The Salesforce AI Bottleneck Nobody Talks About

On September 10, a nonprofit called Per Scholas announced a paid fellowship for graduates to work on Salesforce AI projects at New York City nonprofits. It was covered as workforce development news. It might be more useful to read it as a symptom.

Per Scholas is not publicly traded. But the thing it's trying to fix is.

Salesforce sits at a $200 billion market cap, trading at roughly 20 times trailing earnings, and has rebranded itself around an AI strategy called Agentforce. The numbers Salesforce itself touts are impressive: Agentforce annual recurring revenue reached $1 billion. Combined AI and data products bring in $3.4 billion a year. The company grew revenue 11.2% year-over-year, and has been beating earnings estimates for several consecutive quarters, with the last reported quarter at $2.91 per share versus a consensus of $2.78.

The market appears to be paying attention. Stock is up about 26% over the last 20 trading days, up from a 52-week low of $146 to its current level around $243.

But there's a structural problem that no quarterly earnings call fully addresses: AI tools don't work without people who can govern them. And the people who can do that are in short supply.

The Salesforce ecosystem is not just a software platform. It's a talent-dependent system. Customers buy Salesforce, but they need administrators to configure it, architects to design it, and specialists to maintain it. The software's value to any given customer depends almost entirely on whether those people exist inside or around that organization.

Here's what the talent data shows. Demand for Salesforce Technical Architects jumped 27%. Supply grew 4%. Thirty-four percent of organizations say the architect shortage is their primary delivery blocker — ahead of budget and strategy. Meanwhile, Salesforce job postings on Glassdoor nearly doubled, from about 14,000 in May 2024 to over 31,200 by September 2025, with no corresponding increase in qualified senior candidates.

The gap is not evenly distributed. Entry-level administrator supply grew 47% — in part because training programs like Per Scholas have made the baseline certification more accessible. But the admin role has evolved. It's no longer just clicking and configuring. Fifty-three percent of admins say too much is expected of them. About one in five are solo admins running entire orgs alone. Only 13% of entry-level admins feel confident handling technical debt, compared with 44% of experienced ones.

And technical debt is the single largest obstacle to making AI work. Fifty-six percent of admins list it as their top challenge. Only 2% of admins describe their org as clean. Thirty-one percent say debt levels are high enough to slow daily work.

This matters for AI because AI amplifies everything. It does not fix bad data — it makes bad data look confident. Salesforce's own strategy acknowledges this: the "secret sauce" for effective AI agents is unified, trusted data. Without it, you get what one analyst called "confidently wrong results."

The talent shortage translates directly into wasted customer spending. Revenue Cloud licenses operate at 30% capacity in some organizations because there's nobody with the CPQ expertise to fully configure them. Service Cloud automation sits unconfigured. Marketing Cloud licenses go idle. These aren't edge cases. They're structural.

There's a paradox here that's worth sitting with.

AI tools are automating the entry-level tasks — basic reporting, data maintenance, standard configuration — that used to be the training ground for tomorrow's senior architects. At the same time, the senior roles that actually deliver enterprise AI value require five or more years of experience to develop.

So the technology that is supposed to accelerate adoption is also eroding the pipeline that creates the people who make adoption work. This doesn't mean the pipeline is closing. But it means the problem is self-reinforcing in ways that won't show up on a product roadmap.

Per Scholas' fellowship is a small attempt to patch this hole. It takes Salesforce Administrator graduates and places them in eight-month paid positions at nonprofits, where they gain hands-on experience with AI-enabled tools like Agentforce and Prompt Builder. It's real work in real environments. The program is narrow — New York City nonprofits only — but the fact that it exists, and that it requires nonprofits to pass an AI readiness assessment before receiving a fellow, says something about the state of the ecosystem. The nonprofits themselves aren't ready for AI. They need training before they can train the trainees.

This is not to say Per Scholas or Salesforce are failing at talent development. Salesforce has partnered with Per Scholas since 2014, providing technology, funding, and volunteers. The nonprofit now has over 35,000 graduates and operates in 25 locations nationwide. These are real outcomes for real people who move from pre-training incomes of roughly $20,000 to post-training incomes of $54,600.

The point is different. The talent ecosystem Salesforce depends on is a real constraint on its AI growth story, and it's a constraint that gets less attention than product launches and ARR milestones.

What does this mean for the investment case?

Salesforce is a $200 billion company. Its AI revenue of $3.4 billion is meaningful but small relative to total revenue, which runs about $10 billion per quarter — roughly $40 billion annually. Agentforce at $1 billion ARR is a growth story, not yet the core. The stock's 20x trailing P/E reflects a mature business that the market is trying to price as a growth AI business.

If customers can't fully use the AI tools they're paying for because of talent constraints, the risk isn't that Salesforce stops selling licenses. It's that expansion slows, utilization stays low, and the AI premium in the valuation proves premature. Revenue could still grow steadily at 11% while the AI narrative under-delivers on the multiple expansion that the market is pricing in.

Conversely, if the talent pipeline eventually catches up — through programs like Per Scholas at scale, through internal upskilling, or through AI itself becoming simple enough that it doesn't require architects — the constraint dissolves. The $200 billion valuation could hold. But that outcome depends on a labor market problem resolving in a specific direction.

The thing to watch isn't the next Agentforce announcement. It's whether Salesforce customers are actually getting value from the AI they've already bought. If utilization stays low and expansion revenue stalls, the market will eventually separate the AI story from the underlying business. And when it does, the price that stock commands for the underlying business alone might look different than it does today.