AI-Wiki

How to Move from AI Experimentation to AI Transformation

8 min read

Dutt, Rapoport, Chatterji, Weeratunga & Satcher — How to Move from AI Experimentation to AI Transformation

TL;DR

A practitioner-oriented HBR article (30 April 2026) co-authored by Bain & Company partners (Arjun Dutt, Gene Rapoport) and OpenAI’s Economic Research team (Aaron Chatterji — Chief Economist; Gawesha Weeratunga; Harrison Satcher). Combines Bain’s enterprise-AI-deployment book with OpenAI’s vantage point as a provider to “over one million businesses.”

Core diagnostic: most firms get stuck in a “micro-productivity trap” — treating AI as plug-and-play SaaS with isolated use cases. Individual task gains do not translate to firm-level results because the surrounding workflow still depends on tacit knowledge, manual handoffs, or legacy systems.

Two failure modes:

  • Offering lock-in — using AI to optimize existing offerings.
  • Process lock-in — using AI to automate current processes without rethinking them.

Successful firms instead “reinvent the business”: organization-wide, future-focused, outcome-oriented. Bain client EBITDA gains: 10–25% for transformation-mindset deployments.

Four-step framework (with worked cases at Lowe’s + OpenAI partnership and a Fortune-1000 manufacturer the authors call FabricationCo):

  1. Narrow possibilities strategically (4–5 critical domains)
  2. Reimagine workflows across the organization
  3. Engage those closest to today’s process
  4. Measure what matters (concrete business outcomes, not “efficiency” — and continuous evals for AI behaviour)

Key claims

The micro-productivity trap

“Individuals often realize productivity gains on key tasks, but those gains often stall at the firm level when the surrounding workflow still depends on tacit knowledge, manual handoffs, or legacy systems not built for AI. AI can accelerate a task, but unless companies address workflow bottlenecks, productivity gains may not translate into business value.”

Two named lock-ins, often co-occurring:

Lock-inWhat it looks like
Offering lock-inAI applied to optimize the existing product/service rather than reframe what value the firm provides
Process lock-inAI applied to automate current processes step-by-step rather than rebuild around an outcome

The opposite posture — what the article calls the AI-transformation mindset — explicitly assumes “we live in a world in which powerful AI tools exist” and rebuilds workflows on that premise.

Four steps

1. Narrow possibilities strategically

  • Resist the urge to deploy AI everywhere; identify 4 or 5 critical domains.
  • Top-4 domains across Bain’s client work and decision-maker feedback: software development, customer support, knowledge worker efficiency, marketing. (Industry-specific selection still required.)
  • Diagnostic questions:
    • Which parts of the business will benefit most today and tomorrow (assuming continued capability progress)?
    • Where is the highest concentration of resources doing repeatable work?
    • What are high-value, low-effort use cases (often where workers are bottlenecked)?
    • Buy vs build? In-house vs partner?

FabricationCo case: week-long cross-functional workshop with frontline operators + managers across divisions. Mapped end-to-end workflows for both customer and internal roles. Surfaced 14 discrete AI use cases with tens of millions in aggregate potential value. Now on track to realize ~$30M in additional profit.

Lowe’s case: prioritization framework based on technology maturity, use-case size, readiness for change, and dimensional risk (including brand risk). Vision settled on as “democratizing expertise across the organization” — operationalized as two AI interfaces launched March 2025: Mylow (online customers) and Mylow Companion (in-store associates).

“How customers shop with us, how we sell, how we work” — Lowe’s broader framework, per Chandhu Nair (SVP of stores, data, AI, and innovation).

2. Reimagine workflows across the organization

“It’s the process redesign — not the technology — that is the most challenging part of AI deployment for firms, and also creates most of the value.”

Key argument: many seemingly department-specific work products are actually cross-functional. At FabricationCo, “even seemingly straightforward activities like quoting customer jobs required coordination across sales, design, and operational teams” — making them prime candidates for reinvention rather than incremental automation.

Selection questions:

  • Where is the highest value (time, effort, usage)?
  • Which processes are most ready (repeatability, data quality, technology)?
  • Where is variation across business units highest?

FabricationCo quoting workflow (worked example):

  • Old: design engineers spent several hours producing initial designs as input for every bid quote — including bids unlikely to convert (>50% in some segments).
  • New: 20-minute cost estimates by non-designers for early-stage bids; full engineered designs reserved for higher-probability opportunities.
  • Result: ~15× faster quote generation; reduced wasted effort; improved win rates.

3. Engage those closest to today’s process

  • Successful transformations engage both top-down and bottom-up perspectives.
  • Cross-seniority collaboration between leaders close to the work and outstanding individual contributors who can re-envision the process.
  • Prototyping culture in all areas (not just the tech team) — hackathons, micro-sprints. Even simple wireframes can convert sceptical/anxious team members.

FabricationCo: pilot regions chosen for strategic importance + first-mover willingness; visible successes in pilots generated confidence in not-yet-deployed regions; demos and feedback sessions extended to non-pilot regions to enable fast cross-region scaling.

Lowe’s Mylow Companion: targeted store pilots starting in 1–2 departments per store (e.g. plumbing or electrical) with in-app and on-floor feedback loops to refine prompts, AI guardrails, UX. Iterative expansion across departments and stores. Now fully deployed across 1,700+ Lowe’s stores.

4. Measure what matters

“‘Efficiency’ or ‘productivity’ are too vague; instead, those benefits should be tied to key business outcomes, with metrics that allow for comparisons with non-AI methods.”

Two measurement layers:

  1. Business-outcome metrics — comparable to non-AI baselines.
  2. Evals (continuous evaluation suites) — AI systems “like humans, can produce different results from the same request”; require sustained measurement against a tolerance range.

FabricationCo metrics:

  • Win rates (AI-generated quotes vs non-AI)
  • Quote turnaround time
  • Margins on downstream material/factory costs
  • Volume + accuracy of priced bids
  • Result: +10 percentage-point increase in win rate within 3 months of deployment.

Lowe’s evals approach:

  • SMEs created prompt + expert-validated response sets (“how Mylow Companion should reply”).
  • Evaluated multiple models against the validated set; iterated prompting.
  • Associates regularly audit both outputs and the system’s intermediate steps.

Lowe’s outcome metrics:

  • Drivers of business results: basket size, conversion rate, incremental sales.
  • Drivers of success: associate knowledge, customer satisfaction, ease of check-out.
  • Result: conversion rate more than doubles when customers engage Mylow online; customer satisfaction +200 basis points when associates use Mylow Companion.

The boardroom imperative

“AI transformation must be prioritized at the very top levels of the company, with all executives and functions participating, because leaders who have an organization-wide view are critical to transformative deployments.”

Failure pattern in lagging firms: leaders recognize AI as important but delegate it to technology groups/divisions without specific goals or metrics, just vague “improve productivity” or “focus on high-potential sources of value.” From that delegation, initiatives often fail.

Headline numbers

MetricValue
Bain client EBITDA gains (transformation mindset)10–25%
FabricationCo additional profit (on track)~$30M
FabricationCo quote-generation speedup~15×
FabricationCo win-rate increase (3 months)+10 pp
Lowe’s Mylow Companion deployment1,700+ stores
Lowe’s Mylow online conversion rate>2× (more than doubles)
Lowe’s Mylow Companion CSAT lift+200 basis points

Methodology notes

  • Practitioner-oriented HBR Generative AI piece; not an empirical study.
  • Author vantages: Bain enterprise-AI-deployment work + OpenAI’s view as a provider to >1M businesses + Duke business / public-policy academic perspective (Aaron Chatterji).
  • Two case studies: Lowe’s (named, with public OpenAI partnership) and FabricationCo (anonymized Fortune 1000 manufacturer, Bain client).
  • Acknowledgements: Bain’s Mike Coxon and Daan Kakebeeke contributed to writing.

Quotes worth saving

“AI can accelerate a task, but unless companies address workflow bottlenecks, productivity gains may not translate into business value.”

“It’s the process redesign — not the technology — that is the most challenging part of AI deployment for firms, and also creates most of the value.”

“[Successful firms] take an outcome-oriented approach, centering on what outcomes those processes serve and rebuilding their workflows starting from the premise that we live in a world in which powerful AI tools exist.”

“Going from lagging to leading requires that companies avoid siloed initiatives and enable instead a more wholesale transformation of their business.”

Graph View

Backlinks