Past the Question of Whether

To figure out the shape of the frontier, you will need to experiment.

Ethan Mollick, Co-Intelligence, Chapter 3

In Chapter 3 of Co-Intelligence, Mollick describes AI as a General Purpose Technology, like electricity or the internet, whose effects ripple across every domain it touches. Inside that ripple, he and his coauthors observe what they call the Jagged Frontier of AI. He uses the image of a fortress wall. Some towers and battlements jut out into the countryside. Others fold back toward the center of the castle. Everything inside the wall, AI can do. Everything outside, AI struggles with. The complication is that the wall is uneven and invisible. Two tasks that look equivalent from the outside (writing a sonnet versus writing an exactly fifty-word poem) sit on opposite sides of the wall. AI handles the sonnet. AI cannot reliably hit the fifty-word constraint because it conceptualizes language in tokens, not words. Without experimentation, carriers cannot tell which of their workflows fall inside the frontier and which do not.

What the Research Shows

The frontier is not just metaphor. Fabrizio Dell'Acqua, Ethan Mollick, and a team of researchers at Harvard Business School, MIT, and Warwick Business School ran a controlled experiment with the Boston Consulting Group involving nearly 800 consultants. Half worked with GPT-4. Half worked without. Inside the frontier, the AI-augmented group completed tasks 25% faster and produced work judged roughly 40% higher in quality. Consultants below the average skill threshold improved by 43%. Above-average consultants improved by 17%. AI compressed the skill gap. Then the researchers added a single task designed to look like the others but to require resolving contradictory data, the kind of task the AI could not solve. On that task, AI users got 19 percentage points fewer correct solutions than the control group. They had fallen asleep at the wheel.

Inside and Outside the Frontier in Workers' Comp

In workers' compensation, the Jagged Frontier maps to specific workflows that carriers can identify today.

Centaur work has a clear line between person and machine.

Ethan Mollick, Co-Intelligence, Chapter 6

In Chapter 6 of Co-Intelligence, Mollick describes two approaches to working with AI that he names after Greek myth. The Centaur, with its clean division between human torso and horse body, represents work where the person decides which tasks go to AI and which stay with the human, then keeps them separate. The Cyborg, with its blurred boundary between organism and machine, represents work where the person and the AI continually interact, with the human revising AI output, the AI revising human draft, and the loop repeating until the work is done. Mollick is clear that both modes have a place in mature operations. The choice between Centaur and Cyborg depends on the workflow being done. The same operator may run one mode in the morning and the other in the afternoon.

What the Research Shows

The Dell'Acqua BCG study identified both modes in practice. Some consultants treated the AI as a delegated colleague, handing off discrete tasks and reviewing the results. They operated as Centaurs. Others held a continuous conversation with the AI, drafting, refining, and re-prompting throughout the work. They operated as Cyborgs. The study found that both modes produced gains over the non-AI control group when the work sat inside the Jagged Frontier. The mode did not determine the gain. The fit between mode and task did. A Centaur approach worked best when the task could be cleanly handed off. A Cyborg approach worked best when the task required iterative refinement, ambiguity resolution, or domain judgment.

Centaur Workflows in Carrier Operations

In workers' compensation, Centaur workflows are the ones where a clean handoff is possible.

Cyborg Workflows in Carrier Operations

Cyborg workflows are the ones where the work is genuinely collaborative between human and AI throughout.

AI is better than the Best Available Human that you can access.

Ethan Mollick, "15 Times to use AI, and 5 Not to," One Useful Thing, December 2024

Mollick introduced the Best Available Human standard in an October 2023 post on his One Useful Thing Substack and reinforced it on X in October 2024. The framework is built around access. His question is whether AI is better than the help a person can reach in a particular moment, with the budget they have and the constraints they face. The answer depends entirely on who is available at that moment.

For an executive with a paid coach on retainer, the best available human is the coach. For a worker without access to a coach, the best available human at midnight on a deadline may be no one at all. For a carrier's overnight FNOL desk, the best available human is whoever is staffed on that desk between 10pm and 6am. The relevant comparison is contextual.

Why This Standard Matters Now

Workers' compensation already has a thinning bench. The US Bureau of Labor Statistics projects the insurance industry will lose roughly 400,000 workers to attrition by 2026. The number of insurance professionals aged 55 and older has risen by 74% over the past decade. Industry survey data shows approximately 28% of claims adjusters intend to retire within the next five years. Industry turnover has risen from the historical 8 to 9% to 12 to 15%.

The Best Available Human standard takes the bench as it is. The assessment runs on empirical comparisons against the operator on the desk at the moment the work arrives. When the question is whether AI matches the average overnight adjuster on completeness, tone, and accuracy of standard FROI fields at 11pm on a Friday, the answer is increasingly yes.

The Canonical Carrier Example

A national carrier's overnight FNOL operation is the canonical Best Available Human case. Claims come in around the clock. Compensable injuries do not wait for business hours. The carrier staffs an overnight desk to handle the volume between 10pm and 6am, with experienced supervisors available on call for escalations.

The work the overnight staffer does is mostly pattern-following. Capture the standard FROI fields. Match to jurisdiction. Confirm the employer's policy is in force. Open the file. Set initial reserves at the carrier's seed amount. Generate the acknowledgment letter to the claimant in the appropriate jurisdiction's language. Hand off to the day-shift adjuster.

An AI tuned for FNOL intake handles most of those steps at a quality that matches or exceeds the average overnight staffer on completeness, tone, and accuracy of standard fields. The supervisor remains on call for the complex cases. The staffing question changes shape.

The Question That Replaces "How Many"

The traditional carrier staffing question is "how many." How many adjusters do we need on the overnight desk to handle the FNOL volume? The Best Available Human standard replaces that question with a different one. What should the experienced adjusters and supervisors be doing instead?

When AI takes routine FROI intake, the experienced overnight staffers move into the work AI cannot do well. Complex compensability triage on suspicious early reports. Coaching newer adjusters on day-shift handoffs that need context the AI cannot capture. Reviewing the AI's outputs on cases that are borderline. The institutional knowledge that lives in experienced staff gets pointed at the work that genuinely needs it. Pattern-recognition tools handle the rest.

This is the same logic that applies to underwriting, medical bill review, and compliance functions. The standard is the same. The application varies by workflow.

The Falling-Asleep-at-the-Wheel Trap

The Best Available Human standard comes with a caution. Dell'Acqua's research found that when AI output is polished enough, humans stop checking it. They fall asleep at the wheel. In workers' compensation, falling asleep at the wheel produces missed bureau filings, incorrect reserve postings, and compensability calls that should have been caught. The standard clarifies where the skilled human's attention pays back most. The duty of skilled human oversight remains in force, especially on cases that look routine but carry hidden complexity.

Always invite AI to the table.

Ethan Mollick, Co-Intelligence, Chapter 3

In Chapter 3 of Co-Intelligence, Mollick offers four principles for working with AI. He names them as ground rules for operating in a world where the AI a reader has access to is already different from the AI the author had when writing the book. The four rules are:

1. Always invite AI to the table.
2. Be the human in the loop.
3. Treat AI like a person but tell it what kind of person it is.
4. Assume the current model is the worst AI you will ever use.

The four rules are written for individuals. They translate cleanly into governance for carriers. Each rule maps to a specific question a carrier compliance, operations, claims, or technology function is already trying to answer.

Rule 1 in Practice: Shadow AI Policy

Mollick's first rule is to invite AI to the table on everything, barring legal or ethical barriers. The carrier version of the rule is governance for shadow AI.

Shadow AI is what happens when employees use AI tools the carrier has not formally adopted. McKinsey's 2025 Superagency in the Workplace research found that 48% of US employees would use AI more often if they received formal training, and 45% would use AI more if it were integrated into their daily workflows. The implication is direct. Employees are using AI today, and they are doing it on personal phones, on personal accounts, and outside the carrier's compliance perimeter when the carrier does not provide sanctioned tools.

The Rule 1 carrier response calls for sanctioned tooling with clear permitted use cases, paired with a transparent inventory of what AI is in use and where. Prohibition pushes shadow use deeper into the organization. Sanctioned tooling brings it into the light.

Rule 2 in Practice: Supervisory Standards on High-Stakes Outputs

Mollick's second rule is to remain the human in the loop, particularly on high-stakes decisions where AI output cannot be trusted unverified. The carrier version is supervisory standards for AI output that touches SIU referrals, denials, and reserves.

SIU referrals generated by AI need to be reviewed by an experienced investigator before they go to the unit. Denials drafted by AI need to be reviewed by a licensed adjuster and, in many jurisdictions, by a supervisor. Reserve recommendations produced by AI need to be reviewed by the claim's assigned supervisor against the carrier's reserving methodology.

The reason is the falling-asleep-at-the-wheel finding. When AI output is polished and consistent, humans stop checking. The Rule 2 governance discipline is a structural counter. The carrier defines categories of AI output that are never released without human sign-off, and operationalizes that sign-off as a workflow gate, not a checkbox.

Rule 3 in Practice: Prompt Libraries with Persona Discipline

Mollick's third rule is to treat AI like a person but tell it what kind of person it is. The carrier version is prompt library governance for any AI-generated content that reaches claimants, providers, or bureaus.

A claimant communication drafted with a generic prompt produces generic output. A claimant communication drafted with a prompt that specifies the AI is a licensed claims adjuster in Florida writing a treatment authorization status update at week six of an active claim produces output that fits the regulatory and tonal context. The persona work matters because workers' compensation language is jurisdiction-specific and audience-specific.

The Rule 3 governance discipline is a centralized prompt library, owned by claims operations, with persona templates for routine communication types (FROI acknowledgments, treatment authorization decisions, return-to-work communications, bureau filing narratives, denial letters). The library updates as the model updates, as jurisdiction-specific case law shifts, and as the carrier's tone-of-voice standards evolve.

Rule 4 in Practice: Procurement Timelines That Anticipate Model Replacement

Mollick's fourth rule is to assume the current model is the worst AI you will ever use. The carrier version is procurement and contract design that anticipates rapid model replacement.

A carrier signing a five-year contract for an AI claims tool on a 2024-vintage model is buying yesterday's capability at a multi-year commitment. The Rule 4 procurement discipline points toward contract terms that allow for model upgrades without renegotiation, vendor commitments to keep underlying model versions current with public benchmarks, and exit clauses that recognize a 12-to-18-month meaningful obsolescence cycle in AI capability. These are industry-aspirational terms today rather than industry-standard ones. Carriers asking for them now are setting the bar that vendors will eventually have to meet.

The same logic applies to internal AI investments. A carrier building an in-house solution on a fine-tuned 2024 model is fixing capability at a moment that will look quaint within two years. The Rule 4 discipline plans for model replacement at the architecture level, not the maintenance level.