The AI native firm runs into a human it cannot upgrade
By Professor Ashley Braganza
A firm can now delegate the bulk of its production to software and keep only its senior people. The pyramid that defined professional services, a thin layer of partners resting on a broad base of juniors, can be inverted. The junior base becomes a fleet of agents. This is not a forecast. It is an operating model already running, and it exposes a constraint that no amount of compute removes. The limit on agentic work is not the machine. It is the person left to review it.
My guest on The AI Adoption Podcast is Chris Donahoe, Co-founder of Stillpoint, a corporate affairs and public relations firm headquartered in Washington DC and built to be AI native from the start. He spent close to a decade at Edelman, where he founded its first AI advisory practice. At Stillpoint he has staffed the firm with senior experts alone and handed everything else to agents.
Destroying assumptions matters more than installing software
Chris is precise about where an AI native firm begins, and it is not with the technology. The starting point is the willingness to ask why the business works the way it does, why an industry has always done things a certain way, and whether that legacy approach still creates value. He describes the exercise as one of destroying assumptions. Many companies, he argues, are using AI to speed up an operating model that ought to be questioned rather than accelerated, and the opportunity many of them miss is to ask whether the old model still deserves to exist. The distinction carries a sharp edge. Automating a process that no longer creates value simply produces waste faster and at greater volume.
At Stillpoint the assumption discarded is the talent pyramid itself. In a legacy professional services firm, a thin layer of senior partners with deep expertise rests on a broader band of mid-level staff, which rests in turn on a wide base of juniors who carry the production of the work. Chris has inverted this completely. The human talent is partner and managing director level only, former chief communications officers and heads of practice, with no junior staff. The support infrastructure beneath them is agentic. Work that once went to mid and junior colleagues now goes to teams of agents that he scopes, directs, and manages in a similar fashion. He also insists this only holds with constant experimentation, at the level of the team rather than the individual, because the capabilities shift week to week and a firm that is not testing new workflows will not see the next one arrive. The model is not a productivity tweak bolted to the old structure. It is a different structure, and it redefines what the humans in it are for.
The productivity promise stops at a biological limit
The intuitive expectation is that agents expand a person's capacity almost without limit. Chris admits the firm fell into that trap early, then ran headlong into its correction. Agents can take a piece of work to ninety per cent done at remarkable speed and in remarkable volume. The last stretch, the review, the exercise of judgment, the refinement that makes output fit to send to a client, still falls to a human whose reading speed and attention have not improved since ChatGPT was released. He calls this the biological bottleneck. Production capacity has risen sharply; the human capacity to absorb, coordinate, and approve it has not moved.
One implication is that value is not created evenly across a process. Some points in a workflow generate the bulk of the value and others generate very little, so the discipline is to spend human attention where the value concentrates and to commoditise the rest. Chris spends the majority of his own time on the first mile, the scoping and the context that direct a task, and the last mile, the refinement that makes work customer ready, and he is blunt that the AI outputs are not something he would send to a client untouched. The middle, the production, goes to the machines. The complication is that this bottleneck reshapes decisions that appear unrelated to it. If review time is the scarce resource, the opportunity cost of spending that time elsewhere climbs. Chris has cut the number of internal meetings the firm holds, because an hour of four people in a room or online is an hour of review and refinement foregone, and the volume that hour could have cleared is far larger than it was a few years ago. The bottleneck does not stay inside one workflow. It reprices human attention across everything the firm does.
Every professional is now a manager, whether they asked to be or not
The consequence for individuals is a shift in the skills that command a premium. Chris looks for three capacities at once: deep real world expertise in a field, a generalist's range to integrate several disciplines into a single body of work, and a worked in fluency with AI tools, close to muscle memory. As teams on a project shrink, each person has to cover more ground, so the combination matters more than any single part of it. None of these, he notes, can be installed in an onboarding session. They come from curiosity, experimentation, learned pattern recognition, and years of doing the work, which is precisely why the combination is scarce.
The second consequence is managerial. On his account, even a solo operator is now a manager, because a team of agents still needs scoping, clear delegation, useful feedback, synthesis across outputs, and the judgment to know when work is finished or needs a final pass. Those were once the tasks of a management layer; they are now the baseline for everyone. Yet the orchestration is far from solved. Chris works across Claude, ChatGPT, Gemini, email, and proprietary systems at the same time, and finds it increasingly hard to track what he has delegated and where each task sits. He is candid about the tools rather than diplomatic: strong as Claude is at formatting a Word document or an Excel model, he finds it turning into a poor copywriter, so he routes near final drafts out to ChatGPT for editing and back into Claude to implement the changes. No off the shelf system coordinates this hybrid of people and agents. Whether agents can be kept aligned across functions without a person holding the thread is, on the current evidence, an open question.
There is a sharper edge to this for anyone early in a career. Chris is not opposed to junior talent, but he is blunt that offering to do the work more cheaply no longer holds, because an agent is cheaper still, faster, and does not take a holiday. The value a newcomer must now bring is curiosity, a habit of experimentation, and the kind of AI literacy that cannot be taught in a week. The graduate who arrives with that instinct is a colleague he wants; the one whose only pitch is a lower rate is competing against software built to undercut it.
The unresolved question this model raises is one of succession. If the winning firm employs only senior experts and delegates the rest to agents, the pipeline that produced those experts thins. Chris is honest that no one has solved this, and floats fellowships with clients as a possible answer. On this evidence, the constraint on AI adoption is not the ambition of the technology but the supply and the stamina of the people who must direct it. Build the fleet of agents. The bottleneck, and the judgment, remain human.
Listen to the full conversation
Listen to the full conversation with Chris on The AI Adoption Podcast.
YouTube: https://youtu.be/Fo3IiD-odJM
Comments