The Roles AI Is Quietly Rewriting (And What Leaders Should Do About It)

A BCG and Harvard Business School field experiment found that consultants using GPT-4 completed 12.2% more tasks, worked 25.1% faster, and produced more than 40% higher-quality results on tasks that fell within the model's capable range. On a complex managerial task outside that range, AI made participants 19% less likely to produce correct solutions.

Most leaders ignore that second number.

AI does not uniformly improve performance across a team. It improves performance on specific categories of work and actively degrades it on others. Organizations figuring this out are not just "using AI more." They are auditing which tasks belong to which category, redesigning roles accordingly, and quietly shipping a version of their team that looks and operates differently than it did eighteen months ago.

Organizations that are not figuring this out are running the same job descriptions they wrote in 2022 and wondering why their AI pilots keep stalling before they reach production.

The Assembly Layer Is Already Gone

Work AI has absorbed fastest is the work that was always the most time-consuming and the least intellectually demanding: drafting a first version, summarizing a long document, pulling data into a report, formatting findings into a slide, writing meeting notes. Junior and middle layers of most teams spent a material portion of their week on exactly this work.

That assembly layer is not disappearing from org charts yet. But it is disappearing from the daily experience of the role. And when the task composition of a job changes, the hiring profile should change with it. Skills that made someone good at the old job are not the same skills that make someone good at the new one.

In 2026, the Federal Reserve Bank of New York published a job-postings analysis that found little sign of a broad, AI-driven labor demand collapse at the aggregate level. Labor markets have not cratered. But the research also documents that roles with higher AI-exposed task content are under more pressure at junior levels. Change is happening inside the role before it shows up in headcount numbers.

That dynamic is what makes the shift easy to miss. Org charts look the same. Titles are the same. Budget for that headcount is approved and spent. But the actual work the person is doing has changed, and in many cases, performance expectations have not caught up.

What the Research Actually Says About Task Fit

Ethan Mollick, whose work informed the BCG and Wharton-connected field research, has been consistent on a point most AI coverage glosses over: performance gains from AI are real and large, but they are task-conditional. Inside the frontier, the tool helps considerably. Outside it, the tool actively misleads.

For leaders, the implication is not "use AI" or "don't use AI." It is: know your frontier.

Tasks that sit inside the frontier for current AI tools include drafting and rewriting, summarizing documents, classifying and extracting structured information, building code from a description, and supporting multistep reasoning in bounded, well-defined problems. Gains in this zone are documented and repeatable.

Tasks outside the frontier include making correct judgments on novel or high-stakes problems, knowing when the output is wrong without an external check, and carrying accountability for decisions with legal, financial, safety, or strategic consequences. In this zone, AI assistance can reduce the quality of output by making the human less likely to think critically about what they are reading.

Treat this as a design constraint to build around, not a flaw to work around. Job descriptions, workflows, and review processes should reflect which zone each task sits in.

How This Changes the Actual Job

Looking at a single role before and after is the most concrete way to see this.

A junior analyst role used to center on producing first drafts from scratch: pulling data, building reports, summarizing findings, preparing meeting notes, answering ad hoc requests. Speed at these tasks was the primary performance signal.

In an AI-enabled team, that same role centers on something different. First drafts come from the tool. Human work is to verify the output against source data, flag anomalies, document assumptions, build reusable templates and prompts so the next version runs faster, and escalate anything that requires a judgment call to a manager who can carry the accountability for it.

Hiring signals have changed. Not "can you produce a report quickly" but "can you catch a plausible-looking error, and do you know when to escalate."

Research from the field experiments also shows that lower-baseline workers often gain disproportionately from AI assistance, narrowing performance gaps in some settings. Worth noting for teams that have historically struggled to develop junior talent: AI assistance may compress the time it takes a new hire to reach competence on bounded tasks, freeing managers to focus development conversations on judgment rather than mechanics.

But this only works if the role is designed to use that assistance well. A job description written for a world without AI will not produce the behaviors you need in a world with it.

The Frameworks Worth Applying Here

Patrick Gilbert covers the structural version of this in Never Always, Never Never, particularly in the chapters on the Resource Gap and what he calls the second strand of the AI Double Helix framework. First strand is internal efficiency: using AI to do existing work faster and cheaper. Second strand is external value: using AI to do things your team could not do before.

Most role redesign conversations are first-strand conversations. They are about efficiency: the junior analyst can do more in less time. That is real, and it matters.

Second-strand thinking is where the organizational design question gets more interesting. If AI can draft, summarize, report, and route, then human capacity previously consumed by those tasks is now available for something else. What that something else should be, and whether your current role structure captures it or wastes it, is the question worth answering.

Gilbert's position is direct: bandwidth is no longer a permanent constraint in the way it was. Gaps that used to require headcount to fill can now be filled differently. Leaders who are honest about the gaps in their programs have fewer excuses than they did before, and more decisions to make.

The AI resource gap framework applies here: the constraint was never ambition, it was capacity. AI changes the capacity side of that equation. AdVenture Media (adventuremedia.ai) applies this framework directly with clients navigating role redesign, particularly in the pilot-to-production transition.

To locate where your team sits on this spectrum, the AI Maturity Ladder is worth reading alongside this piece.

The Pilot-to-Production Problem

McKinsey and Deloitte reporting from 2024 through 2026 consistently identifies a gap between AI experimentation and AI production. Firms experiment at high rates. Fewer reach a state where the work is embedded in a real business process with a clear metric and an accountable owner.

Model capability is not the failure mode. Operating change is.

When a pilot succeeds, it usually succeeds because the model can do the task. When the same tool fails to reach production, it fails because no one owns the workflow, the data is inconsistent, the review process is undefined, the guardrails are absent, and the measure of success is tool adoption rather than business output.

Relevance to role design is direct. Deploying an AI tool and telling your team to "just use it" does not change the role. It adds a tool to an unchanged role. Behaviors you need, specifically verification, prompt hygiene, workflow documentation, and escalation judgment, will not appear on their own.

Role redesign that actually works tends to have a few things in common. It is specific about which tasks the AI handles and which tasks the human handles. It names a review standard, not just a review process. It measures output quality and cycle time, not tool adoption. And it assigns explicit ownership for the workflow, including what happens when the output is wrong.

Training that fails tends to be a one-off prompt workshop, detached from real work, with no policy, no quality control, and no one accountable for the outcome.

For leaders thinking through what this looks like in practice, the guide to building an AI-first team covers the structural moves in more detail.

> The pilot-to-production gap is not a technology problem. It is a role design problem dressed up as a technology problem.

The Artifact: A Role Audit You Can Run in 90 Minutes

The following is a working template. You can run it with a single team or use it as a structure for a broader org review. Copy it verbatim or adapt it for your context.

---

AI Role Audit: Task Composition Review

What you need: One team, one hour with the team, thirty minutes to synthesize.

Step 1: List the recurring tasks in this role (15 minutes)

Ask the person in the role to list every task they do at least once per month. Do not filter. Get the full list.

Step 2: Sort each task into one of three columns

• AI-owned: AI produces the first output, human reviews it. (Drafting, summarizing, classifying, formatting, first-pass analysis)
• Human-owned: Human does the thinking and produces the output. AI may assist with one step but is not driving. (Judgment calls, exception handling, stakeholder relationships, strategy decisions)
• Unclear: Neither of you knows yet where this task sits.

Step 3: For every task in the AI-owned column, answer these questions

• What is the review standard? How does the human know the output is correct?
• Who is accountable when the output is wrong?
• Is there a prompt, template, or workflow that makes this task repeatable? If not, who builds it?
• What does "done" look like, and can it be measured?

Step 4: For every task in the Unclear column, run a quick frontier test

Ask: Is this task bounded and well-defined, or does it require judgment about novel situations? Bounded tasks likely belong in the AI-owned column with the right guardrails. Judgment tasks belong in the human-owned column.

Step 5: Identify the capacity that is being freed

Tally the hours per week currently spent on AI-owned tasks. That capacity is now available. Name specifically what the human should be doing with it. If you cannot name it, you have a strategy gap, not a role gap.

Step 6: Rewrite the role description

A revised role description for a role with significant AI-owned tasks should include:

• What AI tools are approved and expected to be used
• The verification standard for AI-generated outputs (what checks are required before a deliverable is shared)
• What the human is accountable for that AI cannot be (judgment calls, escalation decisions, stakeholder communication)
• How performance is measured (output quality, cycle time, error rate, not tool adoption)
• Who owns the prompt library and workflow documentation for this role

Step 7: Set a review date

AI capability is changing. What is outside the frontier today may be inside it in six months. Schedule a task composition review at the six-month mark.

---

Questions to bring to your next leadership meeting:

• Which of our current roles are still designed for a world without AI tools?
• Where are we measuring tool adoption instead of output quality?
• Who in the organization owns the review standard for AI-generated work?
• What are we now capable of that we could not afford to do eighteen months ago? Are we doing it?

---

What to Do First

Pick one role on your team, ideally the one with the highest proportion of drafting, reporting, and assembly work, and run the audit above. Do not start with a policy, a training program, or a company-wide initiative. Start with one role, one task list, and honest answers to the questions in Step 3.

Your job description rewrite is the artifact. Everything else is preparation for it.

If you want to see where your organization sits before you run the audit, the AI Maturity Ladder post gives you a clear starting point. And if your AI rollouts keep stalling before they reach production, the analysis in why AI rollouts fail is worth reading first.