Updated November 2026 · Written and maintained by the Progression Agency strategy team
Most failed attempts to hire AI developers start with a job post that names a tool instead of a problem. AI work sits in layers: data, retrieval, model calls, trained models, evaluation, the product people touch, cost and guardrails. The layer where your problem lives decides which engineer you need, what to test and how to engage them. This page maps the layers, gives interview questions that separate production experience from demos, and sets out market rate bands. Progression’s own developers take on this work, and the team is based in New York.
On this page · 14 sections
- A job post written from a demo
- What AI developers do, layer by layer
- Artificial intelligence developers in the wage data
- Gen AI developers and LLM application skills
- Machine learning engineers for models you train
- The evaluation set as a hiring test
- Interview questions for AI engineers
- Engagement models when you hire AI developers
- AI developers for hire: market rate bands
- Who owns code written with AI tools?
- Data rules before an engineer touches production
- An engineer, a team or advice first?
- Hiring AI developers from Progression
- A one page brief for an AI hire
Need a finished product instead? A finished AI product, scoped and delivered as a project rather than engineering capacity, is covered on our AI app development page.
Examples marked “Illustrative example.” are hypothetical; no real company, candidate or job post is shown.
A job post written from a demo
Picture a hypothetical company whose support chatbot demo answered ten friendly questions well. The job post it wrote asked for a prompt engineer who knew one chatbot API, to finish the demo in four weeks.
The feature that customers would actually use needed something else: retrieval over thousands of help articles kept current, a set of real tickets to test answers against, cost tracking and a clean hand off to a person. Prompting was one small layer of it. A post written from the demo would have hired for the easiest part of the work.
- Start from the feature, not the tool.
- Collect the hard questions before writing the post.
- Name the layer where the demo breaks.
- Hire for that layer first.
What AI developers do, layer by layer
AI features are built from eight layers, and engineers tend to be strongest in two or three of them. The table sets each layer beside the work, the evidence worth asking a candidate for, and one question that shows whether they have done it in production.
| Layer | The work | Evidence to ask for | One question to ask |
|---|---|---|---|
| Data | Pipelines that collect, clean and permission the content an AI feature reads | A pipeline they built that refreshes on a schedule | How did you keep stale or restricted records out of answers? |
| Retrieval | Chunking, embeddings, vector and keyword search, ranking | Before and after numbers on answer quality from a retrieval change | When did hybrid search beat vector search for you, and why? |
| Model calls | Prompts, tool calls, structured outputs, model choice | A prompt with its version history and test results | How do you stop a model inventing a field the schema does not have? |
| Trained models | Fine tuning, classic machine learning, feature engineering | A model they trained, with the baseline it beat | When did you decide a smaller trained model beat a large general one? |
| Evaluation | Test sets, graders, regression runs before each change | An evaluation set and the report from it | How do you grade an answer with more than one correct form? |
| Product | The interface, streaming responses, feedback capture | A shipped feature with real users | What did users do that the demo never showed? |
| Operations | Cost per request, latency, logging, fallbacks between models | A cost dashboard or a bill they brought down | What did one conversation cost, and how do you know? |
| Guardrails | Input checks, output filters, permissions on tools and data | A design that limits what an agent may read or change | What can your agent do if a user tells it to ignore its instructions? |
Which layer is your problem in?
Run twenty real, hard examples through what you have. If answers cite the wrong document, the problem is retrieval. If the right document comes back but the answer is wrong, look at model calls. If nobody can say whether a change helped, the missing layer is evaluation, and that is the first hire.
One engineer or several?
A single feature on clean data often needs one engineer strong in model calls and evaluation. A feature that reads several systems, serves customers and runs at volume usually needs a data engineer and a product engineer beside them.
- Wrong document cited: retrieval.
- Right document, wrong answer: model calls.
- No way to tell if a change helped: evaluation.
- Slow or expensive at volume: operations.
- Answers leak data or take unsafe actions: guardrails.
Artificial intelligence developers in the wage data
No single federal occupation is called AI developer. The work spreads across several occupations that O*NET OnLine, run for the Department of Labor, tracks with wage and outlook data. The three below are the closest.
| Occupation on O*NET | Median wage, 2025 | Employment, 2025 | Projected growth, 2025 to 2035 |
|---|---|---|---|
| Computer and information research scientists | $67.45 an hour, $140,300 a year | 38,600 | Much faster than average (7% or higher) |
| Software developers | $65.38 an hour, $135,980 a year | 1,717,800 | Much faster than average (7% or higher) |
| Data scientists | $57.80 an hour, $120,230 a year | 275,600 | Much faster than average (7% or higher) |
Why salaries and bill rates differ
The O*NET figures are employee medians, before benefits, payroll tax, recruiting, equipment and management time. A bill rate covers those costs plus the bench of colleagues behind the engineer, so the two numbers measure different things. Compare total cost for the same months of work, not salary against rate.
Anyone weighing whether to hire an artificial intelligence developer as an employee should add recruiting time to that comparison, since the role spans skills few candidates hold together. For app roles and metro wage tables, see hire app developers.
Gen AI developers and LLM application skills
Most work on generative features sits in the model call, retrieval and evaluation layers. Gen AI developers who ship to real users share a recognizable set of habits.
- Version every prompt and test it before release.
- Ask for structured output and validate it against a schema.
- Ground answers in retrieved documents and show the citation.
- Log every model call with inputs, outputs, cost and latency.
- Design a fallback when a model is slow, down or wrong.
- Keep model choice behind an interface so it can change.
- Treat user input as untrusted, including text inside documents.
What does a retrieval engineer test first?
Whether the right passage comes back at all. Before touching the prompt, a retrieval engineer measures how often the correct document appears in the top results for a set of real questions, then changes chunking, search type or ranking and measures again.
Assistants grounded in your own content are covered on our AI chatbot development page, and connecting models to existing systems on our AI integration services page.
Machine learning engineers for models you train
Some problems suit a trained model better than a general one: forecasting, scoring, classification at high volume, or anything with years of labeled history. That work calls for machine learning engineers who handle features, training, validation and deployment.
Fine tune, prompt or train from scratch?
Prompting a general model is fastest to try. Fine tuning helps when a narrow task needs a consistent format or vocabulary and you have good examples. Training a smaller model from your own data fits high volume, low latency or tightly scoped predictions. An engineer who can explain that tradeoff with numbers from past work has done it.
- Prompting: fastest start, highest cost per call at volume.
- Fine tuning: consistent format on a narrow task with good examples.
- A smaller trained model: cheap and fast once built, needs labeled data.
- Classic machine learning: tabular data and clear targets.
The evaluation set as a hiring test
The single most useful exercise when screening candidates is to ask them to build a small evaluation set for your feature and run two versions of a prompt against it. It shows judgment about hard cases, honesty about failures and a habit of measuring before changing.
How big should a first evaluation set be?
Big enough to include the hard and ambiguous cases, and small enough that a person can read every result. Start with fifty to two hundred real examples and add each new failure found in use.
- Pay for the exercise.
- Use real, anonymized examples.
- Ask for the grading method in writing.
- Read the failures, not just the score.
Interview questions for AI engineers
When you hire AI engineers, ask about something that went wrong. Production experience shows in specific numbers, named failures and the steps taken to catch the next one.
| Interview question | A strong answer covers | A thin answer sounds like |
|---|---|---|
| Walk me through a feature where the model’s answers got worse after a change. How did you notice? | An evaluation set run before release, the failing cases, the fix and a rerun | Users complained, so we rolled back |
| How would you test a support assistant before it meets a customer? | Real tickets including hard ones, expected answers, a grading method, a pass bar agreed in advance | We tried a lot of questions and it looked good |
| What does one conversation cost in your last system? | A number, how it was measured, and what they changed to lower it | The API is cheap |
| When would you not use a large language model? | Cases where rules, search or a small trained model does the job with less cost and less risk | You can always use one |
| How do you keep private data out of prompts and logs? | Redaction before the call, provider data settings checked, log retention set, access limited | The provider handles that |
| An agent can send emails. What stops it sending the wrong one? | Narrow tool permissions, confirmation steps, rate limits, a log a person can review | The prompt tells it to be careful |
| How do you choose between two models? | The same evaluation set, cost and latency compared, and a fallback plan | We use the newest one |
| What did you hand over at the end of your last engagement? | Code, prompts with history, test sets, run books, accounts in the client’s name | The repository |
Questions about cost
Ask what a request or a conversation cost in the candidate’s last system, how they measured it and what they changed. A senior candidate will also know which costs grow with users and which stay flat.
Questions about failure
Ask for the worst answer their system gave a real user and what changed afterward. Look for a test case added, a guardrail tightened or a permission narrowed, not a promise to be more careful.
Know the layer, not the title?Describe the feature and where it breaks today. You get a free proposal naming the engineers, the first month’s work and the price.
Get a free proposalTell us about a project you’d like help with.
Engagement models when you hire AI developers
The right engagement follows the layer and the certainty of the work. A well defined feature can be a fixed scope project; work that will change as you learn suits a monthly engagement.
| Engagement | Fits when | How it is billed | What stays with you |
|---|---|---|---|
| Hourly | A short, well defined task or a review | An hourly rate | The code and notes from each task |
| One dedicated engineer | One layer needs steady work beside your team | Monthly, with a notice period | Code, prompts, test sets and access in your name |
| A small AI team | Several layers move at once, such as data, retrieval and product | Monthly, sized to the roadmap | The same, plus the delivery plan |
| Fixed scope project | The feature and its pass bar can be written down in advance | A project price in the written proposal | The finished feature and its handover |
| Advice first | The use case or the data is not settled yet | A short scoped engagement | A written plan to build from |
Can an AI developer join our existing team?
Yes. Engineers work in your repositories, ticketing and chat, follow your review process and attend your stand ups. Repositories and accounts stay in your name from the first day.
For monthly capacity across other stacks, see hire dedicated developers; for cloud and release roles, hire AWS developers and hire DevOps engineers.
AI developers for hire: market rate bands
Rates vary with the layer, the seniority and the engagement. The bands below run from the typical low to the published high in the US market.
| Engagement | Typical low to published high | Notes |
|---|---|---|
| US agency or contract development rate | $100 to $200 an hour | Bill rates, not employee wages |
| Mid level dedicated developer | $8,000 to $13,000 a month | Monthly engagement |
| Senior dedicated developer | $12,000 to $18,000 a month | Monthly engagement |
| Squad of three or four | $30,000 to $55,000 a month | Several layers at once |
| AI proof of concept | $15,000 to $40,000 | One question tested on your data |
| AI chatbot or LLM assistant build | $30,000 to $150,000 | Retrieval, integrations and evaluation included |
| Production agent, one workflow | $40,000 to $150,000 | Tool permissions and logging included |
Progression’s price is set in the written proposal after scope and is subject to change. A buyer comparing offshore bill rates with these bands should compare total cost for the same evaluation results, since time zone overlap, rework and handover quality move the real figure.
What moves the rate?
Seniority, the number of layers one person covers, the overlap with your hours, and whether the work touches regulated data or production systems.
Who owns code written with AI tools?
Your contract decides who owns the work, and copyright decides what is protectable. In Part 2 of its report on copyright and artificial intelligence, released January 29, 2025, the U.S. Copyright Office concluded that generative AI outputs can be protected only where a human author has determined sufficient expressive elements. For how coding assistants and agents are used and reviewed day to day, see our guide to AI in software development.
The Office’s announcement says that can include a human-authored work perceptible in the output, or creative arrangements or modifications of it, but not the mere provision of prompts, and that using AI to assist creation or including AI generated material in a larger human-generated work does not bar copyrightability. In practice, engineers review, edit and arrange what coding tools suggest, and the contract assigns the resulting work to you in writing.
- Assignment of all work to you, in writing.
- Repositories and model accounts in your name.
- Prompts, test sets and configuration treated as deliverables.
- A record of which tools were used on the codebase.
Data rules before an engineer touches production
AI features read more of your data than most software, and they send some of it to a model provider. Settle the rules before access is granted.
- List the data the feature may read, and the data it may never read.
- Check the model provider’s retention and training settings for your account.
- Redact personal data before it reaches a prompt where the task allows.
- Set retention on prompt and response logs.
- Keep keys in a secrets manager, never in code.
- Give tools and agents the narrowest permissions that work.
Should an AI developer start with production access?
Start in a copy with realistic, scrubbed data. Production access follows once the evaluation set passes and the permissions are written down.
An engineer, a team or advice first?
If the use case is clear and the data is reachable, hire engineers. If several features compete or the data is scattered, a short advisory engagement settles the order first; see AI consulting services. If the goal is an agent that acts across systems, our AI agent development page covers the build.
- Clear feature, reachable data: engineers.
- Several candidate features: advice first.
- Whole product with a launch date: a project team.
- One layer behind in an existing team: one dedicated engineer.
Hiring AI developers from Progression
Progression’s own developers cover data, retrieval, model calls, trained models, evaluation, product, operations and guardrails, and work from New York with on-site sessions on request. Your team interviews each engineer before they start.
- Scoping call on the feature and the layer.
- Written proposal with roles, first month’s work and price.
- Interviews by your team.
- A two week paid trial on your backlog.
- Monthly engagement with a notice period.
- Handover of code, prompts, test sets and accounts.
A one page brief for an AI hire
A short brief turns a vague request into a proposal with named roles. Eight lines are enough.
- The feature, in plain words.
- The layer where it breaks.
- Twenty hard, real examples.
- The systems involved and who grants access.
- Data that may never leave your environment.
- A budget figure.
- Hours of overlap.
- The measure of success.
What this page leaves out: the Progression Method, proprietary and reserved for clients. Reach out to learn more
Getting found in search
AI, AEO and what is changing
Paid media and lead generation
Websites and design
Choosing and working with an agency
Software and app development
Website design by industry and type
Web development, platforms and hosting
Social, content and brand
By industry and by situation
Frequently asked questions
What do AI developers do day to day?
Can I hire AI developers by the month?
What do AI developers for hire cost per hour?
What do artificial intelligence developers earn as employees?
Is an AI developer different from an AI engineer?
What should a gen AI developer know?
Do we need a machine learning engineer or an LLM developer?
Which skills should we test when we hire an AI engineer?
Which interview question shows real LLM experience?
How do we hire AI engineers without an AI team of our own?
Who owns code a developer writes with AI coding tools?
What data access does an AI developer need?
Are the AI developers Progression’s own team?
Can an AI developer for hire work inside our repositories?
What should the first month of an AI engagement produce?
What is an evaluation set?
How many engineers does a first AI project need?
Should we hire engineers or buy an AI product?
Should an AI developer have production access on day one?
What goes in a hire brief for AI work?
Is a fixed price possible for AI work?
Know the layer, not the title?Describe the feature and where it breaks today. You get a free proposal naming the engineers, the first month’s work and the price.
Get a free proposalTell us about a project you’d like help with.
Get a free marketing proposal
Tell us what you are trying to grow and we will come back with a plan, not a pitch deck. Same-day reply on weekdays.
