AI Is Changing PR, But Most Underestimate Its Impact

There are a number of interesting arguments taking hold in the PR industry about AI. Two of them are being made by colleagues I admire and respect greatly.

The first, broadly defined, is AI will make routine work cheap. Expertise will become easier to access. But judgment, experience, relationships and strategic counsel will become more valuable. Machines will handle production while experienced practitioners move up the value chain.

Andy West, Managing Consultant at West of Center, made a version of this argument recently in a post titled When Expertise Becomes Cheap, Judgement Becomes Expensive. He argues that as generic capability becomes widely available, value shifts toward experience, context and judgment. His advice to agencies follows logically: don't compete on producing more stuff. Build businesses around the things humans do particularly well.

The second is we are digging ourselves a training deficit for the next generation of practitioners who will be starved of first-hand experience as low-level work is handled by machines.

Stephen Waddington, Director at Wadds Inc., has been thinking about this consequence. PR traditionally trained junior people through production work: monitoring, reporting, drafting and pitching. AI can now perform a meaningful share of that work. Waddington worries, reasonably, that removing the bottom rungs of the career ladder could eventually leave the industry without enough experienced advisers at the top.

Both arguments make sense within the way PR works today. In fact, I think both identify real consequences of AI. Where I part company with West and Wadds is in what happens next. But both still take the practitioner as the basic unit of production and the unit of analysis. AI does some work, people do the rest, and the industry reorganizes the boundary between them.

What if AI changes more than the division of labor? What if it changes the amount of cognitive work a communications organization can economically perform in the first place?

Drucker's Unfinished Problem

Peter Drucker identified the problem remarkably clearly in 1999.

Writing in California Management Review, he argued that management's great contribution in the twentieth century had been a roughly fiftyfold increase in the productivity of the manual worker in manufacturing. He thought management's central challenge in the twenty-first century would be doing something similar for knowledge workers.

At the time, computers had become common. The internet was spreading. Enterprise software was moving into almost every business function. The basic unit of knowledge production was still a person.

Consider a PR agency in 1999. People researched, read the news, analyzed competitors, developed ideas, wrote pitches, called journalists and assembled reports. Senior people did work, reviewed work, made decisions and counseled clients.

By 2020, the tools were dramatically better. Search, media databases, monitoring, analytics and collaboration had all improved. The production model has remained surprisingly unchanged. People are still the processors. Software mostly helps humans move information around.

Now, AI has introduced something different. Machines can now participate directly in work we previously classified as cognitive labor. Machines can read, compare, classify, summarize, research, draft, interrogate datasets, analyze patterns, generate alternatives, critique work and increasingly perform multi-step assignments.

Current systems have obvious limitations. They make mistakes and still require supervision. However, for the first time, we have machinery that attacks Drucker's productivity problem directly. That should make us reconsider what productivity means.

Software Development Is Giving Us a Preview

Software engineer Steve Yegge has spent more than 30 years writing software, including long periods at Amazon and Google. After initially being skeptical of AI coding, he has become an aggressive advocate for coding agents.

What's interesting isn't that Yegge thinks AI can help programmers write code faster. He thinks programming itself is moving to a different level of abstraction.

Yegge views software engineering as a layered activity in which code generation is increasingly transferred to machines. Developers move toward design, coordination and integration as AI takes on more of the underlying production.

He takes the argument further. Yegge has been experimenting with systems in which a developer coordinates many agents working in parallel and has predicted the emergence of "super-engineers" capable of running 50 to 100 coding agents simultaneously.

The numbers may prove wildly optimistic. The production model Yegge contemplates is more interesting. One developer writing code faster is productivity software. One developer directing dozens of machines that write, test, review and revise code starts looking like a different production system.

And software is already encountering the consequences. When ten agents can produce code simultaneously, integration becomes the bottleneck. Coordination and quality control become more complex as technical debt accumulates faster because productive capacity has exploded. Once the means of production become abundant, the organization has to adapt to that abundance.

PR should pay attention.

Is Judgment Really the Safe Ground?

This brings me back to West's argument. If expertise becomes cheap, he argues, judgment becomes more valuable.

Maybe.

Good judgment is certainly valuable today. Clients don't hire senior advisers because they can write a press release faster. They pay them to understand the situation, challenge assumptions and recommend what to do.

But why should we assume judgment remains scarce?

Judgment isn't a single mysterious human capability. Much of what we call judgment involves gathering evidence, recognizing patterns, comparing alternatives, recalling analogous situations, estimating consequences and testing assumptions.

Machines can already participate in many of those activities. That doesn't mean AI has human judgment. It means the boundary between machine production and human judgment may not stay where we have conveniently drawn it today.

Software development is demonstrating this in real time. A few years ago, programmers dismissed AI-generated code as unreliable autocomplete. Basically, slop. Then the capabilities moved to functions, then applications. Now developers are experimenting with agents that plan work, implement it, review it and coordinate with other agents to ship production code.

The human hasn't disappeared. The human role keeps moving. There is no particular reason communications should be exempt from the same process.

The Apprenticeship Problem Points Somewhere Else

Waddington's concern about junior practitioners is harder to dismiss. Monitoring coverage, building media lists, writing first drafts and pitching journalists weren't merely cheap labor. They were how people learned PR.

Take that work away and something important disappears. But the preservation of inefficient work as the foundation of training new practitioners seems to be the wrong response.

We don't require accountants to perform arithmetic manually before allowing them to use spreadsheets. Pilots train on simulators. Technology changes the work, and training changes with it.

AI might eventually allow junior practitioners to acquire skills and experience faster. An early-career communicator could analyze hundreds of crisis responses instead of seeing three during the first five years of a career, compare thousands of journalist interactions, examine how narratives changed across competing companies and test alternative messages against simulated stakeholders.

None of that substitutes for sitting across the table from an angry CEO or convincing a skeptical reporter that a story is worth another look. But neither did compiling clip books and taking meeting notes.

The question is what apprenticeship should look like when access to analysis, simulation and feedback becomes dramatically cheaper.

PR Has an Expensive Cognition Problem

A modern communications department is surrounded by more information than its people can possibly process. Teams cannot read every story about their market, continuously analyze every competitor's claims, listen to every relevant podcast and earnings call, track every shift in customer language or maintain a current map of the questions buyers are asking across search, social and AI systems.

Most importantly, they cannot continuously compare what the company says about itself with what journalists, analysts, customers and competitors are saying about it.

So we sample.

We subscribe to monitoring platforms. We build dashboards. Agencies produce monthly reports. Executives make decisions from fragments of a much larger information environment. We've become so accustomed to operating this way that we think of it as simply how communications works. Yet much of this is merely the rationing of expensive cognition.

If machine cognition continues to get cheaper and more capable, which it will, then the opportunity isn't limited to doing today's work faster. We will be able to produce and consume much more information and analysis to make better decisions.

A communications organization could continuously model the issues surrounding a company, track changes in competitor narratives, detect emerging questions, analyze the sources influencing AI-generated answers, test claims against evidence and identify gaps between what a company wants to be known for and what its information environment actually supports.

Much of this is technically feasible today, even if AI reliability varies by task. The larger constraint is increasingly organizational.

Cheap Cognition Changes the Economics

When mechanical power became cheaper, we didn't just stop using human and animal muscle to move things. We found vastly more things worth powering.

William Jevons pointed out how improvements in efficiency lower the effective cost of using a resource, which increases its total consumption, known as Jevons paradox. If AI makes cognition dramatically cheaper, we should at least consider the same possibility, that we will consume vastly more cognition.

When computation became cheaper, businesses didn't run the same number of calculations and pocket the savings. They found millions of additional things worth calculating. When storage became cheap, companies started storing almost everything. In 1945, a billion floating point operations cost roughly $1.3 trillion. In 2025, the same computation cost about one cent (Exponential Tracker).

What happens if cognition follows a similar curve?

The answer may not be smaller PR departments producing the same amount of work. It may be communications organizations consuming 100 times more analysis than they could previously afford and producing 100 times the work.

If we simply automate existing agency workflows, the prize is efficiency. Perhaps an eight-person account team becomes four people and a collection of agents. But cheap cognition gives us another option, where the four-person team and AI perform vastly more work than the original eight-person team ever could.

Competitive analysis that happened quarterly could happen daily. Narrative analysis could move continuously across thousands of sources. Emerging issues could be detected before someone thinks to commission a report. Buyer questions could be mapped and remapped as markets change. Claims could be continuously checked against available evidence. Communications strategy could respond to an information environment that the organization can actually observe rather than periodically sample.

The unit of work itself changes.

Palantir offers a useful example of where this logic can lead. In defense and intelligence, Palantir integrates enormous volumes of fragmented information, from intelligence sources to operational systems and battlefield sensors, into a common environment where software can correlate information, identify patterns and move findings from analysts to operators.

The lesson for PR is that by automating intelligence analysts Palantir changed the amount of intelligence an organization could economically collect, connect and analyze before a human has to make a decision. That is a much more interesting ambition for communications.

From Knowledge Workers to Knowledge Production

Drucker framed the great management challenge as increasing the productivity of the knowledge worker.

Twenty-seven years later, I think we may need to modify the question.

What happens when the knowledge worker is no longer the only unit capable of producing knowledge work? Drucker insisted that our job as managers and leaders is to focus on achieving outcomes and creating value for the business, its customers and stakeholders. Machines can now create some of that value, and they get better by the day.

Yegge is seeing one answer in software development. Palantir gives us a glimpse of what happens when machine analysis becomes part of the operating architecture rather than a tool handed to individual employees.

PR is still largely talking about individual productivity. Teaching 100 people to use ChatGPT better may improve productivity. It doesn't necessarily produce a new operating model. The harder work is organizational.

Which communications activities should operate continuously rather than periodically? Which decisions can machines support and which require human authority? What information should systems collect continuously? What happens when competitive analysis that once took three weeks can run every morning? What should a communications organization know that it doesn't know today because finding out has always been too expensive?

Those questions lead somewhere quite different from "AI in PR."

In my own work I have been using the term computational communications for this emerging model, although the terminology matters less than the idea.

Communications has historically been organized around human capacity because human cognition was the limiting factor. Agencies sold hours. Departments allocated people. Research was episodic. Analysis was sampled. Strategy was periodically refreshed because continuously reconsidering everything was impossibly expensive.

Remove that constraint, even partially, and some assumptions holding the current model together start to look arbitrary.

This is where I ultimately differ from the argument that AI makes human judgment more valuable or that the central problem of our craft is rebuilding the apprenticeship ladder. Both may be true. But they are the subtext of a larger change.

The more consequential question is what we would build if we were designing the communications function today, knowing that cognition itself is becoming a machine output and the scale of it is increasing exponentially.

I'm not sure the answer looks much like the PR agency or communications department we see today.