Every enterprise leader we talk to is being told to add AI talent. Most of them are being handed a budget line and a title, and very little else. Then the req goes out, four hundred resumes come back, and the hiring manager discovers that "AI" on a resume can mean anything from fine-tuning a model to prompting a chatbot to standing up a GPU cluster.
The phrase itself is the problem. "AI talent" is a category, not a role. Asking to hire it is like asking to hire "cloud talent" in 2012. Are you looking for the person who racks the hardware, the one who secures the workload, the one who wires it into your business systems, or the one who tells you what the output means? Those are four different people with four different backgrounds, and the market prices them differently.
The market data says the same thing
This is not just our read. When TechTarget asked seven IT leaders which roles are hardest to fill right now, none said "AI specialist." They said data engineers, AI architects who have actually fit the pieces together at scale, and people with deep AI infrastructure expertise. Doug Gilbert, CIO and chief digital officer at Sutherland, put it directly: "Data engineering is probably the hardest and most critical role to hire for right now, because bad data leads to higher hallucinatory rates" (TechTarget).
Meanwhile, roughly 51% of job postings that require AI skills now sit outside traditional IT and computer science roles, and postings that include AI skills carry about a 28% salary premium (Lightcast). Read those two facts together and the picture is clear: the AI skill requirement has scattered across the org chart, and the labor market is charging you a premium for a word that no longer specifies what you are buying.
So the first discipline is refusing to write the req until you know which layer you are staffing.
Start from what AI actually is in your environment
Here is the framing that has kept our hiring honest: AI is an app.
It runs on compute. It moves across a network. It reads and writes data. It authenticates, or it should. It has an attack surface, and increasingly that attack surface can act on its own. Strip away the novelty and an AI workload behaves like the most demanding application your environment has ever hosted, with two new properties: it is non-deterministic, and it can take actions.
That reframe matters because it tells you where to recruit. If AI is an app, then the people who know how to land a hard app in a real enterprise are the people you want, and those people already exist. They came up in infrastructure and security. We target that background deliberately, because the failure modes we see in the field are infrastructure and security failure modes, not model failure modes.
There is a limit to the framing, and it is worth naming so nobody over-rotates. Calling AI an app understates the data layer. Applications are shaped by their code; AI systems are shaped by the data and context you feed them, and that plumbing does not look like traditional app deployment. Gartner has predicted that through 2026, organizations will abandon 60% of AI projects that are not supported by AI-ready data (Gartner). So: AI is an app, and its most important dependency is a data estate most enterprises have not built yet.
The four skill sets we actually hire against
| Layer | What this person actually does | What we screen for | Market reality |
|---|---|---|---|
| AI infrastructure | Sizes and lands compute, storage, and network for training and inference. Understands GPU fabric, east-west traffic, power and cooling constraints, and what happens when the pilot becomes production. | Has built at scale and can show the scars. Familiarity with reference architectures, not just vendor slides. | Among the hardest and most expensive to hire. Competing with hyperscalers and neoclouds. |
| AI security | Treats the model and the agent as workloads with identity, permissions, and blast radius. Governs what tools an agent can call, what data it can reach, and what it can do without a human. | Security fundamentals first, AI fluency second. Understands non-human identity and tool permissioning. | Emerging discipline. OWASP now publishes a separate Top 10 for Agentic Applications, on the premise that agent security is not LLM security (OWASP). |
| Harness and applied AI engineering | Builds the layer that gets data out of real applications and into a form a model can use. Orchestration, context management, retrieval, evaluation, and the unglamorous pipeline work in between. | Software engineering discipline plus systems thinking. Comfortable with messy source systems. | The market is now naming this role. Cognizant said in August 2025 it would train and deploy 1,000 "context engineers" over the following year to do exactly this kind of work (Cognizant). |
| Analytics and business consulting | Turns model output into a decision a business owner will act on. Frames the question, defines the measure, and knows when the answer is not trustworthy. | Business judgment and communication over algorithm depth. Can sit across from a CFO. | Widely available in title, narrow in practice. Most candidates can build a model; fewer can defend a recommendation. |
We have hired against all four. The infrastructure and security roles came from our existing bench, because that is our core. The harness work required a different skill set entirely, so we invested in building it. And most recently we brought in an analytics-focused profile aimed at the business side, to expand what our consulting practice can take on.
The unicorn is the trap
When a resume claims all four layers, treat that as a screening signal rather than a find. Depth in one of these areas takes years, and the candidates who are strongest in infrastructure usually have the least patience for business framing. That is fine. You are building a team, not hiring a person.
The practical version of this for a CIO writing a req:
- Name the layer before you name the title. If you cannot say which of the four you are staffing, you are not ready to post.
- Screen for production scars, not tool lists. Anyone can name the frameworks. Ask what broke and what they changed.
- Hire the data and integration roles earlier than feels comfortable. They are the constraint, and the market agrees they are the hardest to fill.
- Put security in the design conversation, not the review gate. An agent with credentials is an identity problem, and identity problems are cheap to prevent and expensive to discover.
Integration is the actual work
A widely cited MIT NANDA preprint found that roughly 95% of enterprise generative AI pilots delivered no measurable P&L impact. Its stated reason was not model quality. The authors called it a learning gap: tools that "don't learn, integrate poorly, or match workflows" (MIT NANDA, via Fortune). That paper is a preprint with a thin sample and its methodology has been criticized, so treat the 95% as a directional signal rather than a measurement. McKinsey's 2026 global survey points the same direction. Among the small group of organizations attributing meaningful EBIT impact to AI, 73% had fundamentally redesigned workflows, against 25% of everyone else (McKinsey).
None of that is a talent problem you solve with one hire. It is a systems problem you solve by staffing four layers and then doing the harder thing: making them work as one delivery model. Infrastructure that the security team designed with, a harness that the analytics team can actually query, and a consulting motion that closes the loop back to the business owner.
That is the part we are still working on, and we would rather say so plainly than pretend it is finished. If you are staffing an AI effort right now, the useful question is not who to hire. It is which of the four layers is your weakest, and what happens to the whole program if you leave it that way.
Business Technology Architects works with enterprises on AI infrastructure, security, and the integration work in between. If you are sizing an AI effort and want a second read on where your gaps are, we are happy to have that conversation.