The future of AI will depend as much on people who can implement it as on those who invent it.
getty
It’s Monday morning at a regional health system. The prior authorization team left behind a queue several hundred requests deep. Each one represents a patient waiting, a physician’s office calling, and a payer applying its own rules. It is why a woman waits a week and a half for imaging her doctor ordered on a Tuesday.
Now, imagine that Monday differently. Every request in that queue has been read, matched against the relevant payer’s criteria, and sorted. These first forty are clean and ready to submit. These next twelve are missing clinical documentation. These last three will be denied on the current record and need a physician to make the case. Work that otherwise consumed a team for a week is staged for human judgment before the first coffee. The tools that do this are commercially available right now.
However, none of it happens automatically. Someone has to connect the systems so the model has something usable to work from. Someone has to decide which submissions a nurse must review and which can go out on their own, establish what patient data may leave which environment, and build a way to measure whether patients actually got care faster.
And someone has to make sure the whole thing still runs six months after the person who built it takes another job.
Three questions follow from that Monday morning, and not one of them is technical: Who decided a model could make that call? Who would notice if it started getting it wrong? And when the model is wrong about a patient, who answers for it — the vendor, the analyst, or the executive who approved the pilot?
None of those answers comes with the software. Some organizations have someone whose job is to supply them. Most do not.
Those are the questions that determine whether any of this works. They are not the questions being asked. The national conversation wants to know which jobs AI will change, which skills will matter, which industries will move fastest. Those are the right questions. They are not the only ones, and they are not the urgent ones. One experiment starting this fall is built on exactly that question.
The Market Already Named The Missing Job
Look at what the market is already saying. Read the job postings coming out of the companies building frontier AI models and, alongside the research scientists, you find a different category: forward-deployed engineer, AI transformation lead, solutions architect. None of these roles exists to make the models smarter. They exist to help customers use the models that already work.
Every major technological revolution eventually shifts from invention to implementation, and advantage moves from those who invent the technology to those who reorganize work around it. AI is reaching that turn quickly. Enterprise attention has already shifted from which model to license toward agents, workflows, and deployment. AI’s next scarce resource is not intelligence. It is organizational capability.
Implementation has become AI’s next frontier—not in the lab, but inside organizations.
getty
Why Most Pilots Die
Now, notice what that health system has going for it. Capital. A technology budget. A compliance function. A competitive reason to move quickly. And, it will still probably fail, because most organizations do. The GenAI Divide, a 2025 report from MIT’s NANDA initiative, found that roughly 95% of enterprise generative AI pilots delivered no measurable impact on profit and loss. The headline number has drawn fair criticism, and the report bills itself as preliminary findings drawn from a review of some 300 publicly disclosed initiatives, interviews with 52 organizations, and surveys of 153 senior leaders gathered at industry conferences. But the diagnosis is the part that has held: the failure traced not to the quality of the models but to how poorly organizations integrated them.
Then consider the institutions with none of those advantages. A national governing body running youth programs on volunteer labor and a membership system built in 2009. A county agency with procurement authority and no implementation capacity. A nonprofit whose entire technology function is one person who also runs payroll.
AI is becoming a commodity. Institutional capability is not. That asymmetry is where competitive advantage, and institutional inequality, will be decided.
The next AI divide will not separate organizations by who holds the most powerful model, but by who can put models to work and keep them working.
The People Who Sit In Between
Inside most organizations, the constraint is not imagination. What is missing are the people who sit between technology and operations. Translators. Workflow designers. Change managers. Evaluators. People who know enough about the technology to make it useful and enough about the institution to make it relevant.
This is the emerging AI deployment workforce. Without it, AI becomes one more resource that well-capitalized organizations use strategically and everyone else uses sporadically, or not at all.
This is not a new pattern. The economics are well documented. In their work on the productivity J-curve, economists Erik Brynjolfsson, Daniel Rock, and Chad Syverson found that general purpose technologies deliver little measured productivity at first, because their payoff depends on complementary intangible investments: new processes, managerial experience, retrained workers. Realizing such a technology’s potential, they write, requires “a fundamental rethinking of the organization of production” itself.
The obvious objection is that this problem is temporary. Models keep getting easier to use, agents increasingly manage their own integration, and the skills commanding a premium today will be commoditized within a few years. That is likely true of the technical work. The rest of it does not get easier, because the rest of it was never technical. Who is accountable when the system errs, and what evidence counts as proof it helped, are governance questions.
No model answers them for you.
I have watched this play out from the other side of the desk. During twenty-five years leading universities and nonprofits, admissions taught me the same lesson repeatedly: human attention does not scale. AI loosens that constraint, and that is exactly where the risk sits. A model can now read every file in the pool. What it cannot tell you is who answers when a decision turns out wrong. The institutions that handle this well will not be the ones that adopt fastest. They will be the ones that decided in advance what a human being is for.
The risk is that institutional inequality widens. Large corporations will not only have better tools; they will build teams around them, with training budgets, data infrastructure, and someone senior whose job it is to care whether any of it works. Public-serving institutions will have mission urgency and thin capacity. The result is a paradox: the organizations closest to public need are least able to benefit from the technologies most capable of extending their reach.
The Experiment Starting In October
That is the gap Claude Corps, whose first cohort begins in October, is built to address. It is a 12-month paid fellowship that embeds early-career technologists full-time inside nonprofits working on housing, public health, food security, and workforce development, building AI into the daily work and training the staff who will inherit it. Anthropic has committed $150 million and leads the program’s overall strategy, CodePath recruits, trains, and employs the fellows, and Social Finance leads evaluation and is building the financial vehicle to scale it.
The first cohort places roughly 100 fellows, with later cohorts in January 2027 and August 2027 building toward 1,000 fellows and more than 400 host organizations. Fellows are CodePath employees earning $85,000 with benefits, and spend five hours a week in continued training. Each host organization receives a one-time $10,000 implementation grant, administered through Social Finance, and each fellow carries up to $2,500 in Claude licenses and API credits. The program budgets separately for the work of adoption. That is an admission that access was never the binding constraint.
Claude Corps matters not because it solves the problem but because it reframes it. AI equity has been a question of access: who gets the tools, at what price, on what terms. This asks something harder — how an organization develops the capacity to convert those tools into performance that lasts.
The industry that builds these models has already priced that capability. Claude Corps is a bet that it can be built deliberately for institutions that will never compete for those people in the labor market.
When organizations build the capability to use AI well, communities—not just technology—benefit.
getty
The Only Test That Matters
Claude Corps rests on a simple premise. Early-career talent, trained well and supported closely, can move mission-driven organizations from curiosity about AI to workflows that survive the person who built them. Hosts commit to supervision and a plan for sustaining the work after the fellowship ends.
Which points to the only evaluation question worth asking. A fellowship should not be judged by whether a fellow completes a useful project; most competent fellows will. It should be judged by whether the organization is stronger after the fellow leaves. Do the workflows hold? Does anyone keep measuring once the person who introduced the practice is gone? That is the test Social Finance’s evaluation is built to run.
The evaluation carries a second mandate, and it is the more consequential one. Social Finance is not only measuring whether the fellowship model works; it is designing the financing structure that would let it outlive a single company’s commitment. Philanthropy, government, and employers each have reason to fund a pipeline of implementation talent that none of them can build alone. But a shared instrument requires evidence of what that talent produces and for whom: which organizations became measurably more effective, and what that was worth.
“Access to AI is the starting point, not the destination,” said Kirstin Hill, President and CEO of Social Finance. “The organizations that create the most value from this technology will be the ones that treat it as an institutional capability to be built rather than a product to be purchased. What we hope to learn from Claude Corps is what it takes for host organizations to sustain that capability, and how to make AI adoption and implementation affordable for the institutions that need it most.”
In that sense, Claude Corps is less an end than a hypothesis—an experiment designed to test whether organizational capability can be built deliberately, measured rigorously, and ultimately scaled.
Where This Could Go Wrong
The fellowship model carries real risk. A fellowship is not a substitute for sustained investment, and hosts will be tempted to treat fellows as free labor or temporary tech support. Success requires a senior leader with the authority to change a process rather than merely approve a pilot. It also requires avoiding the trap of binding an organization’s AI capability to a single provider, a genuine limitation if this is meant to become infrastructure for a field rather than a pipeline for a company. The design makes that tension concrete: to host a fellow in the first cohort, an organization must already be a Claude for Nonprofits Team or Enterprise customer. The subscription is discounted by up to 75% for nonprofits, but it means the first wave of capability-building reaches organizations that have already chosen a vendor. Anthropic has said it intends to open-source elements of the training curriculum, host playbook, toolkit, and core infrastructure so that other funders and operators can build comparable programs — a commitment that, if honored, is the difference between a demonstration and a template.
None of that argues against the experiment. It argues for watching it closely. The announcement was covered in June; the experiment starts in October, and what happens in between is the part worth watching. The most useful innovations arrive not with answers attached but with a structured way to test the questions.
Program Or Infrastructure
If the design works, the next challenge is arithmetic. Capability built one fellow at a time will not reach the thousands of institutions that need it. A fellowship funded by one company is a program. A standing corps, financed by philanthropy, government, and employers, is infrastructure: a durable pipeline of people whose profession is helping institutions convert technology into capability, available to the organizations that will never staff such a function on their own.
The regional health system from Monday morning does not need another AI model. It needs the organizational capability to make the one it already has work safely, responsibly, and consistently.
The leadership question is no longer whether AI will change work. It is who will have the capability to shape how it changes. Investing in AI without investing in the people and systems required to use it well is not transformation. It is procurement.
Disclosure: The author serves as an advisor to Social Finance, one of the organizations described in this article.
Leave a comment