How to hire AI engineers in 2026: in-house, freelance, agency, studio

There are five ways to hire AI engineers in 2026. You can employ them, contract freelancers through a marketplace, rent engineers from a managed talent or nearshore firm, commission a development agency, or work with a product studio that brings a fractional CTO. An in-house AI engineer in the US costs about $223,000 a year once employer costs are added to the midpoint of Robert Half’s 2027 salary range. Technical roles took a median of 76 days to fill in Ashby’s hiring data, while marketplaces say they can match a freelancer in two or three days. The options differ most in who owns the architecture and the evals once the first version ships.

Demand for the role is high. LinkedIn put AI engineers first on its 2026 list of the fastest-growing jobs in the US, and 92% of the technology leaders in Robert Half’s survey were offering higher pay for relevant AI skills.¹² The role is also young, so a title says little about what someone has run in production. In LinkedIn’s data, people who moved into AI engineering had a median of 3.7 years of prior experience, most often as software engineers, data scientists or full-stack engineers.

9io wrote this post, and as a product studio with a fractional CTO we are one of the five options. Figures for the other four come from public sources and each company’s own pages as of 9 October 2026, and each company’s claims are labelled as its own. The decision rule at the end includes the cases where a studio is the wrong choice.

Five ways to hire AI engineers, compared

This table compares the five as of 9 October 2026, and the sections after it give the sources.

Hourly figures make the routes easier to compare. This table converts each one to a cost per hour, as of 9 October 2026.

Rates don’t show how many hours each route needs for the same result, or how much time someone on your side spends directing and reviewing the work.

AI engineer salary in 2026, and what an in-house hire costs

The US Bureau of Labor Statistics has no occupation for AI engineers. The nearest are software developers, with a median wage of $135,980 in May 2025, and computer and information research scientists, at $140,300.¹¹¹²

Robert Half’s 2027 Salary Guide, released on 1 October 2026, lists AI engineer as a role of its own. Its ranges are US national averages of base pay for new hires, built from the placements Robert Half makes and checked against job-posting data.²¹³

Levels.fyi adds self-reported pay, including stock and bonus. On 9 October 2026, US software engineers tagged ML or AI had a median total compensation of $250,000 across 5,941 submissions, with a median base of $185,400.¹⁴ All US software engineers had medians of $196,925 in total and $160,000 in base, so the ML and AI tag carried 27% more total pay and 16% more base.¹⁵

Salary is only part of the cost. In June 2026, private employers spent $73.69 per hour worked on compensation for professional workers, of which $50.98 was wages and $6.78 was paid leave.¹⁶ An annual salary already includes paid leave, so the multiplier for a salary is $73.69 divided by $57.76, or 1.28. At Robert Half’s $175,000 midpoint, an AI engineer costs about $223,000 a year, or $18,600 a month, which is about $122 per hour actually worked. That leaves out equity, equipment, management time and recruiting. Arc, for example, charges 20% of the annual salary for a full-time hire, which would be $35,000 at the midpoint.¹⁷

Time is the other cost. Ashby’s 2026 recruiter productivity report draws on anonymised data from its recruiting platform, covering 109 million applications and 247,000 jobs from January 2021 to March 2026. The median technical role took 76 days from opening to first hire, against 56 days for business roles, and 40 days from application to accepted offer.¹⁸ Notice periods and the weeks a new engineer spends learning your codebase come on top.

Your hire owns the architecture and the evals, which is the strongest argument for hiring. If they are your first AI engineer, nobody else on the team can review their eval set or question a model choice, and an interview panel without AI experience struggles to tell production work from demos.

Hiring pays once the work is steady. By our arithmetic, at Go Fractional’s $156 average for a fractional lead AI engineer, $223,000 buys about 1,430 hours a year, or 27.5 hours a week, every week. Below that, renting the time costs less in cash. At the $125 median rate, the line moves to about 34 hours a week.

How to hire freelance AI developers on Upwork, Toptal and Arc

Marketplaces are the quickest way to put someone on an AI task. They fit a prototype built to test demand, or a bounded task that your own technical lead directs. As of 9 October 2026:

• Upwork says hiring an AI engineer on its marketplace generally costs $35 to $60 an hour, and gives $3,000 to $12,000 as a typical project price for a generative AI application with RAG or fine-tuning.⁴ Its Basic plan adds a 5% fee to each payment and a one-time contract fee of $0.99 to $14.99. Business Plus charges 10% and includes what Upwork calls Expert-Vetted talent, the “top 1%” of its freelancers.⁵
• Toptal says fewer than 3% of the more than 200,000 people who apply each year are accepted, and that clients typically hire an AI developer in about 48 hours. Each engagement starts with a trial of up to two weeks, which isn’t billed if the client isn’t satisfied.¹⁹ Toptal doesn’t publish talent rates, and charges a flat $79 monthly subscription once you proceed with matching.²⁰ We compare Toptal with Arc, Lemon.io, Gun.io and the managed options in Toptal alternatives in 2026.
• Arc says AI developers hired through it typically charge $60 to $100+ an hour, that it matches a freelance developer in about 72 hours, and that full-time hires take about 14 days.⁷ Its pricing page lists freelance hires at $15 to $110+ an hour.¹⁷

A freelancer makes the design decisions inside each task. Someone on your side has to own the architecture across tasks, and evals exist only if the contract asks for them. Upwork’s own hiring guide tells clients to define how model performance will be measured and accepted before the project ends, and to agree who owns the datasets, trained models, prompts and source code.⁴

The AI-specific risks follow from contracts sized to a task:

• A $3,000 to $12,000 project price buys a working application. Check whether it also buys an eval set, monitoring and spend limits, because those keep it working after launch.
• Production data and API keys end up with one person. Keep data in your accounts, give scoped access, and issue keys in your company’s name that you can revoke.
• When the contract ends, the reasons behind each prompt and threshold leave with the freelancer unless they are written down in the repository.

Managed talent and nearshore firms such as Andela and BairesDev

Managed talent firms place their engineers on your team or run a team for you, and nearshore firms do the same from time zones close to yours. They fit a company that already has a technical lead and needs several engineers for months. As of 9 October 2026:

• Andela says it has 17,000 certified “AI-native” engineers and has trained more than 200,000 people, and offers to add AI engineers to your team or deploy a fully managed one.²¹ Its pricing link leads to a form for booking a discovery call.
• BairesDev says on its Clutch profile that it has 4,000+ senior engineers across the Americas working in their clients’ time zones, and that it receives more than 2 million applications a year and hires fewer than 1%. The same profile lists $50 to $99 an hour and a minimum project of $50,000.⁶ It sells staff augmentation, dedicated teams led by its own technical project managers, and full outsourcing.²² It says it typically stands up a team in two weeks, and that its engineers typically cost 30% to 50% less than the fully loaded cost of a local hire.²³

Ownership depends on the model you buy. Under staff augmentation, the engineers work to your lead, who owns the architecture and the evals. On a dedicated team, the firm’s project manager runs delivery against your priorities, and under full outsourcing BairesDev says it takes “full ownership of the execution”. The further along that line you go, the more the contract needs to name the eval set, the prompts and the runbooks as deliverables kept in your repository.

The AI-specific risk is continuity. Ask what happens to the eval set, and to the knowledge behind each prompt, when an engineer moves to another account. Ask how much of the working day overlaps with yours, because a model provider’s outage needs someone awake who knows the fallback. Production data should stay in your accounts, with access granted to named engineers and removed when they leave.

Outsourcing AI development to an agency

An AI development company, or agency, builds to a contract, usually a scope with milestones and a fixed or estimated price. It fits a bounded project outside your core product, where both sides can agree an acceptance test before work starts. Clutch’s AI pricing guide, updated 21 September 2026, draws on verified client reviews. It says most AI development firms on Clutch charge $25 to $49 an hour and most reviewed projects cost $10,000 to $49,999. The average project cost $120,595, with a typical timeline of 10 months.³ GoodFirms puts professional agencies in general higher, at $75 to $250 an hour.⁸

The agency designs the architecture within the scope. Evals, monitoring and spend controls are part of the build only when the contract names them, and at handover you take over a system someone else designed.

Two risks are particular to AI work. The first is the gap between a demo and a product that holds up under real use. In June 2025, Gartner predicted that over 40% of agentic AI projects would be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. It also estimated that only about 130 of the thousands of vendors selling agentic AI are real, and called the rebranding of older products “agent washing”.²⁴ Ask any agency to show a system that has run in production for months, and what its evals caught.

The second is model churn after handover. Anthropic gives at least 60 days’ notice before it retires a publicly released model, and on 30 September 2026 it gave notice that Claude Sonnet 4.5 retires on 30 November.²⁵ OpenAI’s hosted Evals platform becomes read-only on 31 October 2026 and shuts down on 30 November.²⁶ A system handed over with its evals in a hosted tool, or pinned to a retiring model, needs work within weeks. The contract should say who reruns the evals when a model changes, and who pays for that time.

A product studio with a fractional CTO

A product studio builds a product and keeps it running with a small team. The version compared here includes a fractional CTO, who owns the architecture, the vendor choices and the evals for part of each week. 9io is one. We are a small team of senior engineers with a fractional CTO, a designer, a product manager and QA, and these are the terms we work on.

• The proposal names the people who would do the work, their hours and each person’s rate. There is no management fee, and we bill only for the hours we work, invoiced twice a month for work already done, with 10% of each invoice held until you accept the milestone.
• Work usually starts within a couple of days of agreeing a proposal, with a working environment and a login near the start and a short written update every Friday.
• Code, history and infrastructure definitions live in your repository from the first commit, and you own everything as you pay.
• AI features are tested on every change, and pipelines can be traced. Under a 90-day warranty, defective delivered work gets fixed, and you can scale up, scale down or stop at a month boundary with 30 days’ notice.

This post doesn’t quote our rates. As a market reference, Go Fractional’s benchmark on 9 October 2026 put fractional CTOs at an average of $209 an hour and a median of $200, for a typical 15 hours a week. Across the middle half of rates, that comes to $9,000 to $15,000 a month.¹⁰ Our post on what a fractional CTO costs compares that with a full-time hire, and our AI engineering service describes the work we take on.

Per hour, fractional CTOs and senior US engineers sit at the upper end of the hourly table above. A small team has a ceiling on how much it can run in parallel. The fractional CTO is part-time, and until you hire your own technical lead you depend on the studio’s people for decisions. A studio suits a company whose product is AI and that has no CTO yet, from the first build until it hires one. At that point the fractional CTO can write the role, interview for it and hand over.

What makes a good AI or LLM engineer in 2026

Most of the work in an AI product happens around the model call. LinkedIn lists LangChain, RAG and PyTorch as the most common skills among AI engineers,¹ and a list of frameworks says little about whether someone can keep a feature working. The five areas below say more, and each links to a write-up of that work on products we build and run.

Evals. A good AI engineer builds a set of real inputs with known good outputs before tuning anything, and runs it on every prompt, model and retrieval change. They also compare what a pipeline produced with what it planned to produce. One of our content pipelines reported 0 dropped while shipping 26 of 51 planned diagrams, because work given up after throttling was never counted. The eval set belongs in your repository in a form that any provider’s models can run against.

Latency and cost engineering. They know where the time goes before the first token, and what one request costs. An in-app assistant we run went from 20 to 30 seconds to about 3 seconds to first word once its context went into the prompt and tools were forced only for real actions. On cost, a backfill job made 419,057 vision calls it didn’t need, about $1,100, and rate limits don’t stop a job that runs steadily below them. A spend ledger per job and daily caps catch that kind of bug early.

Retrieval. They measure retrieval separately from generation, using labelled queries and the documents that should come back, and they apply the user’s permissions before ranking. They also know when keyword search or a database query is enough.

MCP and tools. Tools are where an assistant acts on other systems. The Model Context Protocol is the common way to expose them, and its current specification is dated 28 July 2026.²⁷ When we built one MCP server that works in Claude, ChatGPT, Cursor and VS Code, almost everything that broke was in OAuth and protocol-version handling, which a demo against a single client never exercises.

Guardrails. The OWASP Top 10 for LLM applications (2025) starts with prompt injection, and also covers excessive agency and unbounded consumption.²⁸ In practice, that means tools with the least privilege they need, a person approving actions that write or spend, retrieved text treated as untrusted input, and spend caps on every job. The MCP specification itself says tool descriptions should be treated as untrusted unless they come from a trusted server.²⁷

Interview questions that separate AI engineering from demo-building

Ask about systems that ran in production, and listen for numbers. Each question lists what a strong answer contains.

1. Show me an eval set you built. Where did the cases come from, and what did it catch? A count, a source such as real user inputs labelled by someone who knows the domain, a pass threshold, and a regression it stopped.
2. The model you depend on retires in 60 days. How do you switch? Run the eval set on two or three candidates, compare quality, cost per task and latency, send a share of traffic to the new model, and keep a fallback.
3. Time to first token is 12 seconds. How do you find out why? Timing each step before changing anything, from retrieval and tool calls to reasoning tokens and the network. “Switch to a faster model” skips that step.
4. What does one request cost, and how would you stop a job spending $1,000 overnight? Tokens in and out times the price, a ledger per job and model, caps that fail small, and jobs that skip work they have already done.
5. How would you tell whether retrieval or the prompt is the problem? Labelled queries with expected documents, recall measured before any prompt tuning, and permission filters applied before ranking.
6. A document in the index tells the assistant to email your customer list. What stops it? Least-privilege tools, a person approving writes, retrieved text and tool output handled as untrusted, and a log of every action.
7. When would you not use an LLM here? Cases where a rule, a database query or a small classifier is cheaper and more predictable. We ask ourselves the same question, and if a form or a database query solves the problem, we use that.
8. What have you shipped that is still running, and what broke? A specific incident, the numbers around it and what changed afterwards.

The same questions work on vendors. Ask an agency or a studio to answer them about a system it built, and to show you the eval set.

How these figures were gathered

• Salaries are US base pay for new hires from Robert Half and May 2025 wages from BLS. Levels.fyi figures are self-reported and include stock and bonus, so they run above base-pay surveys.
• Employer costs use BLS’s June 2026 figures for professional workers in private industry. The multiplier is total compensation divided by wages plus paid leave, and hourly cost is annual cost divided by the hours worked once leave is taken out.
• Time to hire is Ashby’s median from its own platform, and companies on other recruiting tools may differ.
• Vendor figures are each company’s own claims, read on 9 October 2026, which we haven’t verified. BairesDev’s live site showed us a bot check, so we read its Clutch profile and the Internet Archive’s copies of its pages from late September.
• Not covered: pay outside the US beyond Clutch’s rate bands, equity, quality differences between people or firms, and hiring data of our own. We didn’t survey anyone or run a hiring test.

Which option fits your stage

Decide who will own the architecture and the evals first, then buy the hours around that person.

• Before you have a technical owner, avoid hiring a junior AI engineer to work alone, and avoid any fixed-price build without an eval set in the scope. Both leave nobody able to tell whether the next change made things worse.
• When you hire, plan for about 11 weeks to fill the role, and keep outside help until the new hire has taken over the eval set and the runbooks.
• With sensitive or regulated data, any of the five can work if the data stays in your accounts, API keys are in your company’s name, and the people with access are named in the contract.
• Whichever you choose, put code, prompts, eval sets and infrastructure definitions in your repository from the first week, and agree in writing who reruns the evals when a model changes.

Frequently asked questions