Why AI Pilots Keep Stalling Before They Scale
Boardrooms across the world have signed off on ambitious AI transformation programmes. Hundreds of pilots have been launched, productivity tools have been rolled out to entire workforces, and proof-of-concept demonstrations have impressed senior leadership. Yet the fundamental question remains unanswered: why aren't those gains showing up on the balance sheet?
The answer, according to a landmark closed-door summit convened at Harvard Business School, has almost nothing to do with the quality of the underlying technology. The bottleneck is organisational, not algorithmic. Researchers from the Frontier Firm Initiative (FFI), a joint endeavour between Harvard's Digital Data Design (D^3) Institute and Microsoft, have identified seven structural frictions that are preventing AI transformation from moving beyond isolated experiments into enterprise-wide operating models.
By The Numbers
- A global investment bank documented more than 250 applications connecting large language models to its enterprise systems, yet has not achieved organisation-wide transformation.
- A global payments network reported that more than 99% of its employees actively use AI copilots, but finance leaders struggle to identify where the productivity gains appear on the balance sheet.
- A global apparel firm automated more than 18,000 finance processes and still could not convert those wins into a standard operating model.
- A large asset-servicing institution is currently running more than 100 AI agents and is planning for a future involving tens of thousands.
- A global professional services firm operating in more than 170 countries found the same process being executed in dozens of different ways depending on geography.
The summit brought together senior leaders from roughly a dozen global organisations spanning healthcare, banking, and industrial manufacturing. These were not AI sceptics. They were enthusiastic early adopters who had deployed widespread access to tools like Microsoft 365 Copilot, ChatGPT Enterprise, and GitHub Copilot. Their frustration was not with the technology itself but with their inability to convert individual productivity wins into systemic business transformation.
The primary obstacle to progress is rarely model quality or data availability, but rather the 'last mile' of transformation where technical capability must meet organisational design." , Frontier Firm Initiative, Harvard Business School
The Seven Frictions Blocking AI Transformation
The FFI's research, combined with direct testimony from summit participants, identified seven recurring structural problems. Together, they explain why so many organisations find themselves what researchers describe as "pilot-rich but transformation-poor."
1. The Proliferation of Pilots
The absence of a repeatable path from proof-of-concept to standard operating model is the first major friction. A global food and beverage company successfully launched AI pilots across its 185-country footprint. An apparel firm automated thousands of finance processes. Yet neither could make those wins the default way the company operates. Localised success does not automatically propagate upward.
2. The Productivity Gap
Individual productivity improvements are real but routinely fail to materialise at the organisational level. Time saved by AI tools tends to be reabsorbed into low-value activities, such as additional internal meetings or unnecessary email chains, rather than being structurally redirected toward higher-value work. Without deliberate role reclassification and budget redesign, the productivity gains remain invisible to the finance team.
3. Process Debt
AI acts as a diagnostic tool that rapidly exposes brittle, exception-ridden processes accumulated over decades of acquisitions and organic growth. At one large healthcare insurer, workflows were so fragmented that AI surfaced inconsistencies faster than it could resolve them. Re-architecting the workflow itself, before deploying AI on top of it, requires what the FFI calls techno-functional leadership: people who understand both business logic and technical constraints well enough to redesign processes from scratch.
4. The Tribal Knowledge Identity Problem
Tacit knowledge held by long-tenured employees is frequently undocumented and, critically, is often protected because it confers professional status. An engineering consultancy at the summit framed this explicitly as an identity problem rather than a reskilling issue. For decades, expertise meant being the person who knew. AI now asks those same individuals to externalise their judgement and encode it into systems, a request that can feel existential rather than merely operational.
5. Governance in an Agentic World
Traditional governance models built around human-in-the-loop controls were designed for isolated, single-task automation. They collapse under multi-agent architectures where dozens or hundreds of AI agents are coordinating actions across systems simultaneously. A global bank described an emerging accountability gap that raises questions more reminiscent of human resources than IT: how do you onboard, evaluate, secure, and eventually retire a digital worker?
6. Architectural Complexity
Most enterprises operate a patchwork of AI capabilities across multiple cloud providers and application stacks. One global apparel company described spending months simply getting agents across SAP, Microsoft, and Google environments to communicate reliably. A large industrial manufacturer noted that platform evolution now outpaces project timelines, tempting teams to reset initiatives every time a more capable model is released.
7. The Efficiency Trap
Framing AI primarily as a cost-reduction tool has narrowed the ambition of many transformation programmes. Several summit participants compared the early positioning of AI to a new form of offshoring, which triggered defensive behaviour from middle management and constrained what the C-suite felt it could attempt. A data and analytics advisory firm at the summit warned that a relentless focus on efficiency risks hollowing out the human capabilities, such as judgement and storytelling, that differentiate high-value work.
The most significant gains are likely to come from rethinking value creation rather than merely shaving minutes off existing tasks." , Frontier Firm Initiative, Harvard Business School
The Blueprint for the AI-Native Firm
Despite these frictions, organisations making the most meaningful progress are converging on a shared operating model. The FFI has synthesised these into four core strategic shifts that together constitute what it calls the frontier firm blueprint. These are not optional tactics. They are the structural counterweights to the seven frictions described above.
Clean-Sheet Process Redesign
Leading firms have stopped bolting AI onto legacy workflows. Instead, they treat AI as a trigger for rebuilding processes from scratch. The key question is: if we were designing this process today with modern AI agents as first-class participants, what would we build? This approach maps both the strategic planning loop and the execution loop before a single line of code is written. AI agents are identified as workers; humans are designated as orchestrators. The hand-offs between the two are engineered to be seamless from the outset.
Strategic Knowledge Capture
Organisations that move quickly treat tribal knowledge as a strategic asset rather than a liability. They pair senior experts with process designers to systematically externalise accumulated judgement. Critically, this is framed not as replacement but as legacy-building: encoding unique expertise into digital systems that free experts from repetitive tasks to focus on the highest-complexity challenges the firm faces. New roles such as AI process architect and knowledge steward are being created to give this work organisational status and career trajectory.
Managing the Digital Workforce
Model governance is no longer sufficient. Forward-looking organisations are designing what the FFI calls agentic control planes: centralised dashboards that monitor agent performance, security permissions, and accuracy in the same way a traditional HR performance management system monitors human employees. The guiding principle is that agents must be treated as a managed workforce, not a collection of software scripts. This is directly relevant to the growing number of organisations that, like the asset-servicing institution at the summit, are already managing over 100 active agents and planning for tens of thousands.
Role Redesign and Career Pathing
Without explicit role redesign, AI adoption can feel like career erosion rather than evolution. Engineering firms are anticipating a shift toward assurance and systems-thinking roles. Advisory firms are formalising what they call storyteller capabilities to contextualise AI output for clients and stakeholders. Financial institutions are beginning to assign managers to digital workers just as they would to human teams. New job descriptions are being written around learning agility, domain depth, and agent orchestration as core competencies.
| Friction | Root Cause | Blueprint Response |
|---|---|---|
| Pilot proliferation | No repeatable path to scale | Clean-sheet process redesign |
| Productivity gap | Gains reabsorbed into low-value work | Role redesign and career pathing |
| Process debt | Legacy workflows incompatible with AI | Clean-sheet process redesign |
| Tribal knowledge | Identity threat to senior experts | Strategic knowledge capture |
| Governance gap | Human-in-the-loop fails at agentic scale | Agentic control planes |
| Architectural complexity | Multi-vendor fragmentation | Managed digital workforce model |
| Efficiency trap | AI framed as cost-cutting only | Value creation reorientation |
The Asia-Pacific Picture
The frictions identified by the FFI are not unique to Western enterprises. Across Asia-Pacific, organisations face the same last-mile AI transformation challenge, often compounded by additional regional dynamics.
In China, the government has placed AI at the centre of its national industrial strategy, with the latest five-year plan positioning AI as a core driver of economic modernisation. Yet even state-backed enterprises report the same pilot-to-scale gap that frustrated summit participants at Harvard. The ambition is immense; the organisational redesign required to realise it remains a work in progress.
Vietnam has moved faster on the regulatory dimension than almost any other market in the region. The country recently enforced Southeast Asia's first dedicated AI law, creating a compliance framework that will force businesses to address governance questions they might otherwise defer. This is a direct policy-level response to the agentic governance gap the FFI identified at the summit. Vietnam has also invested heavily in education, with AI being taught from primary school level to build the talent pipeline required for genuine transformation.
Japan and South Korea face a distinct version of the tribal knowledge problem. Both countries have deeply seniority-based corporate cultures where institutional knowledge is tightly held and professional identity is closely linked to expertise. The identity friction the FFI describes is arguably more acute in these markets than anywhere else. Nikkei Asia has reported extensively on the challenges Japanese corporations face in restructuring workflows around AI while managing workforce expectations and cultural norms around automation.
In financial services, Singapore's Monetary Authority has been among the most proactive regulators globally in issuing guidance on AI governance for banks and asset managers. The accountability gap around multi-agent architectures that the global bank at the FFI summit described is precisely the kind of issue MAS frameworks are beginning to address. The same architectural complexity problem affects regional banks operating across multiple Southeast Asian markets with fragmented technology stacks inherited from decades of localised operations.
The broader consumer-facing shift is also well underway. AI has already fundamentally changed how Asia shops, creating both the commercial urgency and the workflow disruption that make the last-mile problem so pressing for retail and e-commerce businesses across the region. For organisations already navigating this shift, the FFI's blueprint offers a practical framework rather than another abstract call to action.
The pace of AI-enabled software development is also reshaping expectations. Vibe coding is changing how software gets built across the region, compressing development timelines and raising the stakes for firms that cannot convert technical speed into organisational capability.
What Leaders Must Do Now
The FFI's core finding is unambiguous: the AI last mile is not blocked by technology. It is blocked by unresolved questions about operating models, governance structures, and human professional identity. The technology has already been purchased. The capability is largely available. What is scarce is the leadership willingness to commit to a different way of running the enterprise.
Organisations that are making genuine progress share several characteristics. They are redesigning processes from scratch rather than overlaying AI on legacy workflows. They are treating their most experienced employees as knowledge architects rather than potential redundancies. They are building governance systems designed for agents, not just for individual automation tools. And they are rewriting job descriptions around orchestration, judgement, and contextualisation rather than execution.
- Audit your pilot portfolio: identify which pilots have a defined path to standard operating model and which are experiments with no exit strategy.
- Map your process debt: before deploying agents on top of fragmented workflows, identify which processes need to be rebuilt, not just automated.
- Design for agents, not just automation: distinguish between tasks where AI acts as a worker and tasks where humans must act as orchestrators, and engineer the hand-offs deliberately.
- Reframe the knowledge conversation: position expertise externalisation as legacy-building and amplification, not replacement, to address the identity friction head-on.
- Build an agentic control plane: establish centralised oversight of agent performance, permissions, and accountability before agent numbers scale beyond manageability.
Frequently Asked Questions
What is the "last mile" problem in AI transformation?
The last mile problem refers to the gap between successful, localised AI pilots and enterprise-wide transformation. Most organisations have deployed AI tools and run hundreds of proofs-of-concept, but cannot convert those individual wins into a fundamentally redesigned operating model. The FFI research identifies seven structural frictions, including process debt, governance gaps, and tribal knowledge barriers, as the primary causes.
Why do AI productivity gains not show up on the balance sheet?
Time saved by AI tools is typically reabsorbed into low-value activities rather than being structurally redirected toward higher-value work. Without deliberate role redesign and budget reallocation to capture reclaimed time, the productivity gains remain trapped inside individual workflows and invisible to finance teams. The FFI calls this the productivity gap, and it requires intentional organisational redesign to resolve.
How should companies govern AI agents as their numbers scale?
The FFI recommends building agentic control planes: centralised dashboards that monitor agent performance, security permissions, and accuracy in the same way HR systems monitor human employees. As organisations move from dozens to potentially tens of thousands of active agents, governance must evolve from software management to workforce management, with defined processes for onboarding, evaluating, securing, and retiring digital workers.
The AIinASIA View: The FFI's research cuts through the hype with unusual clarity: the AI transformation gap is a leadership and organisational design problem, not a technology problem. For Asia-Pacific enterprises, the urgency is compounded by regulatory pressure from Vietnam, competitive pressure from China's state-backed AI push, and cultural dynamics around seniority and expertise that make the tribal knowledge friction especially acute. The blueprint is available. The question is whether enough executives are prepared to use it.
Given the frictions your organisation faces right now, which of the seven bottlenecks is hitting hardest, and what have you actually tried to break through it? Drop your take in the comments below.