Source check: September 24, 2026. Use current employer postings to identify the work, then prepare evidence that you can do it. A title, salary headline or model launch does not establish what every employer wants.
This chapter combines a small set of dated primary labor-market reports with specific employer examples. It is not a survey of all AI jobs and does not claim to have analyzed hundreds of listings. The role and preparation tables are a practical classification, not an official or exhaustive occupational standard.
Read the evidence before the headline
Different sources measure different things:
| Measure | What it describes | What it does not establish |
|---|---|---|
| Job postings | Advertised vacancies on a platform | Hires made, unique open seats or offers an individual will receive |
| Payroll employment | People employed in the covered records | Every employer or the cause of a change |
| Advertised salary | Pay range or amount shown in a posting | Accepted pay, total compensation or a national median |
| Survey forecast | Respondents' expectations | A measured future outcome |
| A specific role description | That employer's stated work and criteria | Universal requirements for everyone with the same title |
Three current observations
| Primary source and period | Finding | Appropriate interpretation |
|---|---|---|
| Indeed Hiring Lab, July 8, 2026 | US software-development postings had rebounded almost 15% since late February 2025, but remained about 27.5% below their pre-pandemic level. Senior roles accounted for 71% of the May 2025–May 2026 increase; AI-titled roles accounted for 37%, with overlap | Recovery from a low base is not a uniformly strong market. The two shares must not be added |
| Stanford Digital Economy Lab, August 12, 2026, ADP records through June 2026 | Employment for ages 22–25 in highly AI-exposed occupations was about 19% below the level implied by keeping pace with less-exposed peers. The adjustment appeared mainly in reduced hiring | A relative shortfall is not “19% of all junior jobs were eliminated.” The authors describe patterns, not causal estimates of AI's effect |
| Indeed Hiring Lab, September 17, 2026 | Advertised pay rose faster in more AI-exposed occupations, alongside a shift toward senior postings. The estimated premium shrank under alternative controls; the specification holding occupation/seniority mix constant was not significant at 5% | Do not convert this into a guaranteed AI salary premium. Advertised-pay composition and the chosen statistical model matter |
These studies cover different populations and outcomes. They support careful, role-specific research; they do not prove that one occupation is safe, that a title has disappeared or that every experienced AI specialist can command an offer.
Role taxonomy: classify responsibilities
| Work family | Titles you may encounter | Main responsibility | Useful preparation artifact |
|---|---|---|---|
| AI application engineering | AI engineer, applied AI engineer, LLM engineer | Integrate models into useful, reliable product behavior | A tested application with explicit failure and data boundaries |
| ML/model engineering | ML engineer, model engineer, research engineer | Data, training/adaptation, evaluation and deployment | Reproducible experiment against a meaningful baseline |
| Research | Research scientist, research engineer, member of technical staff | Investigate model behavior or develop new methods | Clear hypothesis, method, evidence and limits |
| Infrastructure and serving | AI platform, inference, ML systems, performance engineer | Capacity, runtime performance, reliability and deployment | A load/cost analysis and operational design |
| Evaluation | Evaluation engineer, backend engineer for evals, research engineer for evaluations | Datasets, graders, measurement and evaluation infrastructure | Versioned evaluation with failure analysis and uncertainty |
| Agent and integration engineering | Agent engineer, integrations engineer, MCP-focused engineer | Tools, execution state, permissions and recovery | A tool workflow that handles duplicates and unknown outcomes |
| Customer deployment | Forward deployed engineer, solutions engineer, applied engineer | Translate customer needs into working deployments | Discovery, integration and rollout with a concrete value measure |
| Security and reliability | AI security engineer, red-team specialist, reliability engineer | Threat analysis, controls, incidents and failure containment | A tested boundary or incident investigation |
| Leadership and delivery | Engineering manager, director, technical lead | Team capability, architecture, ownership and delivery | Truthful examples of decisions, delegation and outcomes |
| Product and program | AI product manager, technical program manager | User outcomes, experiments, prioritization and dependencies | A measurable product decision and delivery plan |
Titles overlap. Research is not confined to frontier labs, and a research-engineer title does not universally require a PhD or publication. “Member of technical staff” does not specify a level by itself. Model prompting, distillation and MCP can be responsibilities without being standalone job titles.
What current employer examples actually show
These examples were accessible on the source-check date. Openings and requirements can change or close.
| Employer example | Stated work or requirements | Preparation implication |
|---|---|---|
| OpenAI — Backend Software Engineer (Evals) | Support-automation evaluation infrastructure, backend services, data integration and reproducible monitoring; the posting names Python, FastAPI and Postgres and asks for backend experience | Connect ordinary backend engineering with rigorous evaluation; do not prepare only model trivia |
| OpenAI — Research Engineer, Frontier Evals & Environments | Model environments, experimental methods, measurement reliability/variance and scalable evaluation systems | Prepare hypotheses, experiments and measurement limits as well as implementation |
| Anthropic — Forward Deployed Engineer, London | Customer collaboration, production LLM applications, Python, communication and technical customer-facing experience | Rehearse discovery, integration constraints, deployment and explanation to a customer |
The interpretation in the last column is this guide's preparation advice. Three examples demonstrate variation; they cannot establish market-wide framework frequencies, demand rankings or a universal required stack.
Skills by scope, not a universal level number
An L5 or L6 at one employer may not mean the same scope as that label elsewhere. Years of experience can be an eligibility criterion, but do not alone determine level.
| Expected scope | Evidence to prepare | A question for the recruiter |
|---|---|---|
| Implement a bounded component | Correct code, tests, debugging and clear interfaces | Which language and implementation depth will be assessed? |
| Own a feature or service | Requirements, release, monitoring and failure recovery | What does end-to-end ownership include here? |
| Lead a broader architecture | Tradeoffs, migrations, cross-team contracts and capacity | Which teams and systems does this role influence? |
| Manage people and delivery | Hiring, coaching, delegation, priorities and accountability | What authority, staffing and operating responsibility come with the role? |
| Own research direction | Experimental design, evidence quality and judgment under uncertainty | What balance of publications, experiments and production work is expected? |
| Own product/program outcomes | User needs, success measures, dependencies and rollout | What decisions does the role own versus coordinate? |
Prepare technical depth relevant to the work. A serving role may need GPU memory and scheduling; an application role may need authorization, retrieval and product evaluation. Neither needs to pretend expertise in every layer of AI.
Turn a job description into a study plan
Read diagram source
flowchart LR
J[Current role description] --> R[Separate requirements from preferences]
R --> E[Map each responsibility to your evidence]
E --> G[Identify gaps and clarify scope]
G --> P[Practice relevant concepts and designs]
P --> M[Mock the confirmed interview format]
- Save the title, employer, location, posting URL and date checked.
- Separate explicitly required experience from preferred experience and vague marketing language.
- Identify the actual outputs: a service, trained model, evaluation system, customer deployment or team outcome.
- Match each responsibility to a real project, experiment or skill demonstration.
- Choose the two or three gaps most likely to block performance in that role.
- Confirm the interview format and permitted tools before rehearsing it.
| Responsibility in a posting | Study here | Practice evidence |
|---|---|---|
| Retrieval and grounded answers | RAG fundamentals | Explain ingestion, evidence, permissions and evaluation |
| Model adaptation | Fine-tuning | Compare adaptation with a prompt/retrieval baseline |
| Evaluation infrastructure | Evaluation-gated CI/CD | Version inputs, aggregate outcomes and decide a release |
| Agent integration | Tools and MCP | Design scoped tools, state and recovery |
| Performance and serving | Serving infrastructure | Calculate capacity and measure realistic latency |
| Customer deployment | Enterprise RAG case study | Connect a customer workflow to measured value and rollout |
| Leadership | Behavioral preparation | Explain actual ownership, decisions and people outcomes |
Interview tip: a small complete project with clear evidence can be more informative than several unfinished demos. This is a preparation principle, not a claim that a portfolio always outweighs a degree or research paper in hiring.
Compensation: compare like with like
Use a posting's own pay label. Base salary, total compensation, sales on-target earnings and contractor revenue are different quantities.
Dated examples, not market medians
| Employer and role | Location shown | Published annual amount | What is excluded or unresolved |
|---|---|---|---|
| OpenAI — Backend Software Engineer (Evals) | San Francisco and Seattle | US$266,000–445,000 base-pay range | Equity is offered separately; individualized pay and other compensation vary |
| OpenAI — Research Engineer, Frontier Evals & Environments | San Francisco | US$295,000–380,000 base-pay range | Equity is offered separately; this is one opening, not a research-role median |
| Anthropic — Forward Deployed Engineer | London | £225,000–255,000, labeled “Annual Salary” | Confirm package composition, level and terms with the recruiter; do not label it total compensation |
All three rows were checked September 24, 2026. They are selected employer examples, not a representative salary sample. Do not compare the GBP and USD figures without an explicit exchange-rate date and an understanding of taxes, benefits and location. Do not infer an offer at the top of a published range.
Build an offer-comparison worksheet
| Component | Record separately | Common error |
|---|---|---|
| Base pay | Currency, period and guaranteed amount | Comparing monthly with annual pay |
| Bonus | Target, eligibility, discretion and payment timing | Treating a target as guaranteed |
| Sign-on | One-time amount, installments and repayment terms | Counting it in every future year |
| Equity | Instrument, units, vesting, exercise/settlement and liquidity terms | Treating a paper value as cash salary |
| Benefits | Retirement, insurance, leave and applicable costs | Ignoring material package differences |
| Work arrangement | Location, travel, on-call and employment type | Treating “remote-friendly” as work from any country |
| Review process | Level, scope, refresh and promotion criteria | Assuming a title guarantees future progression |
For a hypothetical arithmetic example, $180,000 base plus a 10% target bonus is $198,000 in base plus target bonus. A $20,000 sign-on makes the first-year figure $218,000 if paid in that year; it does not make recurring annual pay $218,000. Equity is additional and must be described under its actual terms. This is a comparison worksheet, not personal tax or investment advice.
It is reasonable to ask a recruiter about the budgeted range and level early enough to establish mutual fit. Detailed negotiation timing varies; there is no universal rule that compensation may only be discussed after an offer.
Geography and industry change the role
Do not infer worldwide opportunity from a few US technology-company postings. Research the location and sector you can actually work in.
| Dimension | Verify before comparing roles |
|---|---|
| Country and work authorization | Eligible location, employment entity and sponsorship policy |
| Remote, hybrid or customer-site | Required office presence, travel and time-zone overlap |
| Industry | Domain knowledge, source-data access and consequence of mistakes |
| Operating environment | Managed APIs, cloud infrastructure, on-premises or edge constraints |
| Regulated work | The specific security, privacy or sector requirements and qualified owners |
| Customer responsibility | Integration work, support expectations and continuing ownership |
Regulatory requirements depend on the application and jurisdiction. SOC 2, HIPAA, FedRAMP and the EU AI Act are not interchangeable credentials or rules that every AI engineer must personally “certify.” Use the governance chapter for current distinctions.
Interview process patterns: confirm the actual loop
A process may include recruiter discussion, coding, system design, a project deep dive, experiments, a take-home, behavioral questions or a practical build/review. Employers combine these differently; whiteboards and ordinary programming fundamentals have not universally disappeared.
A documented example is Sierra's April 22, 2026 account. It describes a Plan → Build → Review onsite with a two-hour AI-assisted build, and a system design interview replacing its coding phone screen. It also describes a debugging format being piloted, with the degree of AI use still under consideration. That is one employer's published process, not permission to use AI in another company's interview.
| Ask before the interview | Why it matters |
|---|---|
| Which rounds and durations apply to this role? | Avoid preparing for a generic process that is not used |
| Is it coding, design, experimentation, debugging or a project discussion? | Practice the actual kind of evidence requested |
| Which language, tools, references and AI assistance are allowed? | Follow explicit rules; do not assume everyday tools are permitted |
| How should AI assistance be disclosed? | Keep your contribution and validation clear |
| What is the take-home time budget and expected scope? | Avoid unbounded unpaid work or polishing the wrong artifact |
| What environment or accessibility arrangements are available? | Resolve practical constraints before the timed session |
| What criteria distinguish this level from adjacent levels? | Match the expected responsibility and depth |
For a permitted AI-assisted task, be able to explain and test the resulting code, identify failures, and separate working behavior from unverified assumptions. A generated artifact is not evidence that you understand it.
Specialized responsibilities to watch
| Responsibility | Durable skill underneath | Avoid overinterpreting |
|---|---|---|
| Forward deployment | Customer discovery, integration and delivery | A universal pay premium, travel rule or revenue threshold |
| Evaluation engineering | Measurement, data curation, graders and reliable pipelines | A requirement for one specific evaluation framework |
| Agent systems | Tool contracts, state, permissions and recovery | A need for multiple agents in every design |
| AI reliability | SLOs, incidents, controlled degradation and cost containment | Complete separation from existing SRE/platform work |
| AI security | Threat modeling, isolation, abuse testing and response | Model refusals or a single filter as sufficient protection |
| MCP integrations | Protocol compatibility, authorization and service design | Protocol expertise replacing distributed-systems fundamentals |
| Computer-use workflows | Interface observation, execution verification and constrained actions | A fixed set of product names or guaranteed automation accuracy |
A new label can identify useful work, but its long-term prevalence needs evidence. Learn the underlying contract so your expertise survives framework and title changes.
For hiring managers and engineering leaders
- Define the outcomes and authority of the role before listing tools.
- Separate necessary day-one skills from skills the team can teach.
- Use consistent, relevant assessment criteria and make permitted assistance clear.
- Size the assessment to the actual responsibility and avoid requesting confidential prior work.
- Give evaluation and release decisions appropriate independence without assuming every team needs a separate evaluation department.
- Account for operating work, incident coverage, source ownership and onboarding in staffing.
- Calibrate compensation to the actual level, location and package; do not apply an invented universal “AI premium.”
A practical exercise can be useful when it measures job-relevant skills, but this chapter does not establish that one interview format is universally more predictive than all others.
Interview and career-research questions
1. Does growth in AI-related postings prove that getting an AI job is easy?
No. It measures a platform's advertisements, not your eligibility, competition, hires or offers. Look at current roles matching your location and responsibilities, and distinguish a rebound from the starting level.
2. Can a payroll shortfall be described as an equal percentage of layoffs?
No. A relative employment gap can arise from reduced hiring, separations or other changes. Preserve the study's population, comparison and stated limitations.
3. Why can advertised pay rise even when entry opportunities are weak?
The mix can shift toward more senior or higher-paying roles. Compare like levels and occupations, and distinguish advertised pay from pay received by existing workers.
4. Does “AI engineer” have one globally accepted job description?
No. The title can cover applications, model work, infrastructure or a combination. Classify the actual responsibilities and ask what outcomes the role owns.
5. Does one evaluation-engineer listing establish the salary for all evaluation roles?
No. Product evaluation infrastructure and research measurement can have different scope, locations and levels. A posting range is evidence about that role, not a market median.
6. How should a backend engineer prepare for an AI application role?
Build on APIs, data, authorization and reliability, then add model evaluation, evidence handling and model-specific failure analysis. Select gaps from the actual posting rather than assuming every role requires training a model.
7. What should a serving candidate demonstrate beyond knowing model names?
Workload sizing, memory and scheduling behavior, quality-compatible optimization, load testing, failure recovery and full costs. The exact hardware/runtime requirements come from the role.
8. Is a PhD always required for a research-engineer title?
No universal rule applies. Read the stated qualifications and work. Some roles emphasize experimentation and implementation; others require particular research credentials or publications.
9. How should you compare a one-time sign-on with recurring pay?
Record them separately by year and verify repayment conditions. Do not count the same sign-on every year or treat a target bonus as guaranteed cash.
10. Does remote-friendly imply permission to work from any country?
No. The employment entity, eligible locations, tax/work-authorization rules, travel and time-zone requirements still matter. Confirm the specific arrangement.
11. Can an AI-assisted interview example authorize tool use elsewhere?
No. Follow the actual employer's rules for that round, including disclosure. Ask before using assistance if the rule is unclear.
12. How do you choose between learning a new framework and strengthening a concept?
Identify the task you need to perform. Learn the concept and contract, then the framework features required to implement it. A version-specific skill is useful when the target role needs it, but popularity alone is insufficient.
13. How can a portfolio support an application without overstating experience?
Show a complete, bounded project with a baseline, tests, evaluation and limitations. Label it as a personal project; do not claim production users or business outcomes you did not observe.
14. What should a manager clarify before hiring an AI specialist?
The work, decision authority, current team gap, dependencies, operating burden and assessment criteria. Decide whether hiring, internal development or changed scope best addresses the need.
15. What makes a job-market statement credible?
A dated source, defined population and metric, appropriate comparison and explicit limits. Keep descriptive evidence, causal claims and forecasts separate.
Final summary and notes
| Remember | Practical action |
|---|---|
| Titles vary | Read responsibilities and confirm scope |
| Metrics differ | Separate postings, employment, pay and forecasts |
| Salary needs context | Record currency, location, level and package components |
| Interview rules vary | Confirm format and permitted tools |
| Skills transfer through concepts | Practice contracts, evidence and decisions |
| Claims need provenance | Keep the source URL and date checked |
Use the transition guide to map existing skills, the practice hub to plan study, and behavioral preparation to present your actual experience.