The job titles that mattered three years ago no longer describe the work. Prompt engineer is fading as a title. Data scientist is being redefined. Meanwhile a set of hybrid roles has appeared that sits between product thinking and engineering execution, and most companies do not yet have language for them.
This matters practically. When you post for an “AI engineer” and receive 400 applications, roughly none of them are for the job you meant. The title has become a hiring priority rather than a precise role.
Below are eight emerging AI roles that now exist as distinct hiring categories, what each actually does, and how to tell a real one from a relabelled resume. Rise92 sources across all eight for global engineering teams.
The eight roles at a glance
| Role | Core deliverable | Hire when |
| AI Engineer (applied) | Shipped features using models | You have any AI surface in your product |
| Forward Deployed Engineer | Working system inside a customer’s environment | You sell AI to enterprises |
| AI Product Manager | Evaluation criteria and product definition | You shipped one AI feature and it behaved unpredictably |
| AI Evals Engineer | Test suites for non-deterministic output | You cannot tell whether a change improved anything |
| Agentic AI Engineer | Multi-step agent systems in production | Your use case needs action, not just answers |
| Context Engineer | Retrieval and context pipelines | Your model has the capability but not the information |
| LLMOps / AI Infrastructure Engineer | Reliable, observable, affordable inference | Your costs or latency are out of control |
| AI Solutions Architect | A defensible system blueprint | You are choosing an architecture you will live with for years |
What actually changed in the market
The shift is from training models to deploying and evaluating them. That single change created every role on this list.
The demand data is unusually consistent across sources. AI skills appeared in 2.5% of US job postings in 2025, up 55% year over year and representing roughly a million postings, per Lightcast data in the Stanford AI Index 2026. PwC’s AI Jobs Barometer 2026 counts differently, tracking AI specialist postings, and found 68.9% growth in 2025 against 8.6% for the market overall.
Compensation reflects the same pressure. PwC puts the wage premium for AI skills within the same occupation at 62%, up from 56% a year earlier. Lightcast’s method gives a lower figure of 28%, roughly $18,000 a year. Different methods, same direction.
Two structural facts worth holding onto. Half of all AI postings now sit outside IT departments, which means these roles increasingly report into product and business functions rather than engineering alone. And the machine learning engineer title remains the most common AI role in Europe at 64% of AI engineering postings, so the older taxonomy has not disappeared, it has been joined.
The AI talent shortage underneath all of this is specific rather than general. There is no shortage of people who have used a model. There is a severe shortage of people who have shipped one into production, watched it fail, and fixed it.
1. AI Engineer (applied)
The generalist who puts models into products. Retrieval pipelines, agents, fine-tuning where warranted, and the evaluation loop around all of it.
- Growth: US postings rose 143% year over year in 2025, and LinkedIn ranked it the fastest-growing job title in the US for 2026
- Not the same as: a research scientist. This role ships features, it does not train foundation models
- Real signal: a deployed system with users, and a specific account of what broke
- False signal: course certificates, notebook demos, model architecture theory with no production history
- Reports to: engineering, usually, though increasingly product
Hiring note: this is the one role almost every company with an AI surface genuinely needs. Hire the generalist first, specialise later.
2. Forward Deployed Engineer
Embedded with a customer, shipping code inside an environment they did not build. The data pipelines belong to someone else, the compliance rules are the client’s, and the domain knowledge sits with people who may not speak technically.
- Growth: postings grew over 800% year over year in 2025, and in May 2026 the three largest AI labs each made structural hiring bets on this model within weeks of each other
- Origin: Palantir, which until 2016 employed more forward deployed engineers than software engineers
- Why it exists: enterprise AI deployments fail organisationally rather than technically, and an embedded engineer clears those blockers faster than any other vendor role
- Real signal: has sat in a client’s office, rebuilt the plan after discovering the data was worse than promised, and shipped anyway
- Hardest part to assess: the client-facing layer, not the technical one
Hiring note: the highest salary ceiling on this list and the least competition, partly because most engineers do not want the travel or the ambiguity.
3. AI Product Manager
The product specialisation that treats model behavior, evaluation, and context strategy as first-class product surfaces rather than implementation details.
- Why it emerged: at companies that shipped one AI feature and discovered that putting a model behind a generic product manager did not produce predictable outcomes
- Core difference from a traditional PM: the underlying capability is non-deterministic, so the spec cannot describe the output, only the acceptable range of outputs
- Real signal: has written evaluation criteria for a probabilistic feature and made a shipping decision against them
- False signal: has used the tools enthusiastically and added “AI” to a standard product resume
- Common background: product management, frequently with a technical foundation
Hiring note: hire this role the day your AI feature gets its second unexplained regression.
4. AI Evals Engineer
Builds the measurement layer. Without it you cannot tell whether last week’s change helped, and every release becomes a judgment call.
- Status: growing fast as a formal hiring category rather than a task absorbed by other engineers
- Deliverable: domain-specific test suites, regression detection, and quality metrics that survive contact with real usage
- Real signal: has built an eval suite that caught a regression nobody else noticed
- False signal: describes evaluation entirely in terms of published benchmarks
- Tooling context: the agent evaluation toolchain, including platforms such as LangSmith and Braintrust, has matured enough to support a dedicated role
Hiring note: the most commonly skipped role on this list, and the one whose absence causes the most expensive downstream failures.
5. Agentic AI Engineer
Builds systems that take multi-step actions rather than returning single responses. Tool use, planning, error recovery, and the guardrails around all three.
- Growth: the agentic AI skill went from 0.06% to 0.23% of US postings in a single year, an increase of 280%, bringing the total to roughly 90,000 postings
- The hard part: failure handling. A single wrong answer is a bad response. A wrong action in a multi-step chain compounds into a real incident
- Real signal: can describe a specific agent failure mode they designed around, with the guardrail they built
- False signal: has chained a few framework calls together in a demo
- Adjacent requirement: strong Python and LLM application experience, though not the ability to train models from scratch
Hiring note: demand is running well ahead of genuine production experience here. Assess the failure-handling reasoning, not the framework familiarity.
6. Context Engineer
The role that absorbed most of what prompt engineering was supposed to be, with a much larger scope. Retrieval quality, context assembly, chunking strategy, and getting the right information to the model at the right moment.
- Why the title shifted: prompt engineering as a standalone job title is fading, while the underlying discipline expanded into systems work
- Core insight: most model failures in production are information failures rather than capability failures
- Real signal: can explain a retrieval quality problem they diagnosed and the specific change that fixed it
- False signal: a prompt library and strong opinions about phrasing
- Skill lean: systems and operations rather than deep machine learning
Hiring note: if your model “hallucinates,” you probably need this role rather than a better model.
7. LLMOps and AI Infrastructure Engineer
Deployment, monitoring, cost control, latency, and versioning for probabilistic systems. Traditional operations practice does not transfer cleanly, because the thing being monitored has no deterministic correct output.
- Demand: consistent across company types rather than concentrated at labs, and among the highest-volume hiring categories globally
- Core tension owned: the cost, latency, and quality triangle, which someone has to manage explicitly
- Real signal: has cut inference cost substantially without degrading output quality, and can show how they measured it
- False signal: general DevOps experience with model-serving added as a line item
- Skill lean: systems and operations, foundational machine learning knowledge helpful but not essential
Hiring note: the role that pays for itself fastest, because inference costs at scale are usually larger than anyone budgeted.
8. AI Solutions Architect
The deliverable is a defensible blueprint rather than running code. Which model, which deployment pattern, where the boundaries sit, and what the system will cost at ten times current volume.
- Why it emerged: every multi-agent system requires architectural decisions that an applied engineer is not positioned to make alone
- Compensation: typically sits $20,000 to $40,000 above the applied engineer band in equivalent markets
- Real signal: has made an architecture decision, lived with it for two years, and can explain honestly what they got wrong
- False signal: vendor-shaped recommendations and diagrams with no cost model attached
- Watch for: architects who have never operated what they designed
Hiring note: only hire this separately once you are making multi-year commitments. Before that, your strongest applied engineer covers it.
Which of these you actually need
Most companies need two or three, not eight.
| Situation | Hire |
| First AI feature, small team | AI Engineer only |
| Shipped one feature, unpredictable behavior | AI Engineer + AI Evals Engineer |
| Selling AI into enterprise accounts | Forward Deployed Engineer |
| Multiple AI surfaces, no clear owner | Product specialist + AI Engineer |
| Model has capability but wrong outputs | Context Engineer |
| Costs or latency out of control | LLMOps / Infrastructure Engineer |
| Choosing a multi-year architecture | Solutions Architect, possibly fractional |
| Agents taking real actions | Agentic AI Engineer + evals |
The sequencing that works: generalist first, evals second, specialist third. Hiring specialists before you have a generalist who understands your specific failure modes produces expensive, uncoordinated work.
Why these roles are hard to hire
Three reasons, and only one of them is about supply.
The title problem. Roughly 70% of qualified people apply under the wrong title, according to one 2026 analysis. Your posting for an AI engineer is being ignored by the context engineer you actually need, and answered by people whose experience stops at the demo stage.
The evidence problem. These roles hire on delivery evidence rather than credentials, meaning repositories, deployed agents, and eval suites. Standard resume screening is close to useless here, because the signal is not on the resume.
The access problem. The people with genuine production experience are employed, well paid, and not looking. They are not reading job boards, and cold outreach to them has a low response rate because everyone is sending it.
That last one is the real constraint, and it is why hiring for hybrid AI and product engineering roles usually needs a network route rather than a posting route. The relevant approach is off-market sourcing through closed networks rather than broader advertising.
Where the supply actually is: Pakistan market data counts 300,000+ export-ready engineers across AI, cybersecurity, blockchain, and cloud. The honest caveat is that senior depth in production AI specifically is thinner than the headline figure suggests, which makes the sourcing method matter more, not less.
What these roles cost
Comparative figures, with the usual warning that AI compensation data moves quickly and sources diverge.
| Role band | US market indication |
| AI Engineer, average 2026 | Reported around $206,000, up roughly $51,000 from 2025 |
| Entry level | From around $143,000 |
| Senior | $269,000+ |
| Solutions Architect | $20,000 to $40,000 above the applied engineer band |
| Wage premium for AI skills, same occupation | 62% per PwC, 28% per Lightcast |
Two implications for budgeting. First, the premium is large enough that role definition has direct financial consequence, since hiring a solutions architect for work an applied engineer could do costs you real money annually. Second, building these roles in a different market changes the economics substantially, though only if the employment structure is compliant and the sourcing reaches genuine senior depth. Full cost comparison sits in transparent pricing.
Frequently asked questions
They are real as hiring categories, evidenced by distinct posting growth, distinct compensation bands, and distinct skill requirements. Forward Deployed Engineer postings grew over 800% in a year and agentic AI appeared in 280% more postings, which is not relabelling. That said, several overlap heavily at small companies, where one person covers three of them.
An applied AI engineer if you are building products on existing models, and a machine learning engineer if you are training or substantially fine-tuning your own. The confusion is expensive because the interview processes differ completely, and running the wrong one filters out exactly the people you needed. Decide which before writing the job description.
Once you have shipped one AI feature and discovered that a generic product spec does not produce predictable behavior. The distinguishing requirement is comfort defining acceptable output ranges rather than exact outputs, and making shipping decisions against evaluation criteria. Before that point, your existing product team plus a strong engineer covers it.
Not as a standalone job title, which is fading, though the underlying discipline expanded into context engineering with a much broader systems scope. If you are hiring for prompt work specifically, you probably need someone who can build retrieval and context pipelines instead. The skill persists, the title does not.
The practical takeaway
The taxonomy matters more than it sounds. These emerging AI roles carry different skills, different interview processes, and compensation bands separated by tens of thousands of dollars. Getting the title wrong means running the wrong assessment against the wrong pool and paying for a capability you did not need.
Start by naming the problem rather than the role. Unpredictable output means evaluation. Wrong answers with the right capability means context. Rising costs mean infrastructure. Enterprise deals stalling at deployment means a forward deployed engineer.
Then accept that the people who have actually done this work are employed and not searching, which makes access the binding constraint rather than process. To talk through which of these roles your situation actually calls for, get in touch.
