Health Tech Pilot Agreement: One Deal, Two Ways of Asking a Legal AI

The deal

The company is an Australian pre-seed startup with two founders. One cared for his father as a teenager, and later for his grandfather. The other ran technology at a large dementia care provider. They build ambient sensing for people living with cognitive decline at home: Wi-Fi sensing, a small radar unit and an edge hub. No cameras, no microphones, nothing to wear.

The software is the interesting part. It learns how one specific person moves, sleeps and uses the home, and tells care staff when that pattern changes. A carer who visits on Tuesday and Friday gets to know what happened on Wednesday night.

Their customers are home care providers. In the US, one of the most useful partners is a home care company that runs a large franchise network. Its answer to the founders was short: prove it in a hundred homes, then come back.

That one sentence means three documents. A mutual NDA so both sides can talk. A letter of intent that sets out the pilot and what happens if it works. And a pilot agreement covering 100 real homes with real people in them. The founders want Australian law. The partner is American. Nobody wants to spend a quarter trading redlines.

We treated it as a field test. Producing an NDA isn't the hard part. The question founders face is different: how much do you need to know, and write down, before an AI draft is worth handing to your lawyer? So we ran the same deal twice.

Five traps in an in-home care pilot

Most pilot agreement guides cover the same ground: scope, duration, success criteria, fees, IP, confidentiality. All of that still applies here. In-home care adds five problems, and they became the backbone of how we scored both runs.

1. The sensor sees everyone in the house

A wearable records the person wearing it. A radar or Wi-Fi sensor records whoever is in the room: a spouse, a visiting daughter, the carer, the plumber. Consent can't stop at the client. The agreement needs a household and visitor notice, a named party responsible for delivering it, and a rule for data about people who did not sign anything.

2. The person in the home may not be able to consent

Many of the people this product is built for live with dementia. Some can give informed consent and some cannot, and a diagnosis alone doesn't settle which. In the US the decision usually falls to a healthcare proxy or a power of attorney. In Australia it can be an enduring guardian. In England, the Mental Capacity Act 2005 and lasting powers of attorney apply. The pilot agreement has to say who collects consent, from whom, on which form, and who carries the risk if it was collected wrongly.

3. One sentence can turn a sensor into a medical device

This trap hides in marketing copy rather than in the contract. The US FDA updated its General Wellness guidance on 6 January 2026 and reissued its Clinical Decision Support guidance on 29 January 2026. Low-risk products intended only for general wellness can stay outside device review. A product that claims to diagnose, predict or prevent a specific disease is a device. "Notices changes in night-time routine" is one kind of claim. "Predicts urinary tract infections" is another. Australia's TGA runs its own rules for software as a medical device.

So the pilot agreement needs an intended-use clause and a claims restriction that binds both parties. The startup can watch every word it writes and still end up regulated if the partner's sales team tells a family the product predicts falls.

4. A franchise network is many separate decisions

When the partner is a franchisor, the franchise owners are independent businesses. They deliver the care, hold the client relationships and choose their own vendors. A contract with headquarters doesn't oblige any single office to install anything. What tends to work is a pilot agreement with the franchisor plus a short participation addendum that each pilot office signs.

5. Australian law, American counterparty

The startup wants New South Wales law, which is reasonable. But a judgment from a Sydney court is hard to collect against a company whose assets are in the US. Arbitration usually travels better, because Australia and the US are both parties to the New York Convention on the recognition of arbitral awards. Australian law also brings its own checks: the unfair contract terms regime for small business standard form contracts, the Australian Consumer Law, and whether the Privacy Act's small business exemption is available to a company that handles health information.

Two ways in

Both runs used HAQQ, on the same day, on the same facts.

Ask A: hand over the raw material

We pasted what a founder's advisor has on day one: the transcript of a twenty-minute call with the founder, the founder's outreach email and a short instruction. In plain words, the instruction said: understand the situation and the market, think about every document this founder will need, and explain what a stake could look like if the startup partners with the franchisor. The anonymised version is in Appendix A.

Our first paste was wrong. Alongside our research notes on this deal, it carried the transcript of an unrelated call. HAQQ did not draft anything. It replied:

"The transcript and the appended brief concern two different matters. Before drafting, please confirm which one you want to pursue."

It then asked which output we wanted, whose interests the documents should protect and which jurisdiction applied. That's the right reflex. A drafting tool that blends two clients' facts into one contract is worse than having no tool at all.

With the right material in, HAQQ wrote a research memo and a plan of the documents the founder would need. We then pasted the written brief from Ask B in a single block and followed the next steps HAQQ itself suggested: draft the letter of intent, then research its binding clauses. Five messages in, we had a solid NDA, a letter of intent and a research note on the letter's enforceability. No pilot agreement. That was as much our doing as the tool's. We clicked through the suggestions instead of asking for the document we needed.

Ask B: a written brief, one document per message

For the second run we wrote a counsel-style brief with a general-purpose AI assistant. It says who we act for, sets out the facts, turns the five traps into constraints, and fixes the output format for every document: a founder summary, the clean draft, a negotiation table, open questions and the authorities relied on.

We argued about two things before sending it. The real party names stayed out, replaced by placeholders: the founders had shared their deck in confidence, and a drafting test doesn't need it. And we split the brief into three messages, one document each. Smaller asks are easier to review, and the letter of intent can build on the NDA already on the table. The brief is in Appendix B, exactly as sent, mistakes included.

How we scored it

Before either run, we wrote a 22-point checklist: 4 points on scoping, 8 on the in-home care traps, 6 on Australian and US law, and 4 on quality. Each point scored 0, 1 or 2. Our rule was strict: a 2 needs an actual clause or template in a document, while advice or a non-binding line in a letter of intent scores 1. We also checked every authority each run cited against its source.

The scoring and the citation checks were done with a general-purpose AI assistant and spot-checked by us. No lawyer has reviewed the drafts. Read the numbers as a structured reading of the output, not a legal opinion.

The scorecard

The written brief also cost less: 88.6 credits for the full pack, against 105.6 for a pack with no pilot agreement. Most of the difference is the detour. In Ask A, 54.2 of the 105.6 credits went on turns that produced no document: the clarifying round, the research memo and the research note on the letter of intent.

The headline gap is in drafting, not in spotting. On a looser rule that counts a correct non-binding principle in the letter of intent, Ask A scores 8 out of 8 on scoping and 16 out of 16 on the care traps. It found the issues. It didn't write the clauses, because it didn't get to the pilot agreement.

The citation result is the one we cared about most. Across both runs we checked 135 cited statutes, cases, regulator guidance and treaty references. None was invented. The four errors in Ask B were small: a broken link, a Washington statute cited with the wrong subsections, a threshold described as "below A$3 million" where the statute says "A$3 million or less", and a government web page cited for the right point from the wrong page. Seven could not be checked, mostly because court websites blocked our requests.

What the raw ask found on its own

Ask A is the better test of judgment. It had our rough research notes, but no instructions on what to draft or what to flag. Before it saw the written brief, it did five things we would expect of a careful junior lawyer.

It stopped at the wrong paste, as described above, and asked whose side it was on before writing a word.

It separated diagnosis from capacity. Living with dementia doesn't automatically mean someone can't consent, and a family member is not automatically authorised to consent for them. Our own notes had been looser on both points.

It pushed back on data ownership. Our research notes had assumed the startup could simply own all anonymised data. HAQQ proposed a narrower position: the startup keeps its platform and models, the client data stays with the client and the care provider, and anything used to improve the product must meet a stated de-identification standard.

It flagged a risk in the equity idea. We had floated giving the partner equity that vests as more homes go live. HAQQ noted that where federally reimbursed healthcare business is involved, compensation of that kind "can raise federal Anti-Kickback Statute issues". Our research had missed that entirely. It matters for this founder, and it changed the advice we will give.

It corrected our market research. Our notes said the partner had no in-home sensing offer. HAQQ found that it already sells a sensing service of its own. That changes the pitch: the startup has to explain what it adds over the partner's own service, not just why sensing matters.

What the written brief bought

Ask B produced the documents. The pilot agreement runs to 23 clauses and seven schedules, including a franchise office participation addendum, a data processing schedule, a business associate agreement that applies only if an office is covered by HIPAA, a consent template, a household and visitor notice, and a schedule for converting the pilot into a commercial agreement. It closes with a one-page note for the founder on the five clauses the partner is most likely to contest.

It asked before committing. Before drafting the binding agreement, it stopped with three questions: the parties' legal names, the pilot fee and currency, and which US states the pilot would run in. We answered "you choose". It chose, and labelled every choice as "proposals for approval, not facts about an existing agreement". That matters, because the prices and states it picked are its own inventions. A founder has to treat them as placeholders.

It also went beyond the brief in two places, and said so. It switched the possible-fall and unsafe-exit alarms off by default for the pilot, until a regulatory assessment clears them. And where we asked for a blanket mutual non-solicitation clause, it drafted a narrower restriction on misusing confidential information, and explained that California law limits broad no-hire clauses. A lawyer could reverse either call. Both are defensible, and both are flagged rather than slipped in.

Three mistakes in our own brief

We wrote the brief carefully. It still went out with three errors, and HAQQ caught all three.

The FDA date. Our brief said both FDA guidances were revised on 6 January 2026. Both runs corrected it: the General Wellness guidance is dated 6 January, but "the Clinical Decision Support guidance was superseded on 29 January 2026". We checked the FDA document. HAQQ was right.

The typo. When we edited the brief before sending, the post-pilot price went out as "3$". Ask B asked: "should I use US$3 per home per month, and does that cover software only or also hardware, installation and support?"

The partner's existing offer. This one came from Ask A's research, as described above. Our brief built a positioning argument on a fact that wasn't true.

A detailed brief makes the output better. It also passes your errors straight into the draft. What you want is a model that pushes back when the facts don't hold, and this one did, three times.

Four clauses, as drafted

All four come from Ask B's pilot agreement unless noted. "Startup", "Partner" and "Office" are the defined terms from the draft. Quoted as drafted, with the curly quotes straightened. Not legal advice, and not yet reviewed by a lawyer.

Intended use and the claims restriction (clauses 8.1, 8.3 and the opening of 8.5):

8.1 The agreed intended use is supplementary awareness of household activity and routine changes to support, but not replace, an Office's ordinary non-emergency care workflow.

8.3 Neither Party nor an Office may state or imply that the Platform diagnoses, predicts, prevents, treats, cures or mitigates a disease, injury or condition, including UTIs, dementia deterioration or falls. Neither may imply FDA approval, TGA approval or clinical validation not actually held.

8.5 Possible-fall functions, emergency escalation and unsafe-exit or elopement alarms are disabled for this Pilot unless a signed amendment follows a documented function-specific regulatory assessment and establishes lawful use, revised safeguards and any required approvals.

Consent and household notice (from clauses 9.1, 9.2 and 9.5):

Before collection, the Office must obtain informed, documented participation consent from the Client or, where the Client lacks capacity for the relevant decisions, an Authorised Representative.

Dementia or cognitive impairment is not itself treated as proof of incapacity. The Office must follow applicable decision-specific capacity requirements, verify the representative's identity, scope of authority and any conditions activating that authority, and keep the supporting record.

All adult household members must receive an accessible notice before activation. The Office must arrange notices to regular visitors and a visible notice for other visitors, using Schedule 6. Affirmative consent must be obtained wherever required; entering the Home does not itself constitute consent.

The franchise participation addendum (Schedule 2, from sections 1, 2 and 7):

The Office confirms receipt of the Agreement and agrees directly with Startup to comply with its provisions applicable to an Office, including care responsibilities, Hardware custody, licence restrictions, claims restrictions, consent, data protection, confidentiality, indemnities, liability limits and disputes.

Startup agrees directly to provide the allocated Services to the Office on those terms.

Partner remains the central payer. The Office does not guarantee Partner's subscription fees unless a separate signed arrangement expressly says so.

The Office expressly agrees to NSW law and the Sydney-seated ACICA arbitration agreement in clause 22, including direct enforcement by and against Startup. Clients are not parties to that arbitration agreement.

Ask A's letter of intent reached the same structure in one line: "Partner's signature alone would not bind an Office or guarantee network adoption."

Disputes (from the arbitration clause):

The legal seat is Sydney, Australia; the language is English; and there will be one arbitrator. The tribunal may conduct hearings remotely and should adopt proportionate disclosure and procedure.

A contracting person may seek interim or conservatory relief from a competent court, including a US court where information, a respondent or assets are located, without waiving arbitration.

What a lawyer would still fix

These are first drafts, and they read like it. A lawyer would start here.

Length. The NDAs run to about 4,000 words each and the pilot agreement to about 13,000. That's a lot of paper for a 100-home pilot, and it works against our own instruction that the partner should sign "without a long redline".

The HIPAA test. Both runs treat a business associate agreement as conditional, which is right, but neither spells out the test. A home care office is covered by HIPAA only if it conducts certain standard electronic transactions, such as billing a health plan. A private-pay, non-medical office may not be covered at all. If it isn't, state consumer health data laws and the FTC's Health Breach Notification Rule do more of the work.

Australian consumer law. Under the Australian Consumer Law, a business buying services for up to A$100,000 can count as a consumer, and the statutory consumer rights that come with that can only be limited in specific ways. Both runs flag the rule. Neither applies it to the pilot price HAQQ itself proposed, which sits under that line, or drafts the limited cap the law allows.

One parties block. The party details drift between the three documents, and each letter of intent only takes effect once the NDA is signed, which leaves a gap. A lawyer would fix both in an afternoon.

Clean-up. HAQQ left some raw citation markers in the NDA and letter of intent text. They are easy to strip, but they have to go before anything reaches the other side.

The real stake is the data

The obvious question is how much of the company a partner like this would want. We'd settle something else first: who owns what the pilot produces.

A pilot in 100 homes creates three kinds of data:

• Client data: what happened in one person's home. It belongs with the client and the care provider.
• The startup's platform and models, including every improvement made during the pilot.
• De-identified, aggregated patterns across all the homes. This is what makes the next model better, and what an insurer or health system might pay for one day.

A partner with its own technology platform will want the signals inside its own system, and for client data that's reasonable. It's a problem if the contract quietly hands over the third category, or gives the partner a right of first refusal on a sale of the company. Both runs refused a right of first refusal. Ask B wrote the data split into the agreement; Ask A set it out as a principle in its letter of intent.

How to brief a legal AI on a deal like this

Running both changed how we'd do it next time. The two approaches do different jobs.

Let the raw ask go first. Handed the founder's own material with no framing, HAQQ asked the right questions and found three things our brief had wrong or missed: an outdated date, a fact about the partner, and a regulatory risk in the equity idea. A brief tells the model what you already think. The raw ask is where you find out what you didn't.

Then decide the two things that change every clause: whose side you are on, and which law governs. Ask A stopped to ask both. Ask B had them in the first line.

Write the traps down as constraints. Every point Ask B scored that Ask A didn't was something the brief named explicitly: the household notice, the consent template, the franchise addendum. If a risk matters, say it in the brief.

Ask for one document per message, and name it. The suggested next steps are useful prompts, not a plan. We followed them in Ask A and didn't reach the pilot agreement.

Use placeholders for the parties. It protects whoever shared the facts with you, and it costs nothing in quality.

Ask for a negotiation table. A draft tells you what you want. A table with the other side's likely ask, your fallback and your walk-away point tells you what you will actually sign.

Treat every number the model picks as a placeholder. When we said "you choose", HAQQ chose prices, liability caps and pilot states, and labelled them as proposals. That's the right behaviour, and it still leaves the founder to replace every one.

Key takeaways

• A health tech pilot agreement for in-home sensing needs a household notice, a consent route for people who may lack capacity, and a claims restriction that binds both parties.
• With a franchisor, contract with headquarters and have each pilot office sign a participation addendum.
• Choose your governing law, then choose a dispute mechanism you can enforce where the other side's assets are.
• Negotiate data and model rights before equity. In US healthcare, equity tied to deployment volume needs a separate legal check.
• Use a legal AI twice: raw material first to find what you missed, a written brief second to get the documents. Then a lawyer.

Appendix A: Ask A, anonymised

Our second message in Ask A (the first was the wrong paste), with the parties removed and the transcript summarised:

[Transcript of a twenty-minute call between an angel investor and the founder:
the market thesis, the product, the raise, early conversations with care providers
in three countries, and the franchisor's "prove it in 100 homes" answer.]

The industry is super interesting, and I would like to test generating some of the
contracts for this industry and maybe think about all the documents that this
founder is going to need.

1. I would love for [them] to also understand maybe the stake, for example, in a
   way that you partner with [PARTNER].
2. [Note that the company's website, deck and founder details will follow.]

First, I would like for you to understand the situation and the market research a
little bit.

[The founder's outreach email, with links to the deck and demo video.]

Appendix B: the brief, exactly as sent

Three messages, sent one at a time. The 6 January date in Message 1 and the "3$" in Message 2 are the errors described above. We have left them in.

Message 1 (context and the NDA). We attached the brief as text and added one line underneath it.

You are acting as counsel to [STARTUP] Pty Ltd, an Australian proprietary company registered in New South Wales. Draft for [STARTUP]: protective, but reasonable enough that a US company's in-house legal team signs without a long redline.

**Facts**
- [STARTUP] is a two-founder, pre-seed company. Its product is a passive in-home sensing system: Wi-Fi channel state information sensors, mmWave radar sensors and an edge hub. It uses no cameras, no microphones and no wearables. Software (mobile app, web app, API/event feed) learns each person's normal routine and flags changes to care staff: routine deviation, night movement, bathroom frequency, mobility changes, possible falls, leaving the home. [STARTUP] keeps title to all hardware.
- Users are older adults living at home, many with cognitive impairment or dementia. Customers are home care providers, who pay per home per month.
- [PARTNER] is a large US home care company (Delaware corporation, headquartered in California). It owns a franchise network of independently owned home care offices. Franchise owners deliver the care and choose their own vendors.
- [PARTNER]'s leadership told [STARTUP]: deploy in 100 homes, then come back. Plan: a 6-month paid pilot in 100 homes across 3 to 5 franchise offices, then a preferred-vendor agreement for the network.
- [STARTUP]'s goals: keep its platform, models, improvements and de-identified aggregate data; no exclusivity (if unavoidable: narrow, time-limited, tied to volume minimums); no right of first refusal or first negotiation on a sale of the company; a paid pilot; permission to publish a case study with consent. Possibly later: an advisor relationship with a [PARTNER] executive and performance-based equity for [PARTNER]. These later items are not part of these documents.

**Constraints that apply to all three documents**
1. **Governing law: New South Wales.** [PARTNER] and its assets are in the US. Recommend the dispute mechanism that gives [STARTUP] an outcome it can enforce in the US: for example, arbitration seated in Sydney under the ACICA Rules, compared with NSW courts or US arbitration. Explain the trade-off in 3 to 5 lines.
2. **Not a medical device, not an alarm.** The product must not be positioned as a medical device, a personal emergency response system, a 24/7 monitoring service or a substitute for care. Include an intended-use clause and a claims restriction binding both parties: no statement that the product diagnoses, predicts, prevents or treats a disease or condition (for example UTIs or falls). Explain why with reference to the US FDA General Wellness and Clinical Decision Support guidance (both revised 6 January 2026) and the Australian TGA software-as-a-medical-device rules.
3. **Privacy.**
   - No health information or PHI is exchanged under the NDA or the LOI.
   - In the pilot, consent comes from the client, or from their legally authorised representative where the client lacks capacity (healthcare proxy, power of attorney, guardian).
   - Other household members and regular visitors get a notice, because the sensors detect everyone present.
   - Address HIPAA (business associate agreement if a franchise office is a covered entity) and state consumer health data laws (for example the Washington My Health My Data Act).
   - Address the Australian Privacy Act 1988, including APP 8 on cross-border disclosure, and whether the small business exemption is available to [STARTUP] given that it handles health information.
4. **Australian law checks.** Say whether the unfair contract terms regime for small business standard form contracts, the Australian Consumer Law and the usual limits on exclusion clauses apply. Say whether they apply when the counterparty is a US company.
5. **Franchise structure.** Recommend how to bind independent franchise offices: a pilot agreement with [PARTNER] plus a short participation addendum each office signs, or another structure. Give your reasoning.

**Output format for each document**
1. Founder summary: two short paragraphs in plain English.
2. The full clean draft, with numbered clauses and placeholders in [square brackets].
3. Negotiation table with columns: clause | our position | what [PARTNER]'s counsel will likely ask | acceptable fallback | walk-away.
4. Open questions for the founder.
5. Authorities relied on (legislation sections, regulator guidance), with links where available. Mark anything you are not certain of. Do not invent citations.

Ask clarifying questions only if a gap would change the draft materially; otherwise state your assumptions and proceed.

**Now draft Document 1: a mutual NDA.**
- Term: 2 years, with confidentiality obligations surviving 3 years after expiry; trade secrets protected for as long as they remain trade secrets.
- Standard exclusions.
- No residuals clause.
- No reverse engineering of hardware, firmware, software or models.
- No use of confidential information to build a competing product.
- Feedback is free to use but grants no licence to the other party's IP.
- An express statement that no health information or PHI will be disclosed under the NDA.
- Return or destroy, with a carve-out for automatic backups.
- No obligation to enter any further agreement.
- Injunctive relief.
- NSW law plus your recommended dispute mechanism.

do what this text says ;)

Message 2 (letter of intent).

Using the same facts and constraints, draft **Document 2: a letter of intent** for the pilot and the path to a network agreement.
- **Binding** only on: confidentiality (by reference to the NDA), each party bears its own costs, no exclusivity, governing law and dispute mechanism, and term/termination of the LOI. Everything else is expressly non-binding.
- **Content:**
  - Pilot scope: 100 homes, 3 to 5 franchise offices, 6 months.
  - Roles: [PARTNER] selects and introduces offices and sponsors integration with its care platform. [STARTUP] supplies, installs and supports.
  - Indicative pilot fee of [AMOUNT] per home per month and indicative post-pilot pricing of 3$.
  - Success criteria, agreed in writing before launch.
  - Timeline with milestones.
  - Data principles, consistent with the pilot agreement to follow.
  - Path to a preferred-vendor agreement if the success criteria are met.
- **One non-binding sentence** stating the parties may separately discuss an advisory relationship and [PARTNER]'s participation in a future financing or performance-based equity. No commitment, no right of first refusal, no right of first negotiation.
- Keep it to about two pages.

Message 3 (pilot agreement). HAQQ replied with three questions; our answer was "you choose".

Using the same facts and constraints, draft **Document 3: the pilot agreement**, between [STARTUP] and [PARTNER], with a franchise office participation addendum as a schedule (or your recommended alternative). Cover at least:

1. Definitions.
2. Pilot scope and services.
3. **Hardware loan:** title stays with [STARTUP], risk of loss, installation, return at the end. Warranty that devices hold the radio certifications needed for the US (FCC Part 15).
4. Software licence for the pilot term.
5. Fees and invoicing.
6. Intended use and the claims restriction.
7. **Consent and notices:** client or authorised representative consent; household and visitor notice; templates as schedules.
8. **Data:**
   - Client data belongs to the client and the franchise office.
   - [STARTUP] owns its platform, models, improvements and de-identified aggregate data.
   - [STARTUP] gets a licence to use de-identified data to improve the product and train models, subject to law and a stated de-identification standard.
   - No sale of personal information.
9. Privacy and security, with a BAA and a data processing schedule.
10. Success criteria and a joint evaluation report.
11. Publicity and case study, only with written consent.
12. IP and feedback; no joint IP by default.
13. **Warranties and disclaimers**, including "not an emergency alarm and not a substitute for care or clinical judgment".
14. Limitation of liability: a cap equal to fees paid, or a fixed amount if greater, with standard carve-outs.
15. Mutual indemnities: [STARTUP] for IP infringement; [PARTNER] and the offices for failing to obtain consent and for misuse.
16. Insurance: product liability, cyber, professional indemnity.
17. Term and termination.
18. **Conversion:** option to move to commercial terms at pre-agreed pricing; this agreement is superseded by the commercial agreement.
19. Mutual non-solicitation of staff.
20. No exclusivity.
21. Governing law and disputes.

Then add a one-page cover note for the founder: the five clauses most likely to be contested, and what to concede first.

Sources and further reading

• US FDA General Wellness (6 January 2026) and Clinical Decision Support (reissued 29 January 2026) guidance, summarised by Covington
• TGA: Understanding if changes to software-based medical device regulation affect you
• HHS: Business associates and cloud services under HIPAA
• UNCITRAL: New York Convention on the Recognition and Enforcement of Foreign Arbitral Awards
• AI contract drafting guide
• Writing a matter brief for AI contract review
• How to anonymize a document before you give it to AI