| Company and role | Base salary | Experience | Where |
|---|---|---|---|
| Gusto Enterprise Application AI Engineer | $185k to $205k (Denver); $225k to $250k (SF) | 8+ years | SF hybrid |
| Asana Senior Software Engineer, AI Retrieval | $202k to $223k | 6+ years | NYC |
| Asana Software Engineer, AI Developer Experience | $171k to $190k | 3+ years | NYC |
| Brex AI Engineer, Product | $171k to $240k | Not stated | SF hybrid |
| Mercury Senior Software Engineer, AI | $166.6k to $218.7k | 5+ years | SF / NYC |
| Lyft Senior AI Software Engineer, Risk | $148k to $185k | 6+ years | SF / Seattle |
| Vercel Software Engineer, AI SDK | $196k to $294k | 5+ years | SF hybrid |
| Coinbase Senior SWE, AI Platform | $186k to $218.9k | 5+ years | Remote US |
| Harvey Senior SWE, AI Platform | $185k to $325k + equity | 5+ years | SF |
| Harvey Sr. AI Enablement Engineer | $133.5k to $200.3k + equity | 5+ years | Remote |
| Ramp Applied AI Engineer | $204.4k to $352k + equity | Not stated | NYC hybrid |
| OpenAI Applied AI Engineer, Enterprise | $197k to $278k + equity | Not stated | SF hybrid |
Read from the companies’ own job boards on 14 August 2026. US base salary; equity noted where the posting notes it. Harvey appears twice on purpose.
Post a job called “AI engineer” this year and the inbound is enormous. What’s hard is not getting applicants. It’s that after a dozen calls, most founders still can’t tell which of these people has ever put a language model in front of a paying customer.
That’s the actual problem, and it isn’t a shortage. It’s a title that means at least four different jobs, applied to a market where every engineer with a weekend project can honestly claim it.
The table above is what the job pays right now at companies that publish their bands. Below is what those companies actually ask for, which turns out to be more useful than the title.
There’s a lazy way to read a job called “AI engineer”, which is to assume it wants someone who builds models. That’s mostly wrong, and the postings say so if you read past the headline.
We pulled the full text of 83 current AI engineering postings from 26 companies that build software but don’t build models: Stripe, Figma, Coinbase, Gusto, Asana, Brex, Vercel, Lyft, Datadog and seventeen more. Every posting, the selection rule, and the exact phrase counted for each requirement are published alongside this article, so the numbers below can be argued with. These are the employers most readers of this article resemble. What they ask for is lopsided in an instructive way.
Ten percent. The thing most people picture when they hear “AI engineer” is a requirement in one posting out of ten.
Brex puts it in writing. Its “AI Engineer, Product” listing is for someone who will build an agent that audits customer spend. It puts “experience building products on top of LLMs or agentic systems” under Bonus points. Not requirements. The requirements are the ones you’d write for a senior product engineer.
A company hiring an AI engineer to build an agent considers agent experience a bonus. That is the whole role in one line.
The AI engineer most startups need is a strong product engineer who has shipped one thing on top of a model and can tell you what broke.
The bands in the table run from under $150,000 to over $350,000, and that spread isn’t a market disagreement. It’s several jobs sharing a name.
Across 158 AI engineering titles at ten AI-native companies, they sort into four families. Product or applied engineers build the feature. Platform engineers build the routing, serving and infrastructure it runs on. Enablement, deployment and support engineers help customers get value from what was built. And forward deployed engineers do the building inside a customer’s own systems, which we wrote about separately. Every one of them is an AI engineer on LinkedIn.
The money splits along one line, and it isn’t product versus platform. Of the 51 distinct titles that publish a US band, product roles typically run $204,000 to $385,000 and platform roles $220,000 to $390,000. Enablement and support roles run $180,000 to $260,000. Building pays the same whether it’s the feature or the infrastructure under it. Helping people use what was built pays a third less at the top.
Harvey, the legal AI company, has both a Senior Software Engineer, AI Platform at $185,000 to $325,000 and a Senior AI Enablement Engineer at $133,500 to $200,300 open at the same time. Same word, same company, same seniority.
One builds the model routing and agent architecture the product runs on. The other helps customers get value from it. A hundred thousand dollars lives inside the word “AI”.
For a startup this matters in a specific way. If your posting says “AI engineer” and nothing else, you’ll get applicants from every one of those buckets, and you’ll pay the one you accidentally hire the rate of the one you were picturing.
Say what the person will own. “You will build the retrieval and eval pipeline behind our support assistant” filters better than any list of frameworks.
And keep the premium in proportion. Stack Overflow’s 2025 survey puts the US median total compensation for AI/ML engineers at $189,500 against $175,000 for back-end developers, across 5,239 US respondents. That’s about eight percent.
Real, but nowhere near the doubling the panic around this title implies.
Here’s the split that matters, and it isn’t fine-tuning versus prompting.
The engineers worth hiring have been through the part where the demo works and the product doesn’t. They talk, unprompted, about evals: how they measured whether the thing was getting better, what the test set looked like, what they did when a model update broke a case that used to pass. They know what an LLM call costs at scale and what it does to a page’s latency budget. They have a story about a hallucination that reached a user.
The postings ask for exactly this. Figma’s London role wants “experience curating and developing eval sets and LLMs as judges.” Gusto lists “robust evaluation frameworks (evals), guardrails for code safety, and strategies for optimizing model latency and token spend.” Stripe screens for people who’ve built agents “beyond basic prompt engineering.”
The ones to be careful with have a portfolio of things that work in a screenshot. A wrapper around a chat completion. A retrieval demo over one PDF. They can talk about agents fluently and haven’t run one for a month. Eighteen months ago almost nobody had production experience, so this describes a lot of good engineers. You should still know which one you’re paying for.
Forget the whiteboard algorithm round. It tests something you already tested by reading the CV.
Ask for a war story, then pull the thread. “Tell me about an LLM feature you shipped that behaved differently in production than in testing.” Then: how did you find out? What did you measure? What did you change? What did it cost?
A strong answer is specific and slightly embarrassing. “Users started pasting invoices with two currencies and the extractor picked one at random. We found out from a support ticket, not from monitoring, which was the real problem. We built a set of 60 hard cases, made it part of the deploy check, and it caught the next model upgrade breaking date parsing.” A weak answer is a description of a system with no incident in it. If by the third question you’re hearing “generally” and “typically”, the person hasn’t been paged for this yet.
Give them a real, small piece of your problem, and pay for it. Ten hours, your data, your constraint. If you already have something in production, hand over its mistakes: “here are 200 support tickets and where our current classifier gets them wrong.” If you’re starting from zero, hand over the raw material: “here are 200 tickets; tell me which ones a model could route today, which it can’t, and how you’d know.”
What you’re grading isn’t the accuracy number. Read the write-up for three things. Did they build a way to measure before they built the thing? Did they say what they’d get wrong and how they’d notice? Did they mention that the tickets are messy? Someone who returns a working demo and no test set has told you what their first six months will look like. We wrote up how to run this in our guide to the paid trial project, and it applies here without changes.
Ask about the money. “This runs on a frontier model at 40,000 requests a day. What’s the bill, and what would you do about it?” You don’t need to know the answer yourself. A good one starts by asking how long the prompts are and how much of the input repeats, then reaches for the levers in order: cache what repeats, route the easy requests to a cheaper model, cut the prompt, and only then talk about fine-tuning. A weak one either names a number instantly or goes straight to “we’d fine-tune a smaller model,” which is the expensive option offered first. You’re testing whether they’ve ever had to care.
If you’re adding AI to an existing product and you have one hire, don’t hire a researcher. Hire the product engineer described above and give them a budget for API calls. If the feature works, the second hire can be someone who goes deeper on models. Hire the researcher first and you’ll get a very good notebook and no product.
If you’re in San Francisco or New York and hiring in-house, the table at the top is your market, and ten of the twelve postings in it want you in an office. If you’re not, the same three questions work over a video call, and a senior contractor with production LLM work behind them bills in the $50 to $80 an hour range from Latin America or Eastern Europe. Whether that’s a bargain or a warning sign depends entirely on whether you asked the questions.
If you’d rather not build this interview loop yourself, it’s roughly what we run before a client sees a profile: the war story, the paid trial on real work, the question about the bill. Tell us what you’re building and the people you meet will already have answered all three.
Privacy is important to us, so you have the option of disabling certain types of storage that may not be necessary for the basic functioning of the website. Blocking categories may impact your experience on the website. More information
