The conventional description of the enterprise AI talent problem goes like this: companies want to deploy AI, AI requires machine learning engineers, machine learning engineers are rare and expensive, therefore companies cannot deploy AI.

This is wrong in a way that has led enterprises to pursue the wrong hiring strategies and make the wrong organizational investments for three years.

The talent that is actually scarce in enterprise AI deployments is not ML engineering. It is a cluster of adjacent capabilities that the industry does not have a clean name for: the ability to evaluate AI system outputs rigorously, design human-in-the-loop workflows that actually work, integrate AI outputs into existing enterprise systems and processes, build monitoring infrastructure for AI behavior, and translate between the technical limitations of AI systems and the operational requirements of business units.

Call it AI deployment engineering. The people who are good at it are genuinely hard to find, genuinely valuable, and almost entirely absent from the job descriptions that enterprises post when they decide they need AI talent.

Why ML Engineering Became the Default Hire

The fixation on ML engineers as the hiring target for enterprise AI capability has an understandable origin. ML engineers are the people who appear in the papers and the conference talks. They are the ones with the impressive GitHub portfolios, the competition rankings, and the PhDs from programs that produced most of the foundational AI research. They are legible to technical recruiters and compelling to boards.

They are also, for a large fraction of enterprise AI use cases, not the constraint.

Building a production RAG system that accurately answers questions from an enterprise document corpus does not primarily require ML engineering. The embedding models and vector databases are off-the-shelf. The LLM is accessible via API. What it requires is rigorous evaluation design (how do you know if the answers are correct, and how do you measure that at scale?), data pipeline engineering (how do you keep the document corpus current, structured, and accessible?), integration work (how do you connect the system to the document sources and the user interfaces it needs to work with?), and operational monitoring (how do you detect when the system starts degrading?).

None of those skills are ML engineering. They are a mix of software engineering, data engineering, evaluation design, and operational discipline — skills that exist in abundance in most large enterprises, but which are not being directed toward AI problems because organizations have not thought clearly about what AI deployment actually requires.

The Evaluation Problem

The most undervalued skill in enterprise AI is evaluation design: the ability to create rigorous tests that accurately measure whether an AI system is actually doing what it is supposed to do.

This is harder than it sounds, and its difficulty is not primarily technical. The technical components — building evaluation harnesses, running test suites, tracking metrics over time — are straightforward for any competent data engineer. The hard part is defining what “correct” means for the AI system’s outputs, which requires domain expertise, careful thinking about edge cases, and the ability to anticipate the failure modes that will matter most in production.

A customer service AI that answers product questions needs an evaluation set that tests not just common questions (which the AI will handle well) but the distribution of questions that real customers will actually ask — including misspelled queries, ambiguous requests, requests that span multiple products, questions where the correct answer has changed since the training data was assembled, and questions where no correct answer exists and the AI should acknowledge it does not know. Building this evaluation set requires someone who understands both the AI system’s failure modes and the customer service domain deeply enough to anticipate where they will intersect.

Most enterprises have not invested in this capability. Their AI evaluations are conducted using vendor benchmarks (which measure capability on problems the vendor selected, not problems the enterprise faces), internal demos (which test best-case scenarios, not adversarial ones), and user satisfaction surveys (which measure perceived quality, not actual quality). These evaluations produce optimistic assessments that do not predict production performance.

The organizations that evaluate AI seriously — that build genuine evaluation pipelines, maintain curated test sets, and run systematic evaluations before every significant system change — have AI systems that perform more reliably in production than their competitors’. This is not a coincidence. It is causation.

The Prompt Engineering Maturity Spectrum

One of the more surprising developments in enterprise AI talent is the emergence of a meaningful skill gradient in what the industry calls prompt engineering — the practice of designing the inputs to AI systems to produce reliable outputs.

In 2023, prompt engineering was treated as a curiosity at best and a fad at worst. By 2025, it had become clear that the quality of prompt design separates functional AI systems from unreliable ones in a consistent way that is not fully explainable by model capability differences.

Senior-level prompt engineers — people who can design system prompts that produce reliable outputs across a wide distribution of inputs, who understand how context window constraints affect output quality, who can architect multi-step prompting chains that handle complex reasoning, and who can debug prompt failures systematically — are genuinely rare and genuinely valuable. They are also almost entirely absent from enterprise hiring processes, which tend to classify this work as either an engineering task or a content task rather than recognizing it as a distinct specialization.

The practical result is that most enterprise AI systems are running on prompts that were designed hastily during the pilot phase, were not revised significantly for production, and have failure modes that are well-understood by individual engineers but not documented or systematically managed. When those engineers leave — and AI talent turnover is high — the institutional knowledge about prompt design leaves with them.

The Integration Engineering Gap

The most consequential talent gap in enterprise AI is the least glamorous: integration engineering — the ability to connect AI systems to the existing enterprise data sources, workflow tools, and user interfaces they need to be useful.

This work is technically demanding in a way that is invisible from the outside. It requires understanding the data schemas of legacy systems that may have been built in the 1980s or 1990s, the authentication and authorization models of enterprise platforms that were never designed to interact with AI, the operational characteristics of systems that need to provide low-latency data to an AI that needs to respond in seconds, and the reliability requirements of enterprise workflows that cannot tolerate AI system unavailability.

It also requires organizational fluency: the ability to navigate the vendor relationships, IT governance processes, and team boundaries that make every enterprise integration slower than it should be. The best integration engineers at the most AI-mature companies are not primarily technical specialists — they are people who can work effectively across organizational boundaries, who understand the priorities of the security team, the constraints of the data governance team, and the operational requirements of the business units, and who can design integrations that satisfy all of those simultaneously.

This profile is rare because it requires both technical depth and organizational navigation skill, and career paths in most enterprises have not developed it intentionally. The people who have it developed it accidentally, through years of difficult integration projects, and they are among the highest-leverage AI deployment resources an enterprise can access.

Building Capability vs. Hiring It

The competitive reality for most enterprises is that they cannot hire their way to AI deployment capability fast enough. The external market for AI deployment engineers — defined broadly — is tight, and the enterprises most willing to pay for the talent are the AI vendors and hyperscalers, not traditional industries.

The enterprises that have built real AI deployment capability have done it primarily through internal development: identifying software engineers, data engineers, and domain specialists with the aptitude and interest for AI deployment work, and investing in their development with a combination of project-based learning and structured external education.

This takes longer than hiring. It produces capability that is more deeply integrated into the organization’s specific context and systems. And it avoids the failure mode of the enterprise that hires a team of ML engineers, hands them a business problem, and discovers twelve months later that the team built something technically impressive that nobody in the organization knows how to maintain.

The right question is not “who do we need to hire?” but “what capabilities do we need to build, and who in our organization is the right foundation for building them?” The answer is almost never a team of ML PhDs. It is usually a combination of existing engineering talent with AI deployment skills layered on, domain experts who can evaluate AI outputs, and operations engineers who know how to run production systems reliably. That combination is buildable, if the enterprise is willing to invest in building it rather than waiting to hire it.

Get the best articles from Think Different

One email a week with practical Mac workflows, useful comparisons and honest tech recommendations.