AI development, US
Outsource AI development without losing control of your data, costs or IP
To outsource AI development safely, keep the model accounts, data and evaluation sets in your company's name and hire a team that proves quality with tests before it ships. We are an India-based team that builds LLM features for US companies on OpenAI or Anthropic APIs billed to your accounts, with written evals, spending limits and IP assigned to you.
- AI feature add-on
- $2,200–7,500 indicative
- Build time
- 2–4 weeks per feature
- Model accounts
- Yours, billed to you
- Acceptance
- Eval results you agreed
Updated · By the Codebrix team
What US companies outsource, and what they should keep
Outsourcing works when the job can be written down and measured. It fails when the vendor decides what the AI may tell customers, or holds the only copy of the prompts.
| Reasonable to outsource | Keep in-house |
|---|---|
| Search and answers over your documents (RAG) | The rules for what the AI may and may not say |
| Drafting replies for support or sales staff to approve | Approval of any claim made about the AI in marketing |
| Pulling fields out of PDFs, emails and forms | Access to live customer data in production |
| Internal agents that call your tools and APIs | The model provider contracts and billing |
| The eval harness, dashboards and cost alerts | The final pass mark for each eval |
If you would rather add engineers to your own AI team than buy a fixed-scope feature, see hiring AI developers. Our general method is on the AI integration services page; this one covers a US buyer working with a team in India.
US customer data: where it travels and who sees it
An LLM feature sends text to a model provider on every request. Three questions decide your risk: does the provider train on it, how long does it keep it, and can anyone on the vendor's team see it. We design for the third first: our engineers build and test on synthetic or redacted data, production keys sit in your secrets manager, and production logs live in your cloud account.
| Question | OpenAI API | Anthropic API |
|---|---|---|
| Used to train models by default? | No. OpenAI says API data has not been used for training since 1 March 2023 unless you opt in | No. Anthropic says it does not use inputs or outputs from commercial products for training by default |
| Exceptions | Data you explicitly opt in to share | Feedback you submit, such as thumbs up or down, and data you choose to allow |
| Retention | Abuse-monitoring logs kept up to 30 days unless law requires longer | Feedback conversations kept up to five years; account identifiers removed before use |
| Reduced retention | Zero Data Retention or Modified Abuse Monitoring, subject to OpenAI's approval | Check your agreement with Anthropic |
| Keep processing in the US | Check OpenAI's current options | US-only inference via the inference_geo setting on Claude 4.6 and later models, at 1.1 times standard prices |
If the data includes health information, the rules tighten considerably; our page on HIPAA app development explains how we approach it. Whether a provider's terms are acceptable for your data is a decision for your counsel; we give them the data-flow diagram they need to make it.
Evals as the acceptance test for each milestone
With ordinary software, a milestone is done when the feature works. With AI, it works most of the time, so done needs a number. Before building, we agree with you a set of real, anonymised examples and the score each check must reach. The milestone is accepted when the eval report shows those scores, and the eval set and harness are yours.
| Check | What it measures | Who sets the pass mark |
|---|---|---|
| Correctness | Answer matches the expected answer for each example | You |
| Grounding | Every claim is supported by the retrieved source | You |
| Refusal | The AI declines out-of-scope or risky requests | You and your counsel |
| Latency | Time to first token and to full answer | Agreed together |
| Cost per task | Average and worst-case spend per request | You |
The time difference helps here: we run eval suites during our day and send the report before your morning, so each prompt or model change arrives with evidence, not a promise.
Cost control you can audit
API bills go to your card, so the controls must be visible to you, not buried in our code. Two provider features do most of the work. Anthropic's prompt caching charges cache reads at 0.1 times the base input price on most models (0.05 times on Claude Opus 5.5 and Sonnet 5.5), with a 5-minute cache write at 1.25 times. Its Message Batches API halves input and output prices for work that can wait; Anthropic says most batches finish within an hour and results are available within 24 hours. These figures are as of November 2026; check Anthropic's pricing page.
- Separate API keys per environment and per feature, so a spike is traceable.
- Spend limits and usage alerts set in your provider console.
- A cheaper model for routing and simple tasks, a stronger one only where evals show it is needed.
- Stable instructions and reference text placed first in the prompt so caching applies.
- Overnight and bulk jobs sent through batch processing.
- A monthly cost report per feature, against the estimate in the proposal.
Who owns the prompts, evals and outputs
The contract should assign to your company the code, prompts, eval sets, fine-tuning data and any fine-tuned model configuration created for you, and treat them as your confidential information. That matters more for AI work than usual because of how US copyright treats AI output. The US Copyright Office's January 2025 report on copyrightability concluded that, with current technology, prompts alone do not make a person the author of the output, and that human contributions such as arranging or modifying output are assessed case by case. The report is policy guidance, not a statute.
So the practical protection for your prompts and datasets is the contract and access control, not copyright in what the model writes. Our software development contract checklist lists the clauses; your counsel decides the wording.
What you say about the AI in public
On 25 September 2024 the FTC announced Operation AI Comply, a set of enforcement actions over deceptive AI claims. The agency's chair said there is no AI exemption from the laws on the books. One case concerned a service marketed as a robot lawyer; the FTC alleged the company never tested whether its chatbot performed at the level of a human lawyer.
Price, timeline and how to start
Our published indicative range for an AI assistant or automation added to an existing product is $2,200–7,500 over 2–4 weeks. A new web application with AI at its core sits in our $5,500–22,000 range for web apps. The fixed price follows a discovery call; you pay per milestone in USD. Model usage is billed by the provider to your account, and we estimate it from your volumes before you commit. A paid discovery that ends with an eval set and a working prototype is a low-risk first step. The parent page for US companies covers time zones, contracts and payment.
Frequently asked questions
Is it safe to outsource AI development that touches US customer data?
It can be, if the design limits who sees the data. We build and test on synthetic or redacted data, production keys stay in your secrets manager, and logs stay in your cloud account. The model provider's data terms apply to every request, so your counsel should review them against your obligations before launch.
Will OpenAI or Anthropic train on our data?
Not by default for API customers, according to both companies' published terms as of November 2026. OpenAI says API data is not used for training unless you opt in, and keeps abuse-monitoring logs for up to 30 days. Anthropic says it does not train on commercial inputs or outputs by default, with exceptions such as feedback you submit.
How do we know the AI feature is good enough to accept?
We agree an eval set of real, anonymised examples and a pass mark for each check before building: correctness, grounding, refusals, latency and cost per task. The milestone is accepted when the report meets those marks. You keep the eval set and harness, so any future vendor or in-house engineer can rerun the same tests.
Who owns the prompts and outputs?
The contract should assign prompts, eval sets, training data and code to your company. The US Copyright Office's 2025 report concluded that prompts alone generally do not make someone the author of AI output, so rely on the contract and confidentiality terms rather than copyright to protect that work. Your counsel should confirm the wording.
How do you keep the monthly AI bill predictable?
Separate keys per feature, spending limits and alerts in your provider console, prompt caching, batch processing for work that can wait, and a cheaper model wherever evals show it is good enough. Every month you get a cost report per feature compared with the estimate we gave before the build started.
Can we market the feature as AI-powered?
You can describe what it does truthfully. The FTC has brought cases over unsupported AI claims and said there is no AI exemption from existing law. Keep claims about accuracy or savings within what your eval results and real usage show, and have counsel review them. We supply the eval reports that back the claims.
Sources
- OpenAI: Data controls in the OpenAI platform
- Anthropic Privacy Center: Is my data used for model training?
- Anthropic: Claude API pricing (caching, batch, data residency)
- Anthropic: Batch processing
- US Copyright Office: Copyright and Artificial Intelligence, Part 2 (Copyrightability)
- FTC: FTC announces crackdown on deceptive AI claims and schemes (25 Sep 2024)
Related
United States
An offshore web development company in India for US businesses
Offshore web development company in India for US businesses: USD pricing, Stripe or wire payment, IP assigned to you, and a planned ET and PT overlap.
Read moreAI integration
AI integration services for the software you already run
AI integration services that add LLM assistants, document search, summaries and classification to your product, tested on your data with cost limits set.
Read moreHire developers
Hire AI developers who build, test and ship LLM features
Hire AI developers who build LLM features with retrieval, agents and evals in Python or TypeScript. Paid first milestone, your keys, rate after a call.
Read moreAI automation
An AI automation agency for the repetitive work behind your business
AI automation agency that moves email triage, lead follow-up, CRM updates and WhatsApp replies to AI workflows, with people approving the risky steps.
Read moreGuide for US companies
A software development contract checklist for US companies hiring an outside team
A software development contract checklist for US companies: IP assignment, escrow, acceptance, warranty, liability, disputes, data terms and handover.
Read moreHealthcare apps
HIPAA compliant app development: what we build, and what stays your job
HIPAA compliant app development explained: the Security Rule safeguards we build in, BAAs with AWS, Google Cloud or Azure, and PHI kept to the minimum.
Read more