Hire The Interface

Two press cycles in late April 2026 told the same story, from opposite ends of the same wire.

On April 28, EY's UK and Ireland practice announced it was standing up a dedicated Forward Deployed Engineer team. The press release described senior engineers who would sit inside client delivery teams, write production code, embed governance from the outset, and stay until something ran. The framing was blunt. AI pilots were stalling in the gap between model output and production reality, and the firm needed a new shape of consultant to close it.

The day before, on April 27, Marc Benioff posted on X that Salesforce was hiring a thousand new graduates and interns to build Agentforce and Headless360. In the same press window he told reporters the company was not hiring more engineers in fiscal year 2026, because coding agents had reduced the need. Headcount of Salesforce engineers had been frozen near 15,000 for two years. The hiring story was a freshman class, not a senior recruitment drive.

A consulting firm building a senior, embedded delivery function for AI. A software vendor freezing its engineering bench and pouring the budget into graduates supervised by agents. Two opposite-looking moves, one underlying idea.

The conventional plan for building an internal AI team in 2026 is drawn from the wrong map. The headcount that matters does not sit in a center of excellence. It sits inside the workflow, with a model on tap and a domain it already knows.

The script most executives are running

Walk into a Fortune 500 boardroom this quarter and the AI org design slide looks roughly identical from company to company. There is a new C-suite seat. The IBM CEO Study published in May 2026 found that 76 percent of large organizations now have a Chief AI Officer, up from 26 percent twelve months earlier. That is the fastest growth rate for any C-suite title in modern executive history.

Below the CAIO there is a platform team. ML engineers who own the model registry. Data engineers who own the pipelines. MLOps engineers who own the deployment surface. An AI governance lead who owns the model cards and the bias audits. A handful of applied scientists who run the proofs of concept. Sometimes a "Center of AI Excellence" that pretends to be a service organization and behaves like a research lab.

Compensation flows accordingly. The Chief AI Officer band runs $400K to over a million in total comp at enterprise scale. Senior ML engineers, in the post-Meta-signing-bonus landscape, clear $500K at firms competing with the labs. PwC's 2025 Global AI Jobs Barometer documented a 56 percent wage premium for workers with AI skills, up from 25 percent only a year earlier. The premium doubled in twelve months.

The map looks logical. AI is the new thing. You need a leader for the new thing. The leader needs a team. The team needs a budget. The budget needs governance. The governance needs a committee. The committee meets. The pilots get funded. The pilots get demos. The demos get applause. The pilots do not enter production.

That last sentence is the part everyone keeps trying not to say out loud. Hold it there.

What the frontier labs actually staffed

If you want to know how the people closest to the technology think it should be deployed, study who they paid to hire. They did not hire central platform teams. They hired embedded operators.

A May 20, 2026 MarkTechPost piece walked through the shift. OpenAI, Anthropic and Google had each posted aggressive 2026 hiring plans for Forward Deployed Engineers. Salesforce committed to 1,000 FDE roles. Google Cloud had 59 open FDE postings across four continents. Anthropic ran an Applied AI track that, in practice, looked identical to the Palantir FDE model that had been quietly compounding inside government and energy clients for fifteen years. Senior FDE total compensation at the frontier labs cleared $500K, and at OpenAI the L5 to L6 band ran $700K to $1.28 million.

In November 2025, OpenAI bought a small London consultancy named Tomoro for roughly $80 million, primarily to acquire its 150 Forward Deployed Engineers. Sam Altman, in the announcement memo to staff, was direct. The bottleneck on enterprise revenue was not model capability. It was the headcount that could sit inside a client and turn capability into a running system. He spent eighty million dollars on a hundred and fifty seats because the math worked.

EY's April 28 launch was the first major consulting firm to formalize what the labs had already proven. The press release flagged a specific finding from EY's UK research. 78 percent of organizations claimed AI was fully or mostly implemented. 49 percent admitted their approach was insufficient for autonomous models. The gap between those two numbers is the entire reason a senior engineer who codes inside the client suddenly costs more than a partner who advises from outside it.

Five other Tier 1 consultancies are quietly drafting equivalents. Deloitte, Accenture, PwC, KPMG and Capgemini will all have something with a similar shape by the end of Q1 2027. The Forward Deployed Engineer is not a job title. It is an admission that the bench is in the wrong place.

The Salesforce signal, read carefully

Benioff's April 27 announcement gets misread constantly. It reads on the surface as a story about junior hiring. Read it as a story about org topology.

The Salesforce engineering bench did not grow this year. The Salesforce engineering bench will not grow next year either, by the CEO's own statement. The growth happened at two places. New graduates and interns at the bottom of the pyramid, building product features alongside coding agents. A separate, smaller pool of senior people working closer to customers, deploying Agentforce into real environments.

The middle hollowed out. The senior IC who shipped against a product spec, the staff engineer who refactored the platform, the line manager who owned the velocity report. Those roles did not get the new headcount. The new headcount went to operators near customers and apprentices near agents.

If you accept that the same dynamic is going to play out inside every enterprise that takes AI seriously, the entire planning exercise changes. You are not building an internal AI team. You are reshaping the workforce around two access points to model capability. One sits next to the customer or the operating workflow, with senior judgment and a high willingness to ship into production. One sits next to the model itself, learning the new craft from the inside while the agent does the rote work.

The CAIO is a real role. The center of excellence, in most companies that have one, is a procurement function in a lab coat.

Why building an internal AI team almost always means building the wrong one

The most useful exercise I run with executive teams sounds simple and almost never lands the first time. Take the org chart you have drafted for the AI function. Cover the box that says CAIO. Cover the boxes that say platform, governance, infrastructure. Look at what is left.

If what is left is the people closest to the workflow where the model is supposed to run, you have a chance. If what is left is empty, you are about to spend three years building an internal AI team that produces governance documents and pilots while the actual work continues to happen somewhere else.

The wrong frame is "we need a team for AI." The right frame is "we need to put model access inside the team that owns the work, and we need a small group of people who can sit with them and close the gap to production."

That sounds small. It is not small. It is structural. It changes who you hire, what their reporting line looks like, where the budget sits, how success is measured, and which hires are net new versus which are reassignments.

Three concrete moves.

First, identify the five to ten workflows in your business where a domain expert and a coding-capable operator with model access could plausibly redesign the work from the inside. Not the workflows on the digital transformation slide. The ones that actually carry revenue or cost or risk. The claims adjuster who handles complex cases. The clinical reviewer who works the inbox of appeals. The pricing analyst who fights with promotion logic every quarter. The compliance officer who reads contracts by hand. These are the rooms where the model will land or fail to land.

Second, pair each of those rooms with one Forward Deployed Engineer. Not a consultant. Not a partner from a firm. A person on your payroll, your benefits, your equity plan, whose job description is to build the system that runs inside that room and stay until it runs. Anthropic's own retention numbers tell you why this matters. 80 percent of their two-year hires stay, despite paying below OpenAI's median. People stay where the work is real. They leave where the work is performative.

Third, hire one CAIO if you do not already have one, but write the job spec backwards from the workflows. The CAIO is not the head of a central technology team. The CAIO is the executive who owns the outcomes of the embedded pods and the cross-cutting governance that lets them operate at speed. The platform exists in service of the pods. Not the other way around.

This is the entire architecture. It is small. It is unglamorous. It requires you to defund some of what is already in flight. It scales linearly with the workflows you can actually staff.

The talent market does not look the way the headlines suggest

There is a separate problem that anyone trying to staff this model runs into within the first week. The market for the people you need is heavily warped by the labs.

A May 21, 2026 Euronews piece walked through the bidding war. Meta offered $100 million signing bonuses to senior researchers at OpenAI and Anthropic during the 2025 talent push. OpenAI's average stock-based compensation across its 4,000 employees ran $1.5 million per person, with 46.2 percent of annual revenue going to equity. Anthropic's L4 senior software engineer hit $665K total comp on average. The Pin AI compensation report tracked a 56 percent year-over-year wage premium on AI roles, and the curve was still bending up.

None of those numbers apply to most enterprises. You cannot, and should not, compete head-on with the lab compensation bands. The Anthropic retention story tells you the leverage point. Engineers leave OpenAI for Anthropic at an 8 to 1 ratio because of mission, not money. Money has a ceiling on its retention power. Anthropic retains 80 percent of two-year hires. Meta retains 64 percent. Eighty million dollars in signing bonuses bought Mark Zuckerberg a fleet of researchers who started looking for the exit within a year.

The talent you need for the embedded model is not the researcher who would otherwise go to a frontier lab. It is the senior engineer who is bored shipping CRUD endpoints and excited by the chance to ship inside a domain that matters. The data scientist who is sick of dashboards and wants to build the system that replaces them. The clinical informatics person who already understands the workflow and can be taught to use the model. The actuarial analyst with two years of Python who can be paired with a model and produce work that was impossible six months ago.

These people exist in your headcount today, or one step away from it. They cost less than a frontier-lab researcher. They stay longer. They produce work that compounds because they understand the domain in a way no outside hire can.

The mistake is to look at the lab compensation numbers and conclude that you cannot afford to build. The right read is that you cannot afford to copy the org chart of a frontier lab. The shape is wrong for your business and the price is wrong for your market.

The CAIO problem, stated plainly

The 76 percent CAIO statistic from the IBM CEO Study is the most over-interpreted number in enterprise AI right now. The fact that three quarters of large companies have hired a Chief AI Officer is treated as evidence that AI is taken seriously at the top. It is closer to evidence that AI is being parked at the top.

When a function gets its own C-suite seat without a clear operating model underneath, the function gets isolated. The CAIO writes the strategy. The strategy lives on a slide. The slide is shown to the board. The board approves. The CAIO assembles a team. The team builds a platform. The platform is presented to business unit leaders. The business unit leaders nod. The platform is not adopted. The CAIO is replaced in eighteen months.

This is a story that has played out in three previous waves. The Chief Data Officer wave of 2014 to 2017. The Chief Digital Officer wave of 2017 to 2020. The Chief Innovation Officer wave that sputtered earlier and has been quietly de-titled in most large companies. In each wave the failure mode was the same. A horizontal function was created for a problem that needed to be solved vertically, inside the business units, with strong central support.

There is a version of the CAIO role that works. It is the version where the executive runs interference for embedded teams, makes hiring fast, kills bad pilots quickly, owns the governance surface so the operators do not have to, and reports outcomes in business terms rather than model terms. That version of the role is rare. It is also the version that does not need a "Center of Excellence" because the excellence is centered in the workflows where the work lives.

If you are hiring a CAIO this quarter, write the job description for the second version of the role. If you have already hired the first version, reorganize the function around embedded pods before the title becomes a liability.

What this looks like, six months in

A picture of a working internal AI team in mid-2026, in a company that runs the model described above, is not very glamorous. There is no Center of Excellence. There is a CAIO who spends most of her week with operating executives, not with model vendors. There is a platform team of five or six people who maintain the shared infrastructure that the embedded operators draw from, mostly the eval harness, the deployment templates, the cost monitoring, the procurement plumbing for model access.

The interesting people are not in any of those boxes. They sit in claims operations, in pricing, in supply planning, in clinical review, in compliance, in customer support. They have a Forward Deployed Engineer paired with them. They have a senior domain expert on the team who owns the outcome. They have a model on tap, with the eval infrastructure to know whether it is working. They are shipping changes to the workflow on a weekly basis, with rollback discipline. Most of the time, the work looks like software engineering. Some of the time it looks like operations redesign. None of the time does it look like a research project.

The CAIO can name every one of these pods. She knows the metric each one is moving. She knows the FDE by name. The platform team supports the pods without owning them. The pods own their own production fate. The governance committee meets monthly, looks at risk metrics from each pod, and largely stays out of the way.

This is a small company inside the big company. It is the part of the org that has actually internalized what the technology is. It is also the part of the org that does not appear on any of the off-the-shelf "AI operating model" frameworks the big consulting firms were selling in 2024, because those frameworks were written before EY paid people to staff the inside of the client.

The CTA, stated as plainly as the rest

You can read this and write it off. The frame can sit on a slide and never touch the staffing plan. The CAIO can keep building a horizontal function. The pilots can keep producing demos. The talent budget can keep flowing to a centralized team that the business units do not call. None of this is theoretical. Most companies will run the wrong play. Some of them will be acquired by the ones that ran the right one.

If you want to run the right play, you need help that does not come in a fifty-page report. You need a partner who has stood up the embedded pod model inside companies that look like yours, who can write the FDE job specs, who can pair the first three rooms in your business with the first three operators, and who can do that before the next quarter closes.

That is the work I do at Agor AI Advisory. Architecting the team that runs the model, not buying the model that needs a team.