Meet the New Class of AI Jobs Being Built Inside Ordinary Companies

The next important AI hire at a bank may not be a machine-learning engineer. It could be a compliance specialist who knows how to test an automated decision, a product manager who can redesign a lending process around AI, or an analyst expected to prove that an expensive software rollout has produced more than a collection of impressive demonstrations.

These jobs are beginning to appear because companies have moved beyond the first phase of generative AI adoption. Buying licences and allowing employees to experiment with chatbots was relatively simple. Connecting AI to customer records, internal documents and operational systems is not. Once a tool begins drafting client communications, screening transactions or completing parts of a regulated process, somebody must decide what it is allowed to do, measure whether it works and take responsibility when it does not.

That requirement is creating a class of jobs that sits between technology and the rest of the business. Some carry unfamiliar titles, including AI agent specialist, model evaluator and AI governance lead. Others are familiar positions acquiring a more technical remit: lawyers who understand model risk, marketers who design human-AI production systems and operations managers who can divide work between employees and software agents.

The market is not replacing every existing profession with an AI equivalent. It is inserting a new layer of work into companies that must now manage systems capable of producing, recommending and increasingly acting.

Most new AI jobs are translation jobs

The public image of an AI career still tends to involve model development, advanced mathematics and large technology companies. Demand for machine-learning specialists, data engineers and AI researchers remains strong. The World Economic Forum identifies AI and machine-learning specialists among the fastest-growing occupations through 2030, alongside big-data specialists and fintech engineers.

Yet those roles represent only one part of the employment effect. A company using an existing model does not necessarily need to build its own. It does need people who can translate a business problem into a workable system.

Consider an insurer attempting to use AI in claims handling. The technical question is whether a model can read forms, photographs and customer correspondence. The operational questions are harder. Which claims can be processed automatically? What evidence requires human review? How should low-confidence outputs be handled? Could the system disadvantage a particular customer group? How will an employee explain the decision if it is challenged?

A conventional software team cannot answer those questions alone. Claims specialists understand the process, compliance officers understand the obligations, data teams understand the available information and product managers decide how the pieces fit together.

This explains why many of the most valuable AI roles will be hybrid rather than purely technical. The employee who understands both the model and the business process becomes the connection between what AI can do and what the organisation can safely permit it to do.

The AI product manager

The AI product manager is likely to become one of the more durable positions to emerge from the current cycle.

Traditional product managers define customer needs, prioritise features and coordinate design, engineering and commercial teams. An AI product manager must do that while working with a system whose output is probabilistic rather than fully predetermined.

A normal software feature should produce the same result when given the same instruction. A generative model may produce different answers, make unsupported claims or perform well on routine requests while failing on an unusual but commercially important case.

The job therefore involves deciding where uncertainty is acceptable. An internal research assistant can tolerate a different risk threshold from an automated system sending financial guidance to customers. A marketing tool may be allowed to create a first draft that an employee reviews. A fraud-detection system affecting whether a transaction is blocked requires tighter testing, auditability and escalation.

This role does not require the product manager to train a foundation model. It requires enough technical knowledge to understand model limits, enough commercial knowledge to identify a useful application and enough operational judgement to design the human safeguards around it.

The strongest candidates are unlikely to be those who have merely completed a course in prompt writing. They will be people who can demonstrate that they have improved a process with AI and measured the result.

The workflow and agent architect

As companies begin using AI agents rather than standalone chatbots, workflow design becomes a profession in its own right.

An agent can be given a goal, access tools and complete a sequence of actions. In customer service, it might identify a client, retrieve account information, classify the request, propose an answer and update a record. In procurement, it might review contracts, compare suppliers and prepare a purchase request.

The difficult part is not instructing the agent to “resolve the customer’s problem”. It is defining each action it can take, the information it may access and the points at which a person must intervene.

A workflow architect maps that process. The role combines elements of business analysis, automation, systems integration and organisational design. The architect decides which tasks should remain human, which can be delegated and how work should return to an employee when the system encounters an exception.

This is also where many AI projects will succeed or fail. A company can purchase an advanced model and still create little value if it inserts the tool into a badly designed process. Giving employees a chatbot without changing approvals, responsibilities or information flows often saves minutes rather than transforming the work.

The workflow architect is not simply automating existing steps. A good one asks whether those steps should exist at all.

The model evaluator and AI quality lead

Software testing traditionally checks whether a system performs according to a defined specification. AI evaluation is less straightforward because the quality of an answer may depend on accuracy, relevance, tone, safety and context.

A customer-service response can be factually correct but inappropriate. A legal summary may capture the general meaning while omitting the clause that changes the commercial risk. A recruitment tool might appear accurate overall while performing less reliably for particular groups.

Model evaluators create test sets, define acceptable performance and investigate recurring failures. They may compare models, assess outputs produced in different languages and test how a system behaves when given misleading, incomplete or adversarial instructions.

This work can involve technical tools, but much of its value comes from domain expertise. A general evaluator can check grammar and consistency. A physician is needed to judge whether an AI-generated clinical summary has omitted a medically significant detail. A financial professional can identify when a plausible market explanation is economically incoherent.

The position may sit within product, risk, data science or operations depending on the company. In regulated industries, it is likely to become increasingly formal because firms will need evidence that a system was tested before deployment and monitored afterwards.

Someone must decide what “good enough” means. That is not a question the model can answer for itself.

The AI governance and model-risk specialist

Companies once treated AI governance as a policy exercise: write principles, create a committee and prohibit employees from entering confidential information into public tools.

That approach becomes inadequate when AI is embedded in operational decisions.

Governance specialists inventory the models and AI tools being used across the company, classify them by risk and establish who is accountable for each system. They examine data protection, intellectual property, discrimination, explainability, vendor dependence and regulatory obligations.

The role becomes particularly important when employees adopt tools without formal approval. A marketing department may connect a consumer AI service to customer data. A sales team may use an automated meeting assistant that stores recordings with an external provider. A developer may incorporate model-generated code without understanding its provenance or vulnerabilities.

The governance specialist must find these uses without paralysing the organisation. A regime that requires six months of approval for a low-risk writing assistant will encourage staff to work around it. A regime that treats every application as harmless creates legal and operational exposure.

The work suits people from compliance, law, audit, privacy, cybersecurity and regulated product management. They need enough technical fluency to challenge vendors and understand how a system operates, but their main value lies in converting rules into practical controls.

The European Union’s AI Act is accelerating this demand, but regulation is only one driver. Boards and insurers increasingly want evidence that companies know where AI is being used and who is responsible for it.

The AI security specialist

AI adds security problems that do not fit neatly within conventional cyber defence.

Models can be manipulated through prompt injection, in which malicious instructions hidden in a document or webpage cause an AI system to behave unexpectedly. Sensitive information can leak through poorly designed retrieval systems. Agents with access to email, files or financial systems can take damaging actions if their permissions are too broad.

An AI security specialist examines these new attack surfaces. The job may involve testing models, limiting access to data, monitoring agent activity and designing controls around external tools and model providers.

The risk grows when an AI system is allowed to act rather than merely advise. A chatbot that produces a poor draft creates inconvenience. An agent that can approve refunds, change account details or execute code creates a security event.

This is likely to become a specialism within cybersecurity rather than a wholly separate profession in every organisation. Large companies may create dedicated teams. Smaller businesses will expect existing security professionals to acquire AI-specific skills.

Either way, knowledge of ordinary cyber hygiene will not be enough. Security teams will need to understand how models interpret instructions and how a legitimate business process can be turned against the system.

The data steward and knowledge engineer

Generative AI has made companies rediscover a problem they often postponed: their internal information is fragmented, duplicated, outdated and difficult to retrieve.

A model connected to poor information does not solve that problem. It produces answers from poor information more quickly.

Data stewards establish who owns important datasets, what can be used, how long it should be retained and whether it is accurate. Knowledge engineers organise documents, terminology and relationships so an AI system can retrieve the right material. They decide which policy is current, how product names should be classified and which sources should take precedence when records conflict.

This work is less glamorous than model development, but it can determine whether an internal AI assistant is trusted. An employee will stop using a tool that repeatedly retrieves obsolete policies or confuses similarly named clients. Once confidence is lost, technical improvements may not be enough to restore adoption.

The role also requires organisational authority. Cleaning data is not a one-off project performed by an IT team. Business units must maintain the information they create and accept common standards. The knowledge engineer therefore works as much with people and ownership as with databases.

Companies that claim they are not ready for AI because their data is poor may discover that preparing the data is itself one of the largest sources of new work.

The AI trainer is not a chatbot tutor

The phrase “AI trainer” can describe several different jobs, some more durable than others.

At model companies, trainers may label data, compare outputs and provide specialist feedback used to improve a system. Inside an ordinary business, the role is more likely to involve teaching an AI system the organisation’s rules and teaching employees how to use it.

A trainer might work with customer-service managers to define examples of a satisfactory answer, identify language the company must avoid and establish when the tool should escalate a case. The same person may train staff to verify outputs rather than accepting them automatically.

This differs from providing a generic workshop on writing prompts. Employees need guidance connected to their actual work: which tool to use, which data may be entered, how an answer should be checked and what evidence of the process must be retained.

The role may not survive as a standalone title in every company. Training responsibilities could move into learning and development, product operations or individual departments. The underlying work will remain because AI systems and employee practices both change too quickly for a one-off course to be sufficient.

The AI value and ROI analyst

One of the least glamorous new AI jobs may prove to be among the most important.

Companies have invested heavily in licences, pilots and consulting projects without always establishing what success should look like. Time saved is frequently estimated from employee surveys. Usage is treated as evidence of productivity. A polished demonstration is mistaken for an operating result.

An AI value analyst tests those claims.

For a customer-service tool, the analyst might examine resolution time, repeat contacts, customer satisfaction, errors and escalation rates. For a coding assistant, the measures could include delivery speed, defects, security findings and the time senior engineers spend reviewing generated work. A document-processing system should be judged not only by how quickly it extracts information but by the cost of correcting mistakes.

The analysis must include the full implementation cost: licences, integration, data preparation, security, training, supervision and ongoing evaluation. A tool that saves employees ten minutes a day may not create a financial return if the time is fragmented and cannot be converted into additional output or lower cost.

This role can sit in finance, strategy, operations or a central AI office. It rewards scepticism as much as enthusiasm. The analyst’s job is not to prove that AI works, but to identify where it does.

Prompt engineer may not become the profession once imagined

In 2023, prompt engineer was frequently presented as the emblematic job of the generative AI economy. Some early vacancies advertised unusually high salaries for people who could design instructions for language models.

Prompt design still matters. Clear instructions, examples, structured outputs and testing can materially improve performance. But prompting is increasingly becoming a skill within other jobs rather than a profession on its own.

A lawyer using AI needs to frame a legal research task well. A marketer needs to brief a content system. A product manager must define how an agent should behave. As interfaces improve and models become better at interpreting ordinary language, the value shifts from knowing a collection of prompt techniques to understanding the underlying domain and workflow.

The durable advantage is not the ability to make an AI produce an attractive answer. It is knowing whether the answer is useful, permissible and correct.

Existing jobs will change faster than new titles appear

Microsoft reported in its 2025 Work Trend Index that 78 percent of surveyed leaders were considering hiring for new AI roles. The positions under consideration included AI trainers, data and security specialists, agent specialists, ROI analysts and AI strategists across functions such as finance, marketing and customer service.

That does not mean companies will create a separate department for each one. In many organisations, the work will be distributed across existing teams.

A finance professional may become responsible for AI investment cases. A compliance manager may take ownership of model governance. A business analyst may begin designing agent workflows. A content manager may supervise AI quality and provenance.

LinkedIn’s research suggests that 70 percent of the skills used in most jobs could change between 2015 and 2030, with AI acting as one of the catalysts. Its 2026 labour-market analysis found that US job postings requiring AI literacy had grown 70 percent year on year.

The implication is more demanding than a simple boom in specialist hiring. Workers may not need “AI” in their title, but they will increasingly be expected to understand how the technology changes their function.

Domain expertise is becoming more valuable, not less

The current market creates an apparent contradiction. AI can perform more professional tasks, yet companies also need people with enough experience to assess its work.

An inexperienced employee may use a model fluently but fail to detect a plausible error. A senior specialist may understand the error immediately but lack the confidence to redesign the process around the tool. The strongest candidates combine both capabilities.

Anthropic’s early workplace research found AI use concentrated in parts of occupations rather than across complete jobs. Its initial analysis suggested that only a small share of occupations used AI across most of their tasks, while a much larger group used it for a meaningful minority of work. More recent data indicates that the breadth of task use is continuing to expand, although adoption remains uneven across professions and countries.

This is why the immediate labour-market effect is likely to involve the reassembly of jobs. Some tasks disappear, others expand and new responsibilities are added around supervision, judgement and system design.

Entry-level work presents the more difficult question. Junior employees have traditionally learnt through research, drafting, reconciliation and other tasks that AI can now complete quickly. If companies automate too much of this work without creating a new training model, they may reduce the pipeline from which future specialists are developed.

AI may make an experienced professional more productive while making it harder for an inexperienced one to become experienced. Companies hiring for the new AI economy will eventually have to solve that problem too.

How to prepare without becoming an AI engineer

For most professionals, the sensible response is not to abandon their field and begin training as a model developer. It is to identify where AI is changing the economics and workflow of the field they already understand.

A communications professional could learn how to build an AI-assisted content process with approval, disclosure and fact-checking controls. A compliance specialist could study model inventories, risk classification and the EU AI Act. An operations manager could map a process and determine which decisions are suitable for automation. A financial analyst could develop a method for measuring AI investment returns.

A credible portfolio matters more than a collection of generic certificates. Candidates should be able to show how they analysed a process, chose an appropriate tool, identified the risks and measured the result. Even a small internal project can demonstrate more judgement than claiming broad familiarity with several AI platforms.

Technical fluency helps, particularly an understanding of data, APIs, retrieval systems and model evaluation. It does not replace professional knowledge.

The emerging class of AI jobs is being built at the boundary between the two. Companies have enough access to models. What they lack are people who know where to put them, how to control them and when the result is not worth the cost.