How to Start a Data Annotation or AI Training Data Company for US Clients from Abroad (2026), Without Visiting: US LLC or C-Corp, EIN Without an SSN, Bank Account, Client Security Demands, HIPAA, Biometric Data Laws and Paying Annotators

Every AI model is trained and tested on data that people have labelled. Someone drew the boxes around pedestrians in a self-driving dataset, marked the tumours in a set of scans, ranked two chatbot answers against each other, or checked whether a model's code actually runs. Most of that work is done by teams outside the USA, for buyers inside it.

US AI companies, from large labs to two-person start-ups, buy this work from specialist vendors. Many of those vendors are small companies in India, the Philippines, Poland, Colombia or Vietnam, with a US company at the front so American buyers can sign contracts, pass vendor onboarding and pay by ACH.

This guide explains how to set that up without visiting the USA, and the rules and security demands that matter when you handle US clients' training data.

In short: A non-resident can form a US LLC or C-Corporation without visiting, get an EIN without an SSN or ITIN, and open a US business bank account remotely. With your annotators working abroad, a foreign-owned single-member LLC usually owes no US federal income tax on the fees, but files Form 5472 every year. Expect US buyers to ask for a security questionnaire, often SOC 2 or ISO 27001, and strict NDAs. Medical data brings HIPAA and a business associate agreement. Faces, fingerprints and voices can be biometric data under state laws such as Illinois BIPA. Avoid any project involving defence or export-controlled data unless the client has cleared it.

What AI data companies sell

The market has split in two. Volume work (simple boxes and tags) is price-driven and increasingly automated. Expert and judgement work (RLHF by engineers, doctors or lawyers, model evaluation, multilingual data) pays much more and rewards a specialised team. A small company abroad usually does best picking a niche, such as a language, a medical speciality or a type of sensor data, rather than competing on price alone.

Can you start an AI data company in the USA without visiting?

Yes. All of the set-up is done remotely:

• Form the LLC or corporation online, with a registered agent in the state
• Get the EIN without an SSN or ITIN, using Form SS-4, which the IRS processes manually for foreign owners
• Open a US business bank account without visiting, at a bank or fintech that accepts foreign-owned companies
• Sign client contracts and NDAs electronically
• Run the work on cloud annotation tools or the client's own platform

Expect four to eight weeks for the company, EIN and bank account together. See opening a US company without visiting, the EIN for non-residents guide and the US business bank account guide. Banks ask what you do; "data annotation services for AI companies" with a website, team page and sample work is an easy business to explain. If one turns you down, see why banks reject non-resident applications.

US LLC or C-Corp?

US LLC formation for an annotation company

A single-member LLC is the usual start. The IRS treats it as disregarded, and where all the annotation is done by your team outside the USA, with no US office or US staff, the fees are generally foreign-source. The LLC then usually owes no US federal income tax. It still files Form 5472 with a pro forma Form 1120 every year, with a $25,000 penalty if it is missed. See Form 5472 explained and effectively connected income for when US tax would apply.

A Wyoming LLC with EIN is $249, or $348 with a US bank account, then $121 a year.

US C-Corp formation for an AI data company

A Delaware C-Corporation pays 21% federal tax on its profit, but it is the right choice if you plan to:

• raise money from US investors, many of whom only invest in Delaware corporations
• hire US staff, such as a sales lead or project managers in the USA
• build your own platform and sell software as well as services
• be acquired, which happens often in this industry

A Delaware corporation with EIN is $358, then $578 a year. See LLC vs Inc for non-residents and raising investment as a foreign founder. If you are moving an existing Indian company under a US parent, read the Delaware flip guide.

What US AI buyers will ask you

Selling to AI teams means passing vendor onboarding. Expect these requests, and prepare for them before the first sales call:

• Mutual NDA before any sample data is shared, then a master services agreement with confidentiality and IP terms
• Security questionnaire, often 100 to 300 questions on access control, device security, encryption, staff vetting and incident response
• SOC 2 Type II or ISO 27001. Small vendors often start without one and offer a completed questionnaire, policies and a roadmap. Larger buyers may insist, and the audit takes months, so start early if you target them
• Secure workspace: the client's platform or a virtual desktop, no downloads, no personal devices, no phones on the work floor for sensitive projects, sometimes a dedicated room
• Background checks and signed confidentiality agreements for every annotator
• Quality guarantees: accuracy targets, gold-standard test sets, inter-annotator agreement and rework at your cost
• IP assignment: all labels, outputs and work product belong to the client
• Data processing terms, which make you a service provider or contractor under state privacy laws such as California's CCPA: you may only use the data to do the client's work

A buyer who finds out your annotators used personal laptops, a public AI tool or an unvetted subcontractor will end the contract. Write your security policies down, train every annotator, and keep proof. See contracts with US clients.

US rules that apply to AI training data

Medical data and HIPAA

Annotating medical images, clinical notes or patient recordings for a hospital, health insurer or their vendors usually means handling protected health information (PHI). Under HIPAA, your company becomes a business associate or subcontractor and signs a business associate agreement (BAA). That brings duties on access control, encryption, staff training, audit logs and breach notification. Many AI clients avoid this by sending you de-identified data, which is not PHI, so ask whether the data has been de-identified before you price the project. See the medical billing company guide for more on running HIPAA work from abroad.

Faces, voices and biometric privacy

Several US states regulate biometric identifiers: face geometry, fingerprints, iris scans and voiceprints.

• Illinois BIPA requires written notice and consent before biometric identifiers are collected, and allows people to sue, with damages per violation. It has led to large class actions against companies using face data
• Texas and Washington have their own biometric laws, enforced by the state
• Several state privacy laws, including California's, treat biometric data as sensitive

The duty to get consent usually sits with the company that collects the data, your client. But a vendor that receives, stores or processes biometric data can be drawn into a claim. If a project involves face recognition, face landmarks or speaker identification, ask the client to confirm in writing that consent was obtained, keep data only on the client's systems, and delete it when the project ends.

Children's data

Projects involving data from children under 13, such as kids' app recordings or images, bring COPPA obligations for the client and strict handling rules for you. Many vendors decline this work or accept it only with tight controls.

Export controls and defence work

Most commercial annotation is not export-controlled. A few projects can be:

• Defence and military data, such as aerial surveillance, weapons or satellite imagery for government programmes, may fall under ITAR or the EAR, which can bar foreign nationals from seeing it
• Some advanced technical data about controlled technologies can need a licence to send abroad

If a project touches defence, government or controlled technology, the client must confirm it can legally be done by a team outside the USA. Don't accept it on a handshake.

Content moderation and wellbeing

Moderation and red-teaming expose annotators to graphic, violent or abusive material. US clients increasingly ask about wellbeing programmes: exposure limits, rotation, counselling and opt-outs. Your local labour law also applies. Workers in several countries have sued vendors over moderation work, so build this into your contracts and pricing.

AI-specific laws

US AI rules are mostly at state level and aimed at the developers and deployers of AI systems, not data vendors. Colorado's AI Act, California's training data transparency law and others may lead your clients to ask where your data came from and how it was labelled. Keep records of sources, consent, guidelines and who labelled what, so you can answer.

Your team abroad

Most annotation companies have a full-time core team and a pool of part-time or project annotators.

• Full-time staff are usually employed through a local company in your home country, which invoices the US company under an intercompany service agreement. The US company pays for the services, and the local company pays salaries and local taxes. Transactions between the US LLC and a related foreign company are reported on Form 5472
• Freelance annotators abroad sign a contractor agreement with confidentiality and IP assignment, and give a Form W-8BEN. Foreign individuals working outside the USA don't get a Form 1099. See paying contractors abroad
• Expert annotators (doctors, lawyers, engineers) need contracts that make clear they are not giving professional advice, and that the work is confidential
• Local labour law decides whether a long-term "freelancer" is really an employee, not US law

Indian residents who own a US company must follow the ODI rules under FEMA. See ODI and FEMA rules for a US company.

Getting paid by US AI companies

• Vendor onboarding: give your US company's Form W-9. A US LLC owned by a foreign person still gives a W-9, because the LLC itself is a US entity. If you invoice from your foreign company instead, you give a Form W-8BEN-E. See W-8BEN-E vs W-8BEN
• Payment terms: large buyers often pay net 30 to net 60. Ask for a deposit or monthly billing on long projects, because annotation is labour-heavy and you pay your team first
• Payment method: ACH to your US account is free or cheap. Wires from the USA to a foreign account cost more and take longer. See receiving payments from US clients
• Card payments from small start-ups can go through Stripe. See Stripe for non-residents

Sales tax

Annotation and labelling are services, and most US states do not tax them. A few states tax data processing or information services, Texas for example, and some tax software or digital products. If you sell a platform or dataset, not only labour, sales tax becomes more likely. Most small vendors have no US sales tax obligation, but check once your sales in any state grow. See state taxes for a foreign-owned LLC.

Selling datasets you own

Some annotation companies build their own datasets, such as speech in a regional language or labelled retail images, and license them. This is a different business with different risks:

• you need rights to every item: consent from the people recorded or photographed, and licences for any content you didn't create
• scraped web data carries copyright and terms-of-service risk
• buyers want provenance records and warranties

A licensing business usually fits better in a C-Corporation, and you may need to collect sales tax in some states on dataset licences.

Examples

These examples are illustrative.

Ananya, Hyderabad: radiology annotation. Ananya runs a team of 18, including six radiologists who review the hardest cases. A Wyoming LLC signs contracts with US medical AI start-ups and is paid by ACH. Every client sends de-identified scans; where one wanted to send identified data, Ananya signed a BAA, moved the work onto the client's virtual desktops and wrote a HIPAA policy set. The team works abroad and there is no US office, so the LLC files Form 5472 and owes no US federal income tax. The Indian company employs the team and invoices the LLC under a service agreement, filed under the ODI rules.

Joy, Cebu: RLHF and writing tasks. Joy started with five English writers ranking chatbot answers for a US data platform, and now has 40. A US LLC gives the client a W-9 and gets paid net 30. Joy's team works only inside the client's platform, with no downloads, and each writer signed an NDA and IP assignment. A competitor lost a client after an annotator pasted tasks into a public AI tool, so Joy blocks those sites on work devices.

Andrés, Medellín: Spanish and Portuguese speech data. Andrés records and transcribes Latin American speech for voice AI companies. Because the business licenses its own recordings, every speaker signs a consent and release, and records are kept for each file. Andrés chose a Delaware C-Corporation, raised a seed round from US investors, and hired a US-based sales lead, so the corporation files Form 1120 and pays US tax on its profit.

Kasia, Kraków: model evaluation. Kasia's team of software engineers tests coding models for US AI companies. A US LLC holds the contracts, and the engineers are freelancers in Poland who give Form W-8BEN. When one large client asked for SOC 2, Kasia started with a readiness assessment and gave the client a full security questionnaire and policies in the meantime.

Common mistakes

• Annotators using personal devices or downloading client data.
• Pasting client tasks into public AI tools, which breaches almost every NDA.
• Taking medical data without asking if it is de-identified, or without a BAA.
• Face or voice projects without written confirmation that consent was obtained.
• Accepting defence or government data without checking export controls.
• Pricing per item before testing how long items really take.
• No IP assignment from freelancers, so the client's ownership has a gap.
• Forgetting Form 5472 because the LLC owes no US income tax.

Frequently asked questions

Can I start a data annotation company for US clients from India or the Philippines without visiting? Yes. The company, EIN and US bank account can all be set up remotely.

Do I need an SSN or ITIN to get the EIN? No. Foreign owners apply with Form SS-4 and don't need either.

Will my US LLC pay US tax on annotation fees? Usually not, if all the work is done abroad and there is no US office or staff. It still files Form 5472 every year.

Do I need SOC 2 to sell to US AI companies? Not always. Many start-ups accept a security questionnaire and written policies. Larger buyers and anything involving sensitive data may require SOC 2 or ISO 27001.

Can my team label medical images? Yes. De-identified data is simplest. Identified patient data requires HIPAA safeguards and a business associate agreement.

Do I need a licence to provide AI training data in the USA? No general licence is needed. Your duties come from contracts, privacy laws, HIPAA where relevant and export controls.

How UCB can help

We form US LLCs and C-Corporations for data annotation and AI training data companies outside the USA, get the EIN without an SSN, and help open a US business bank account without visiting, through our formation, EIN and bank account package. Each year we prepare Form 5472 with the pro forma 1120, Form 1120 or Form 1065, through our US tax filing service. We do not provide security audits, HIPAA, privacy or export control advice.

Related guides: IT services and software development, data enrichment and database companies and VA and BPO companies.

Starting an AI data company for US clients? WhatsApp us at +1 315 961 2956 or email enquiries@ucbsolutions.com.