical reframing of enterprise AI value capture as predominantly an organizational, not technical, challenge, quantified at an approximate 5:1 ratio of change management to engineering effort. The three-layer taxonomy and the crew-scheduling case jointly demonstrate that margin is retained where workflows are redesigned end-to-end around proprietary data and processes, rather than where generic tools are layered onto existing processes unchanged.

For practitioners, the practical takeaway is to treat AI initiatives as governance problems first: select a small number of CEO-prioritized, CFO-measurable problems, secure executive and business-unit co-ownership before engineering begins, and enforce staged kill-metric gates rather than open-ended pilots. These mechanisms, more than any particular model or architecture, appear to determine whether an AI initiative moves the P&L.

Sources

• Your AI Pilots Won't Move the P&L. Here's What Will - Adam Nahari, Apollo - Original Creator (YouTube)
• Analysis and summary by Sean Weldon using AI-assisted research tools

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub