Envision —
The Jagged Edge

In 2023, researchers from Harvard Business School sent 758 consultants at Boston Consulting Group to work on a set of real business tasks — some with access to GPT-4, some without. The results, published as "Navigating the Jagged Technological Frontier" (Dell'Acqua, Mollick et al., HBS Working Paper 24-013), produced a finding that should sit at the centre of every enterprise AI strategy: tasks that fell inside the AI's capability frontier showed a 25.1% improvement in quality and speed. Tasks that fell outside the frontier showed the opposite — AI-assisted consultants performed worse than those working without it.

Ethan Mollick, one of the paper's authors, named what he found: the jagged frontier. The edge of AI capability is not a smooth line. It is irregular, counter-intuitive, and invisible from where you are standing. A task that looks hard — writing a nuanced investor letter — may sit deep inside the frontier. A task that looks easy — checking a spreadsheet against an external source — may sit well outside it, because AI hallucinates confidently where it lacks grounding.

The consequence for enterprise leaders is not abstract. If you deploy AI broadly without mapping this frontier, some of your people will become significantly more productive. Others will become confidently wrong — producing polished output that contains errors their judgement would previously have caught. The 25.1% gain and the hidden degradation can coexist within the same team, sometimes within the same role.

What Mollick calls the "jagged edge" is not a fixed property of any AI system. It shifts with every model update, every new capability, every new use case. The frontier that existed in 2023 is already different from the one that exists today. This is not a reason for despair — it is a reason to build learning into the structure of your deployment, not just the launch phase.

The key practical insight from the BCG study is this: the consultants who were told to use AI for every task performed worse than those given guidance on which tasks were AI-appropriate. Instruction about the frontier mattered more than access to the tool. Your job as a leader is not just to license the technology — it is to help your people understand where the edge is, and to build the structures that let them discover and share what they find.

Before you can map the AI frontier in your organisation, you need a task inventory. Most organisations have job descriptions, competency frameworks, and performance objectives — but these are written at a level of abstraction that makes them useless for AI exposure analysis. "Manages stakeholder relationships" tells you nothing about whether that person spends their day writing briefing notes (Zone 2) or in face-to-face negotiation (Zone 1).

Two public datasets solve this problem. The first is O*NET (onetonline.org), maintained by the US Department of Labor. O*NET contains detailed task and activity data for over 900 occupations, updated continuously. Each occupation entry breaks down into specific work activities, tools used, skills required, and knowledge domains. The AI exposure literature has already used O*NET extensively — researchers at Goldman Sachs, Princeton, and MIT have each mapped O*NET tasks against LLM capabilities to produce exposure estimates by occupation.

The second is ESCO (esco.ec.europa.eu), the European Skills, Competences and Occupations framework. ESCO covers 2,942 occupations and 13,890 skills, with explicit competence-to-occupation linkages. It is multilingual, aligned to the European Qualifications Framework, and structured for machine readability. Crucially for UK and EU enterprises, ESCO is what underpins most national training frameworks — so an ESCO-based task map connects directly to the learning infrastructure your L&D team already uses.

The practical move is to pull the O*NET task list for each major occupation in your organisation, then review it with a subject matter expert to confirm which tasks are actually performed in your context and at what frequency. Many organisations have roles that deviate substantially from the standard O*NET profile — a "Marketing Manager" in a 15-person SME does things that an O*NET Marketing Manager in a 10,000-person enterprise does not. The framework gives you a starting point; local knowledge corrects it.

Once you have your task inventory, the next step is to assign each task to one of the four zones described in the next section. This is not a precise science — reasonable people will disagree about whether a given task belongs in Zone 2 or Zone 3. The discipline is in doing it systematically, documenting your reasoning, and updating the map as you learn.

This series uses a consistent framework for thinking about AI and work. Every task in every role can be placed in one of four zones. The zones are not a hierarchy — Zone 4 is not better or worse than Zone 1. They are descriptors of the appropriate relationship between human and AI for a given task. A Finance Business Partner who understands their zone map knows where to apply their energy, where to invite AI assistance, and where to stay firmly in control.

The framework is designed to be used at the task level, not the role level. A single role will contain tasks in all four zones. The proportion varies dramatically — a copywriter will have more Zone 2 tasks than a safety inspector, who will have more Zone 1 tasks. The map is the point, not a single number for the whole role.

The data on employee anxiety about AI is specific and uncomfortable. The CIPD's "People Profession 2024" report found that 45% of UK employees are worried that AI will change their role significantly. Only 18% say their employer has explained what AI means for them. That gap — 27 percentage points of unexplained, unaddressed anxiety — is not a communication failure. It is a strategy failure. You cannot communicate a position you have not taken.

Microsoft's 2024 Work Trend Index found that 75% of knowledge workers now use AI at work. The same report found that AI power users — those using it deeply across their working week — save an average of 30 minutes per day. Leaders, the report notes, feel significant pressure to deploy AI faster, but most lack a coherent plan for doing so. The technology has arrived faster than the strategy has been built.

There are three things that leadership must own, and that cannot be delegated to IT, HR, or an enthusiastic product team. The first is the governance model: who can use which AI tools for which purposes, what data can be processed, what cannot, and what the approval path is for new use cases. The second is the ethics position: what values govern AI use in this organisation — transparency with employees, fairness in deployment, human oversight of significant decisions. The third is the narrative: what does AI mean for jobs, careers, and futures in this specific organisation? Employees will fill the silence with their worst fears if leadership does not occupy that space.

Ethan Mollick argues in "Co-Intelligence" (Portfolio/Penguin, 2024) that treating AI as a brilliant but junior co-worker — capable but needing oversight, fast but not infallible — is the mental model that serves practitioners best. This framing is also the right one for leadership communication. It is honest about both the capability and the limitation. It positions the human as the senior partner, not the person being replaced. And it frames AI literacy as a professional skill to be developed, not a threat to be managed.

The World Economic Forum's "Future of Jobs Report 2025" projects that 170 million new roles will be created and 92 million displaced by 2030 — a net positive of 78 million jobs. Analytical thinking and AI literacy top the list of skills organisations say they need. This is not a reassuring statistic for the 92 million — but it is a clear signal that the organisations that build AI literacy systematically and early are the ones that will navigate this transition without the human cost of those who do not.

The destination is specific: every employee in your organisation has a zone map of their role and a personalised learning path that addresses their Zone 2 and Zone 3 gaps. Not a one-size-fits-all AI literacy course. Not a 45-minute module that covers "what is a prompt." A development arc built around the actual tasks this person does in this role, showing them exactly where AI can help and giving them the skills to use it well.

The LinkedIn "2024 Workplace Learning Report" named AI literacy as the fastest-growing skill on the platform. One in four L&D professionals is already using generative AI to create content. The report also found that learners who used AI-powered coaching were twice as likely to recommend their employer. Learning is becoming an AI-mediated experience, which creates an interesting feedback loop: the tool you are teaching people to use is the same tool that can teach them to use it.

The ESCO competence framework provides the connective tissue between role task mapping and the learning ecosystem. ESCO competences are structured, searchable, and aligned to national qualifications frameworks. If you map your Finance Business Partner tasks to ESCO competences, you can then query your learning platform for content that addresses those competences — and build a curated, personalised path that goes from "here are your Zone 2 tasks" to "here are the specific skills and micro-courses that will help you use AI on them."

This is the vision that this series builds towards across all eight issues. It is not simple — it requires task mapping, learning infrastructure, and ongoing measurement. But the components exist. What is missing in most organisations is the will to connect them deliberately rather than leaving each employee to work it out alone. The McKinsey Global Institute estimated in its June 2023 report that 60-70% of employee time is currently spent on tasks that AI could partially automate. That is not a threat to be managed — it is an enormous reservoir of recoverable time, waiting for an organisation with the strategy to reach it.