CompanyBench places contract AI developers with US companies: LLM and RAG engineers, ML engineers, data scientists and MLOps engineers, from the benches of US-based C2C vendors and Indian IT vendors working offshore. Indicative CompanyBench rates for offshore AI engineers from India run from $25/hr for a junior ML engineer to $65/hr for an AI solutions architect. At least 3 vetted, masked profiles arrive within 48 hours.
This page is for US teams that already know roughly what they want built: a retrieval assistant over internal documents, a model that scores leads or claims, a pipeline that classifies support tickets, or the platform to put an existing model into production. Other stacks are on the US technologies page.
AI consultant or AI developer: what you are actually buying
Most results for "AI consultant" are consultancies selling strategy: workshops, a use-case list and a roadmap. That is useful when you don't know where AI fits. Once you have a use case and data, the bottleneck is engineering, and a contract developer who builds inside your stack is the faster and cheaper hire.
| AI consultancy | Contract AI developers (CompanyBench) | |
|---|---|---|
| You buy | Advice, a roadmap, sometimes a proof of concept | Engineers who write and ship code in your repositories |
| Output | Slides, a prioritized use-case list, a prototype | Working pipelines, services, evaluations and deployments |
| Who directs the work | The consultancy's lead | Your engineering or product lead |
| Best for | No clear use case yet, executive alignment, vendor selection | A defined use case, a backlog, a prototype that has to reach production |
| Commercials | Fixed fee or day rate per engagement | Hourly or monthly rate for approved timesheets |
A common pattern: a consultant or your own lead picks the use case, then contract engineers build it. If you are hiring an "AI consultant" to write production code, write the requirement as an engineering role.
AI roles we fill
"AI developer" is four or five different jobs. Hiring the wrong one is the most expensive mistake on this page.
| Role | What they build | Core tools | Hire when |
|---|---|---|---|
| LLM and RAG engineer | Assistants and agents over your documents and systems: retrieval, chunking, prompts, tool calls, guardrails, evaluations | Python or TypeScript, LLM APIs, vector databases, LangChain or LlamaIndex or plain code | You want a language model to answer from, or act on, your own data |
| ML engineer | Trained models in production: features, training pipelines, serving, monitoring | Python, PyTorch or scikit-learn, SQL, cloud ML services | You have labeled data and a prediction problem: scoring, forecasting, ranking, classification |
| Data scientist | Analysis, experiments and first models; the question before the system | Python or R, SQL, notebooks, statistics | You need to know whether the data can answer the question at all |
| MLOps engineer | The platform: CI/CD for models, feature stores, model registry, monitoring, cost control | Kubernetes, MLflow or similar, cloud ML platforms, Terraform | Models exist but releases are manual, slow or unmonitored |
| AI solutions architect | System design across data, models, security and cost; vendor and model choices | Broad: cloud, data platforms, LLM providers | Several teams or use cases share the same AI stack |
For a first LLM feature, one LLM engineer with solid backend skills beats a data scientist, because most of the work is software: APIs, retrieval, logging and tests. For a forecasting or scoring problem, it is the other way round.
AI developer contract rates
These are indicative CompanyBench rates for AI engineers working offshore from India, by role and experience, all in, with no placement fee. The rate for a specific engineer is agreed upfront and shown on the shortlist; US-based C2C engineers are quoted per requirement.
| Role | Experience | Indicative rate |
|---|---|---|
| Junior AI/ML engineer | 2–4 years | $25/hr |
| Mid-level AI/ML engineer | 4–7 years | $35/hr |
| Senior AI/ML engineer | 7–10 years | $48/hr |
| LLM and generative AI developer, mid to senior | 4–8 years | $42/hr |
| MLOps engineer | 3–8 years | $40/hr |
| AI solutions architect or technical lead | 10+ years | $65/hr |
Budget separately for model API usage, GPUs and vector database hosting. Those bills are yours, run on your accounts, and a good engineer will show you the cost per request before anything goes live.
How to test an AI candidate on real work
AI CVs are easy to inflate: a tutorial notebook reads the same as a production system on paper. Test on a slice of your own problem, with data you are allowed to share.
- LLM and RAG engineers. Give a small set of sanitized documents and a handful of real questions with known answers. Ask how they would chunk and index the documents, how they would measure whether retrieval found the right passage, and what they would log. Strong candidates talk about an evaluation set before they talk about prompts.
- Ask about failure. Have them describe a model or assistant that went wrong in production and how they found out. People who have shipped can name the monitoring that caught it.
- Prompt injection and excessive agency. Ask how they stop a document or a user from overriding the system's instructions, and how they limit what an agent's tools are allowed to do. OWASP's Top 10 for LLM Applications lists Prompt Injection as LLM01:2025 and Excessive Agency as LLM06:2025; anyone building agents with tool access should know the list.
- ML engineers. Give a feature table and a target. Ask how they would split the data to avoid leakage, which metric matches the business cost of an error, and how they would know the model has drifted after launch.
- Data scientists. Ask them to explain a past result to a non-technical manager, including what the data could not tell them.
- MLOps engineers. Ask them to sketch how a retrained model gets from a notebook to production with a rollback, and who approves the release.
Keep the exercise short. You learn more from a candidate talking through your problem for an hour than from a weekend take-home task.
Data access and security in the contract
AI work touches data that ordinary feature work does not. Settle these before the start date and write them into the requirement, so the shortlist only includes people who can work under them:
- Which data, at which stage. Start on synthetic, masked or sampled data. Move to production data only for named tasks, in your environment, with access logged.
- Where data may go. List the model providers and endpoints that are approved. Forbid pasting company data into personal accounts or consumer chat tools.
- Keys and accounts. API keys, cloud accounts and vector databases belong to you and live in your secrets manager. The contractor never pays for or owns a model account used on your work.
- Regulated data. If the data includes health, financial or personal records, say so up front. Your own policies or contracts restrict offshore access to some data sets; it is better to scope the work around that than to discover it in week three.
- Work product. Code, prompts, evaluation sets, fine-tuned weights and training data prepared for you are work product. Ownership, confidentiality and IP terms are set out in your contract.
- A risk framework. If your company needs a reference for AI governance, NIST's AI Risk Management Framework (AI RMF 1.0, January 26, 2023) is voluntary, and NIST AI 600-1, its Generative AI Profile, followed on July 26, 2024.
How hiring an AI engineer works
- Send the requirement: the use case, the role, your stack and model providers, the data involved and where it is allowed to go, overlap hours, remote or on site, and length.
- We confirm it and agree a rate range, then route it to IT vendors whose benches include engineers who have shipped that kind of AI system.
- Within 48 hours at least 3 masked profiles reach you (initials, role, years, certifications, location, rate), each noting the models, retrieval stacks or pipelines the engineer has put into production.
- CompanyBench books each interview call; contact details are not shared in either direction.
- You choose an engineer. You sign with CompanyBench, which signs a back-to-back agreement with the engineer's employer.
- Work starts on your accounts and data, under the access rules you set. Timesheets are approved monthly, one CompanyBench invoice follows, and CompanyBench settles with the vendor.
How we vet AI engineers
- A government ID matched to the person on the profile.
- An assessment in the role you asked for: retrieval quality and evaluation sets for LLM work, leakage and metrics for ML, release pipelines for MLOps.
- Proxy and deepfake interview screening, so the engineer you approve is the one with access to your data.
- Cloud and AI vendor certifications verified when the profile lists them.
- For AI engineers employed in India, a UAN check that rules out dual employment.
- A background check completed before any data access is granted.
Example: a support assistant over internal documents
Example: a US insurance software company wants an assistant that answers customer support agents from its policy manuals and past tickets. The requirement reads: one LLM and RAG engineer with Python backend skills, the company's approved model provider and PostgreSQL with pgvector, masked ticket data only for the first month, remote from India with four hours of overlap with Eastern time, five months. First deliverable: an evaluation set of real questions with known answers, before any prompt work. The shortlist shows three masked profiles with rates and the retrieval systems each has shipped. The product lead gives each candidate a handful of sanitized documents and questions, asks how they would measure retrieval, picks one, and the engineer starts with keys in the company's secrets manager, never in a personal account.
Mistakes US teams make hiring AI contractors
- Hiring a data scientist to build a product feature. An LLM assistant is mostly software engineering. Hire for backend and retrieval skills, and add data science when you need experiments.
- No evaluation set. Without a fixed set of questions and expected answers, nobody can tell whether a prompt change made things better or worse. Make building one part of the first two weeks.
- Judging by a demo. A demo on five hand-picked questions proves little. Ask for results on the evaluation set, including the failures.
- Ignoring cost per request. A design that calls a large model several times per query gets expensive at volume. Ask for a cost estimate per request before launch.
- Production data on day one. Give contractors masked or sample data first, and widen access as the work needs it.
- Nobody owns it after launch. Models drift and providers change models. Agree who monitors quality and cost once the contractor rolls off, or keep an MLOps engineer on.
How the C2C arrangement works for an AI hire
The AI engineer stays on their IT vendor's payroll; the vendor contracts with CompanyBench and CompanyBench contracts with you, so there is one counterparty for data and IP terms. For how C2C engagements run in the US, start at the US hiring overview; C2C, W2 and 1099 compared explains why an AI contractor on C2C is not your employee. Most AI systems also need Python developers and data engineers around them. Ready? Send a requirement.
Frequently asked questions
How much does it cost to hire an AI contractor?
Indicative CompanyBench rates for offshore AI engineers from India run from $25/hr for a junior ML engineer to $65/hr for an AI solutions architect, with LLM developers around $42/hr and MLOps engineers around $40/hr. US-based C2C engineers are quoted per requirement. Model API and GPU costs are separate, on your own accounts.
What is the difference between an AI consultant and an AI developer?
An AI consultant advises: which use cases are worth doing, which vendors to choose, how to govern it. An AI developer builds: retrieval pipelines, models, services and deployments in your code. If you already know the use case and have the data, you need a developer. CompanyBench supplies developers and engineers, not strategy consulting.
Do I need an ML engineer or an LLM engineer?
Use an LLM engineer when a language model should answer from or act on your documents and systems. Use an ML engineer when you have labeled data and a prediction problem such as scoring, forecasting or ranking. Some engineers do both, but test for the one your use case needs.
Can an offshore AI developer work with our sensitive data?
It depends on your policies and the data. Start with masked or synthetic data, keep production data in your environment with logged access, and list approved model providers in the contract. If some data cannot leave the US under your own obligations, tell us and we scope the role or send US-based engineers.
Who owns the prompts, models and evaluation sets?
Ownership of work product is set out in your contract with CompanyBench, so read it before the start. List what counts as work product in the requirement: code, prompts, evaluation sets, fine-tuned weights and prepared training data. Keep model accounts and API keys in your company's name, so access ends when the engagement does.
How do I test an AI developer before hiring?
Use a slice of your own problem. Give an LLM candidate a few sanitized documents and questions with known answers and ask how they would measure retrieval. Give an ML candidate a feature table and ask how they would avoid leakage. Pay for any exercise longer than an interview.
How soon does an AI engineer shortlist arrive?
At least 3 masked AI profiles with rates arrive within 48 hours of confirming the requirement. Data access approvals are the slowest step in most AI engagements, so start them while you interview and have masked or synthetic data ready for the first week.
Tell us what you need
A vetted shortlist within 48 hours of confirming the requirement. No fees until someone starts.
Sources
- OWASP Top 10 for LLM Applications 2025
- NIST AI Risk Management Framework (AI RMF 1.0 and the Generative AI Profile, NIST AI 600-1)
Last reviewed November 22, 2026.
More for United States
- CompanyBench US
- C2C requirements for vendors and bench sales teams
- C2C vs W2 vs 1099: what corp-to-corp means for buyers and vendors
- Hire .NET developers on contract: C#, ASP.NET Core and .NET upgrades
- Hire AWS consultants and cloud engineers on contract
- Hire Azure consultants and cloud engineers on contract
- Hire contract developers and IT consultants by technology
- Hire data engineer consultants on contract

