Most new tech businesses lose money on paperwork they skipped, not on bad work. A client says the chatbot was supposed to do more. A dental office learns you pasted patient names into a consumer chat app. A project runs four months past the quote because nobody wrote down what "done" meant. A one-page contract and a few habits about client data prevent nearly all of it.
This chapter covers what your services contract should say, who owns work you make with AI, how to handle client data, the industry rules that come with health, tax, school and hiring clients, the marketing and calling rules the FTC and FCC enforce, the newer AI laws, and the insurance that backs it all up. It is a map, not legal advice. Where a rule could cost you real money, the chapter says so and tells you when to pay a lawyer.
The contract: what it has to say
You need a short document both sides read before work starts, not a 30-page agreement. Many people sign a master services agreement once and a short statement of work per project. For a first client, one document of two or three pages is fine.
| Clause | What it says | Why it matters |
|---|---|---|
| Scope | Exactly what you will build or do, in a numbered list, and what is not included | "Connect the web form to the CRM and send a welcome email" is a scope. "Automate the office" is not. |
| Acceptance | How the client checks the work, and that it is accepted after a set number of days with no written problems | Without it, a project never ends |
| Price and payment | Fee, deposit, due dates, late fee, and that work pauses if invoices are unpaid | A 50% deposit on setup work is common and fair |
| Change requests | New requests get a written quote before you start them | Stops scope creep politely |
| Tools and who pays | Which software subscriptions the project needs, whose name they are in, and who pays each month | Put tools in the client’s name and on their card, so the system survives if you leave |
| Client data | What data you will touch, which AI tools may see it, how you protect it, and that you delete or return it at the end | Covered in its own section below |
| Ownership | Who owns the finished work, what you keep (your templates and know-how), and a note on AI output | See the copyright section below |
| AI accuracy | AI output can be wrong; the client reviews anything sent to customers; you do not promise a set accuracy rate | Protects you and keeps your marketing honest |
| Limitation of liability | Your total liability is capped, often at the fees paid in the last few months, and excludes lost profits | One bad automation should not cost more than the job paid |
| Termination | Either side can end it with notice; the client pays for work done; you hand over logins and files | A clean exit keeps a referral even when a client leaves |
| Confidentiality | Both sides keep the other’s private information private | Often enough to replace a separate NDA |
| Governing law | Which state’s law applies and where disputes go | Usually your home state |
The Main Street library’s chapter on leases and contracts covers signing basics that apply here too. If you charge a monthly care plan, the contract should also say what the plan covers each month, the response time you promise, and what counts as extra work. The pricing side of that is in Pricing AI and IT services.
Who owns work made with AI
Clients often assume they will own a copyright in everything you deliver. With AI output, that assumption can be wrong, and your contract should say so plainly.
The U.S. Copyright Office set out its view in Part 2 of its AI report, on copyrightability, published January 29, 2025. In short: AI output is protected only where a human author decided enough of the expressive elements. Typing prompts alone is not enough. Human work that remains visible in the output can be protected, and so can a human’s creative selection, arrangement or editing of the output. Using AI as a helper does not by itself block copyright.
What that means for a contract:
Assign whatever rights exist. Write that you assign to the client all rights you hold in the deliverables, "to the extent any exist," instead of promising that every piece is copyrighted.
Say where AI was used. One sentence is enough: "Some drafts, images or code in this project were produced with AI tools and edited by us." Clients who plan to register a copyright need to know.
Keep a record of human work. Save drafts and edit history for logos, scripts, course material and other creative work. That record is what shows human authorship if anyone asks.
Check the tool’s terms. Most AI vendors give the user the rights they hold in the output. Some plans restrict commercial use. For example, the ElevenLabs Free plan does not include a commercial license, and Midjourney requires companies with more than $1 million a year in gross revenue to be on its Pro or Mega plan. Use a plan that allows client work. The plan prices are in Your AI toolkit.
Keep what is yours. Your templates, prompts and reusable workflows stay yours; the client gets a license to use them. That is how you sell a similar setup to the next client.
Client data, NDAs and AI tools
In Bluevine’s survey of 942 owners (April 2026), 33% named data security and privacy as a barrier to AI, up from 23% a year earlier. A clear data plan is part of what you sell.
NDAs. Sign a separate non-disclosure agreement before a discovery call where the client will show you real files, or when a larger client asks. Watch for one-sided terms, no end date, and clauses that would stop you working for anyone else in the industry.
Use business plans for client data. Consumer chat plans and business plans come with different data terms. Business tiers usually say customer data is not used to train models by default and give admin controls. Read the current terms of the exact plan you use, and write which tools you use into the contract.
Take only what you need, in their accounts. An invoice automation needs ten sample invoices, not the whole accounting file; test with masked data where you can. Ask for logins inside the client’s own systems instead of files sent to you, so they can turn your access off at the end.
State privacy laws. About 20 states now have comprehensive consumer privacy laws (IAPP’s count, cited secondhand). Indiana, Kentucky and Rhode Island took effect January 1, 2026. Most apply only above a set number of consumers (Rhode Island’s line is 35,000), so a small client may be below them. But when you work for a larger company, its contract will often pass "processor" duties down to you: follow its instructions, protect the data, help with deletion requests, and report breaches. Read those clauses closely.
Industry rules: health, tax and schools
The best-paying niches in this section come with federal rules. That is why careful operators get paid there.
HIPAA (clinics, dentists, therapists, billing offices). If you create, receive, keep or send protected health information for a covered practice, you are a business associate under HIPAA (45 CFR 160.103). You need a signed business associate agreement (BAA) with the practice before you touch any patient data. Every tool that sees that data needs a BAA too, which rules out consumer chat accounts. Encrypting the data does not remove the need for a BAA. Many AI and automation vendors sign BAAs only on business or enterprise plans, so check before you design the workflow. The industry AI desk game plan shows how to build a business around this.
IRC 7216 (tax preparers). Section 7216 of the Internal Revenue Code makes it a crime for a tax preparer to disclose or use tax-return information without the taxpayer’s consent. Consent has to be a separate written document for each disclosure or use, following Rev. Proc. 2008-35. The civil penalty under section 6713 is $250 per disclosure, up to $10,000 a year. Commentators say sending return data to a public AI tool likely needs a 7216 consent that names the recipient; the IRS has not issued AI-specific guidance. If you run a tax-season document desk, build the consent step into the workflow and let the CPA own the wording.
FERPA (schools). Schools may share student education records with an outside vendor only under the "school official" exception: the vendor has a legitimate educational interest, works under the school’s direct control, and does not pass the data on (34 CFR 99.31). In practice that means the school’s contract terms, not yours.
COPPA (children under 13). If anything you build collects personal information online from children under 13, COPPA applies, including verifiable parental consent. The FTC’s amended COPPA rule had a general compliance date of April 22, 2026. Tutoring sites, youth league sign-up forms and kids’ apps are where this comes up.
Marketing rules: AI claims, reviews and calls
FTC rules on AI claims. The FTC has brought a run of cases over AI claims. DoNotPay paid $193,000 over its "robot lawyer" claims (final order February 2025). Workado’s final order (August 2025) came from claiming its AI content detector was 98% accurate without proof. Air AI’s proposed settlement (March 2026) bans its owners from selling business opportunities after the FTC said it misled small businesses with earnings and refund claims. The safe rule: do not promise income, savings percentages, accuracy rates or "fully automated" results you cannot prove, and never say your bot replaces a lawyer, doctor or licensed agent.
FTC Consumer Reviews and Testimonials Rule (16 CFR Part 465). In effect since October 2024. It bans fake reviews, including AI-written reviews that misrepresent the reviewer’s experience, paying or rewarding reviews conditioned on a positive opinion, undisclosed insider reviews, and suppressing negative reviews. Civil penalties apply per violation. Asking every customer for an honest review is fine. If you sell review-response or reputation services, design them around this rule. The marketing playbook has review-request wording that stays inside it.
TCPA and AI voices. The FCC ruled on February 8, 2024 that AI-generated or cloned voices count as an "artificial or prerecorded voice" under the Telephone Consumer Protection Act. Outbound calls using them need the called person’s prior express consent, and written consent for telemarketing, plus identification and an opt-out. An AI receptionist answering calls customers place is a different, lower risk. Outbound AI sales calling without consent is the danger zone; this section does not recommend it.
Business texting. Texts sent from ordinary 10-digit numbers must be registered with the carriers under A2P 10DLC. Marketing texts also need consent under the TCPA. If you set up missed-call text-back or reminders, the registration is part of your job.
Email. CAN-SPAM covers commercial email: no misleading subject lines, a real postal address, and a working unsubscribe. Mass AI cold email is legal under it but wrecks your domain’s reputation.
AI laws that touch what you build
Most of what a small AI services firm builds, such as automations, document extraction and staff assistants, is not covered by the AI-specific laws below. They matter when your work helps make decisions about people, especially hiring, or when your clients serve Europe.
NYC Local Law 144. Employers and employment agencies in New York City may not use an automated employment decision tool unless it has had an independent bias audit within the past year, a summary of the results is public, and candidates got notice 10 business days before use. Enforcement began July 5, 2023. If you set up recruiting tools for a NYC client, you do not do the audit; an independent auditor does. Your job is to inventory the tools and set up the notices.
Illinois HB 3773. It amends the Illinois Human Rights Act, effective January 1, 2026. It bars using AI in employment decisions in ways that have a discriminatory effect, bars using zip codes as a stand-in for protected traits, and requires notice to workers. It does not require an audit.
Colorado SB 26-189. Signed May 14, 2026, it repealed and replaced Colorado’s earlier AI law (SB 24-205) with a narrower one on automated decision-making technology used in "consequential decisions" about education, employment, housing, lending, insurance, health care or government services. Developers must give deployers documentation, and deployers must give notice, explain adverse outcomes within 30 days, and offer correction and human review. The legislature’s summary dates the developer duties and the attorney general’s rules to January 1, 2027. The attorney general enforces it, with a 60-day cure period until 2030. The hiring-tool notices idea is built on these three laws.
EU AI Act, if you serve EU clients. It entered into force August 1, 2024. Since August 2, 2026, its Article 50 transparency duties apply: tell people when they are talking to a chatbot, label deepfakes, and mark AI-generated content in a machine-readable way (systems already on the market have until December 2, 2026 for the marking part). The heavier "high-risk" duties were pushed back by the 2026 Digital Omnibus to December 2027 and August 2028. A simple website bot for a European client needs a clear "you are chatting with an automated assistant" line and a way to reach a human.
Insurance: what to carry
| Policy | What it covers | Who needs it |
|---|---|---|
| Professional liability (errors and omissions, E&O) | Claims that your work or advice caused a client a financial loss, such as a broken automation that lost orders | Anyone building systems clients rely on. Estimates in our research run about $500 to $1,500 a year for a small firm; get quotes |
| Cyber liability | Costs after a data breach: notification, recovery, some legal costs. Some policies cover a client’s loss caused by your breach | Anyone with access to client systems or personal data; clinics and accountants may require it |
| General liability | Injury or property damage, such as tripping a cable at a client’s office | Anyone who works on site; landlords and venues often ask for a certificate |
Many insurers sell E&O and cyber together. Ask the agent: does it cover work done with AI tools, a client’s loss from my mistake, and contractors I hire? The Main Street library’s chapter on business insurance explains the policy types in more depth, and the one on legal structure covers how an LLC fits alongside insurance.
When to pay a lawyer
You can write a sound first contract from the checklist above. Pay a business lawyer for an hour or two at these moments:
Before your first health, tax or legal client. Have a lawyer review your BAA template or your 7216 consent workflow.
When a client sends you their contract. Large clients use their own paper, often with unlimited liability and broad indemnity. A lawyer can tell you which two clauses to push back on.
Before hiring-tool or AI-decision work. NYC, Illinois and Colorado rules are new, and your notices are close to legal advice. Work alongside an employment lawyer.
When something goes wrong. A data leak, a threatening letter, or a client refusing to pay a large invoice. Call your insurer and a lawyer before you reply.
Free first stops: your local SBDC and SCORE mentors can review a contract in general terms, and many state bar associations run lawyer referral services with a low-cost first consultation. The free help directory lists where to find them.
Questions people ask
Do I need a contract for a small AI automation job?
Yes. Even a $500 job should have a one-page scope, price, payment terms, a data clause and a liability cap. Small jobs are where people skip it and then argue over what was included.
Who owns content I create with ChatGPT or Claude for a client?
Most AI vendors give their rights in the output to you, and you can pass them to the client. But under the Copyright Office’s January 2025 report, purely AI-generated material may not be copyrightable at all. Assign "whatever rights exist," tell the client where AI was used, and keep a record of your human edits.
Can I use ChatGPT with patient data if I remove the names?
Properly de-identified data is outside HIPAA, but the HIPAA standard for de-identification is stricter than deleting names. For anything that is still protected health information, use only tools whose vendor signs a BAA, and sign one with the practice yourself.
Do I need insurance for a one-person AI consulting business?
No federal law requires it, but clients in health, finance and law often ask for proof of E&O and cyber coverage before signing. It also protects you if a mistake costs a client money. Get quotes before your first larger client.
Is it legal to set up an AI voice agent that calls customers?
Outbound calls with an AI voice need the called person’s prior express consent under the FCC’s February 2024 ruling, and written consent for marketing calls. Answering inbound calls customers make is far lower risk. Call recording also needs everyone’s consent in some states, such as California.
Checked November 2026: the Copyright Office report, Colorado SB 26-189 on the legislature’s site, the NYC Local Law 144 page, and the FTC cases listed. Not legal advice. Re-check Colorado’s attorney general rules before January 2027, the FTC’s July 2026 proposed policy statement on AI accuracy, and any vendor’s BAA terms before you rely on them.