Legitimate interests assessment

Version lia-2026-10-23, in effect from 23 October 2026.

Controller	Global Phoenix Solutions Ltd, company number 14855822, registered in England and Wales
Registered office	Ground Floor Flat, 259 Chichester Road, Bognor Regis, West Sussex, PO21 5AH, United Kingdom
Assessment by	Joshua Sandhu, director
Date of assessment	18 September 2026, reviewed and adopted 24 September 2026, amended 12 October 2026 for the contacts our software collects by itself, 13 October 2026 for the follow-up messages, the choice between fixed wordings of the first message and the alerts to the director, 14 October 2026 for the volume caps, the reach of the reading of company websites, the small-business research award source and the country of a company, and 23 October 2026 for the wordings of the first message and of step two (see Changes)
Next review	24 December 2026, or sooner if the purpose, the data, the sources or the sending method changes
Processing assessed	Sending business-to-business email about ProofPack to named people at companies that hold a public regulatory record for an AI medical device, and, from 12 October 2026, collecting further such contacts daily by software from the FDA’s 510(k) summaries, and, from 14 October 2026, sending the same email to named people and shared inboxes at United States companies that hold a federal small-business research award for an AI medical product, found from the NIH and NSF public award databases
Lawful basis relied on	UK GDPR Article 6(1)(f), legitimate interests
Privacy notice for this processing	Outreach privacy notice
Contact	Joshua Sandhu, joshua.sandhu@globalphoenix.co.uk

This assessment follows the three-part test the ICO describes: the purpose test, the necessity test and the balancing test. We publish it in full rather than summarising it, so the people it concerns can read what we weighed.

Where things stand on 14 October 2026. We have sent outreach email since 12 October 2026: the first sending run was at 10:30 that day, and first messages have gone out in the sending runs since. On 8 October 2026 we collected the contact details we write to: 71 work email addresses at 53 companies, 27 of them a named person’s and 44 a company’s general inbox. Each was read from the company’s own website or from a 510(k) summary the FDA publishes (30 of the 71, including all 27 named people, come from those summaries), and the web address of the page it was read from is kept beside it. The director did this with Claude, Anthropic’s AI assistant, doing the reading for him. We also hold 8 addresses we set aside the same day and will not write to: 4 failed our checks, and 4 belong to two Danish companies that ask not to receive marketing (see “Other laws”).

These details are on the director’s computer, which is also the computer that runs our sending software: in the files of the list written on 8 October 2026, in copies of the public pages they were read from, in working copies (one of them a trial copy of our sending software’s database that holds 22 of the contacts and from which no message has been sent), in the outreach database described below, and in the records the AI assistant keeps of the sessions in which the work was done. One named person’s address and one general inbox, both at one of the two Danish companies, were also in our sending software’s code repository, which GitHub holds; on 10 October 2026 we removed them from it and from its history, and GitHub still holds the earlier versions until it deletes them. The outreach privacy notice gives the detail under “Where we keep it”. The outreach database our sending software sends from exists: the director created it on 10 October 2026 by running the software’s setup, which loaded 22 of the contacts into it and added 7 rows to its suppression list for the two Danish companies; it is a file on the same computer, and no message has been sent from it. The software loads only a contact that has a first name and is marked as working at a body corporate, together with the https web address of the record that shows it; it checks that the web address is there and is https, not what the page says. So the 44 general inboxes and 5 of the named people were not loaded, and 22 addresses were. Of the 71, 55 are marked as working at a body corporate, and each of their web addresses is a page on a public authority’s website, such as a company register, read when the list was prepared.

The safeguards below are in two groups: those our sending software already applies, and ten we set ourselves before the first message, each with where it stands on 10 October 2026. We do not send any outreach email until each of the ten is in place.

From 12 October 2026 the list also grows by itself. From 12 October 2026 our software also finds new contacts by itself, once a day, without a person or an AI assistant reading the page. For a company on one of the two FDA lists whose newest 510(k) clearance was decided in the last twelve months, the software reads the 510(k) summary the FDA publishes, which prints a named contact person and a work address, and stores that contact with the web address of the PDF as the source. It keeps at most one contact per company, and never a submission consultant: it keeps only an address on the company’s own web domain. The contact goes through the same checks as every other: a free-mail address is refused, Germany and Austria are refused, a Danish company is held back until the director has checked the Danish register by hand, and the software treats the employer as a body corporate in only two cases. In the first, the employer is a United Kingdom company with an active entry at Companies House, and the software stores the web address of that entry as the evidence. In the second, the applicant’s name on the FDA’s 510(k) record for the clearance ends in a recognised legal form, such as Inc., LLC, Corp., Ltd., GmbH, S.A., B.V. or A/S, and the software stores the web address of that FDA record as the evidence. The second is a reading of the name the applicant gave the FDA and not of a company register, and the software does not check that the company named is still trading. For any other company, including one whose applicant name ends in no such legal form, the status of the employer is not established: the software stores no name or address of a contact there, records only the company, the FDA record number, the country, the role group and the email domain, and writes to no one. The software takes an address only when it appears in full in the text of the PDF, and drops one that a line break, a hyphen or an apostrophe has cut. The director’s collection of 8 October 2026 is described above and was done differently. Today the software runs on the director’s computer. We intend to move it to scheduled jobs on GitHub and the outreach database to Cloudflare’s D1 service, and before either move we change this assessment and the outreach privacy notice. The first sending run was at 10:30 on 12 October 2026, and the volume caps below apply to the contacts found this way as to every other. From 13 October 2026 the software also reads a company’s own website, by itself, for a shared company inbox such as info@ or regulatory@, and never for a person’s address. Since 14 October 2026 it does this for every company on the two FDA lists whose newest clearance was decided in the last two years and for which we hold no contact, unless our exclusion rules rule the company out. That includes a company whose 510(k) summary the software does not read, because its newest clearance is more than twelve months old or because it has only one entry on the two lists; the safeguard “Limits on the collection by software” says how. The twelve months in the paragraph above applies to a named person read from a summary, who may have left the company, and not to a shared inbox, which does not go stale in that way. From 14 October 2026 the software also finds companies to write to in the National Institutes of Health’s RePORTER database and the National Science Foundation’s Awards database, which list the small-business research awards (SBIR and STTR) the two agencies have made. Once a day it reads the awards of the last two years to companies in the United States and keeps an award only if its title or abstract names artificial intelligence, machine learning or a related term (an index term in NIH’s record is not enough), the award’s title, abstract or index terms name a medical product such as a diagnostic, a device, imaging, screening or monitoring, its title or abstract names a clinical setting, and its title or abstract does not describe drug discovery or development, a therapeutic, a vaccine, an antibody, gene or cell therapy, agriculture, veterinary or defence work. The contact for a company known from an award is the NSF award’s principal investigator or a shared inbox from the company’s own website, and the safeguard “Limits on the collection by software” sets out the limits. The volume caps below apply to these contacts as to every other. From 14 October 2026 the software also establishes a company’s country from the FDA’s own records when the FDA’s list of AI-enabled medical devices gives none; “The nature of the personal data” says how.

Part 1: Purpose test

What we want to do. Send a short email, and up to three follow-ups, to a regulatory-affairs, engineering or leadership contact at a company that has recently appeared in a public regulatory record for an AI medical device or holds a federal small-business research award to build one, telling them that ProofPack exists and what it computes.

Why. ProofPack is software that computes the validation statistics a regulatory submission for an AI medical device needs: subgroup performance with confidence intervals and calibration. The people who need it are a small, identifiable group: the regulatory and clinical-validation leads at companies making such devices. There is no trade directory of them and no marketplace where they look for tools of this kind. A direct, factual email is the way to reach them.

The benefit to us. Sales. Global Phoenix Solutions Ltd is a one-person company with no external funding. Customers are how it continues to exist.

The benefit to the recipient. A recipient who is preparing a submission learns of a tool that computes, with confidence intervals and a stated method, the tables their submission needs. A recipient whose company is building an AI medical product learns, before a first submission, of a tool that computes the performance evidence a submission asks for. The email says plainly that the award record is all we know about the company and does not assume it plans a submission. A recipient who has no such need reads one short email and can stop the rest from the stop link in its footer.

Third parties and the wider public. No third party benefits from the processing. There may be an indirect public benefit if better-evidenced device submissions result, but we do not rely on it here. The interest we rely on is our own commercial interest. Recital 47 of the UK GDPR says that processing personal data for direct marketing purposes may be regarded as carried out for a legitimate interest.

Impact if we could not do this. We would have no way to tell the people who need the tool that it exists, other than waiting for them to find the website.

Other laws. The Privacy and Electronic Communications Regulations 2003 (PECR), regulation 22, restrict marketing email to individual subscribers. We send only to corporate subscribers. Our software checks every contact against our exclusion rules when it is loaded, and again before any message to it is built, and excludes: an address at one of the free-mail providers on its list (16 domains, such as gmail.com, outlook.com and yahoo.com); a contact whose country is not a recognised two-letter country code; a contact in Germany or Austria, or at a .de or .at address; and a contact that is not marked as working at a body corporate together with the https web address of the record that shows it, so an organisation whose status we have not established is excluded. The software checks that the web address is there and is https, not what the page says; we read the record when we prepare the list. For a contact the software finds by itself, the record is the FDA’s 510(k) record of the clearance, when the applicant’s name on it ends in a recognised legal form, or the company’s entry at Companies House, as set out under “Where things stand”. For a company known from an award, the record is the award record, when the organisation’s name on it ends in a recognised legal form. A contact’s country is the country recorded for the company: from the FDA’s list, from the FDA’s own records as “The nature of the personal data” sets out, or the United States for a company known only from an award, because the software keeps only awards to organisations in the United States. A company whose country stays unknown is excluded. Recipients in the United States receive the same message and footer; the footer identifies the message as a business-to-business commercial message from us, gives our postal address and a stop link (item 3 below says what must be shown before the first message is sent), and we honour an opt-out on receipt, which is sooner than the 10 business days the FTC’s CAN-SPAM guide sets. We do not send to anyone in Germany or Austria under any basis. We also do not write to a Danish company that the Danish company register (CVR) marks as not wishing to receive marketing (reklamebeskyttelse). Two companies on our 8 October list are marked so: we took both off the list that day, and the setup the director ran on 10 October 2026 put them on our software’s suppression list, 7 rows in all. Our software does not read the Danish register itself, so before we add any other Danish company we check the register by hand.

Ethical issues. We do not think there are any beyond those the balancing test addresses. We do not use any data about a person’s private life, and we do not infer anything about them beyond their professional role.

Outcome of the purpose test. The purpose is specific and is a legitimate interest.

Part 2: Necessity test

Whether the processing helps achieve the purpose. It does. Reaching a named person in the relevant role at a named company is the only form of outreach that reaches the small group described above.

Whether it is proportionate. We think it is, because of how little is processed and for how short a time. The data is a name, a job title, a work email address, the company and the public regulatory record that put the company on the list. For a company known from an award, the public record is the award record; the contact is the NSF award’s principal investigator, whose business email address the NSF publishes on the award, or a shared inbox on the company’s own website. We considered writing only to shared inboxes, which would leave out the person the NSF names; we keep the named person because the NSF itself publishes the address with the award, because we take it only on the company’s own domain, and because we keep at most one contact for a company. NIH publishes no address, so no named person comes from NIH. The contact sequence is at most four messages, timed for days 0, 5, 11 and 18 after the first, and then it stops. From 13 October 2026 the software sends all four steps, and sends none of the three follow-ups to a person who has replied, asked us to stop or been paused (item 9 below).

Less data, or a less intrusive way. We considered:

• Waiting for inbound enquiries only. The website is public and indexed. This is not a reasonable alternative on its own: the people who need the tool do not know it exists, and a one-person company cannot rely on being found.
• Advertising. It would reach a far larger number of people who have no interest in the product, and would be less targeted rather than more; it also costs money the company does not spend.
• Contacting a general company inbox instead of a named person. Where a 510(k) summary names a regulatory contact for the company, we write to that person, because writing to the general inbox reaches the same person by a longer route with a higher chance of being missed. Until 12 October 2026 our software wrote to no general inbox, because each message opened with the recipient’s first name. Since 13 October 2026 it also writes to a company’s shared inbox, such as info@ or regulatory@, with a message that opens “Hello,”, when it holds no named contact for that company, and it never writes first to both at one company.
• Having the director read the pages by hand each time. This is how the 8 October list was made. It cannot be repeated daily by a one-person company, so from 12 October 2026 software does it, within limits: for a named person two sources (the FDA’s own summaries and, for a company known from an award, the NSF’s award record), and for a shared inbox the company’s own website; companies on the two FDA lists and, from 14 October 2026, companies that hold a federal small-business research award, one contact per company, no consultant, and an address only on the company’s own domain.
• Buying a marketing list. We do not. Every address we hold was read from a page or public record we fetched ourselves, and the web address of that page is stored beside it. No address is guessed from a name pattern. Our software refuses to store any contact without the web address it was read from (item 6 below).
• Telephone. More intrusive than one email that can be stopped from the link in it.

Outcome of the necessity test. The processing is necessary for the purpose and there is no reasonable, less intrusive way of achieving it.

Part 3: Balancing test

The nature of the personal data

• Name, job title, work email address, employer and the professional context in which the employer appears in a public regulatory record or holds a federal small-business research award.
• No special category data. No criminal offence data. No financial data. No data about children. No data about people at risk of harm.
• The data concerns people in their professional capacity only.
• It comes from sources that are public by design. Two FDA sources tell us which companies to write to: the FDA’s 510(k) database, read through openFDA, and the FDA’s list of AI-enabled medical devices. A 510(k) record in the database names the company and its device, and also names the person the record gives as the company’s contact for that submission; our collector reads that name with the rest of the record and does not store it. The contact details come from two places: the 510(k) summary the FDA publishes for a device, which names a contact person for the submission with a work address, and the company’s own website, where the company chose to publish the contact. From 12 October 2026 our software also reads those 510(k) summaries by itself, daily, for the clearances decided in the last twelve months of companies on the two FDA lists, and takes from each at most one contact. From 13 October 2026 the software also reads a company’s own website for a shared company inbox, and since 14 October 2026 it does so for every company on the two FDA lists whose newest clearance was decided in the last two years and for which we hold no contact, unless our exclusion rules rule the company out, including a company whose summary gave no contact it could keep and a company whose summary it does not read. It uses the FDA’s medical device identifier database (GUDID, read through openFDA) only to find the company’s web domain, storing no address from it. We check each company’s legal status in official company registers, such as Companies House in the United Kingdom and the Danish register (CVR), and for some US companies in a state’s business records or the company’s filings with the US Securities and Exchange Commission. For the contacts the software finds by itself, it reads a company’s legal status from the legal form that ends the applicant’s name on the FDA’s record and, for a United Kingdom company, at Companies House first. We collect from no MHRA or EUDAMED record. From 14 October 2026 two further public records tell us which companies to write to: the National Institutes of Health’s RePORTER database and the National Science Foundation’s Awards database, which list the small-business research awards (SBIR and STTR) the two agencies have made. For each award it keeps, the software records the company’s name, city and state, the award’s number, title, dates and public web address, and the words that matched. The NSF’s record also publishes the email address of an award’s principal investigator, which the software keeps only when it is on the company’s own web domain and is not a free-mail provider’s, a university’s or a government’s; the NIH’s record publishes investigators’ names and no email address, so the software stores no investigator from NIH. The contact details of a company known from an award come from the NSF’s award record or from the company’s own website. From 14 October 2026 the software also establishes a company’s country when the FDA’s list of AI-enabled medical devices gives none, from the FDA’s own records read through openFDA: the 510(k) record of the company’s clearance; the FDA’s premarket approval (PMA) record, where a US state gives the United States; or the FDA’s establishment registration record, when every establishment registered under the company’s name gives the same country. When the records disagree, or a record under the same company name shows Germany, Austria, Denmark, Switzerland or Liechtenstein beside another country, the country stays unknown and the company is excluded. The software stores the country with the web address of the openFDA record it was read from, never replaces a country already recorded, and stores no person’s details for this.

Reasonable expectations

We have no existing relationship with the recipient. We did not collect the data from them directly. Against that:

• The recipient’s employer has entered a public regulatory process for an AI medical device, and the recipient’s name and role are published for professional contact, by their employer or in a 510(k) summary their employer submitted and the FDA publishes.
• For a company known from an award: the recipient’s employer holds a federal small-business research award for an AI medical product, in a record the NIH or the NSF publishes. A principal investigator who lists a business email address on a federal award expects it to be read by those who read the award record; the email we send is about evidence for the product the award describes, is addressed to a business role, and carries the stop link and the reason for receiving it. The award record does not show that the company intends a submission; the email does not say it does.
• Business-to-business email about a product relevant to the recipient’s published role is a widely understood practice, and the ICO’s PECR guidance says marketing email may be sent to corporate subscribers.
• The first message says, in its footer, where the data came from and why the person is receiving it, and links to the outreach privacy notice and to a stop link.

We conclude that a reasonable person in the recipient’s position would expect that a vendor of validation tooling might write to them, and follow up at most three times in the same thread if they did not reply, would expect to be told why, and would expect to be able to stop it immediately.

Likely impact

The impact on the recipient is one unsolicited email, and up to three short follow-ups in the same thread if they do not reply. There is no profiling, no automated decision with any effect on the person, no sharing with any third party for their own purposes, and no use of the data beyond sending those messages and recording that they were sent. Which of a few fixed wordings of the first message a person reads is chosen from counts for each wording as a whole and not from anything about them (see “Choosing between fixed wordings” below). The likely harm is inconvenience.

Safeguards our sending software already applies

• Volume caps. At most 20 first-contact emails a day, and at most 50 a day of the messages the outreach software counts, which are its first contacts, its follow-ups and the replies sent through it. Email sent from the same mailbox in any other way, by hand or by another tool, is not in that count. While the mailbox is new, the software sends no outreach in the first seven days and at most 10 first contacts a day in days eight to fourteen. Transactional email about a purchase (licence files and quotes) is sent through a different service and is not counted against these caps.
• A short, finite sequence. The software has four message steps and refuses a fifth. Steps two to four are replies in the same thread, and the last says that we will stop there. Item 9 below says how the software times the steps and when it stops a sequence.
• Limits on the collection by software. From 12 October 2026 the software that finds named contacts by itself takes them from the 510(k) summaries the FDA publishes for companies on the two FDA lists, and only for a company whose newest clearance was decided in the last twelve months, and, from 14 October 2026, one named contact from an NSF award record as the last part of this item sets out; keeps at most one contact per company; keeps an address only if it is on the company’s own web domain, so never a submission consultant at a consultancy; takes an address only when it appears in full in the text of the PDF, and drops one that a line break, a hyphen or an apostrophe has cut; stores the web address of the PDF as the source and refuses a contact without it; applies the free-mail, Germany and Austria and body-corporate checks described under “Other laws” and “Where things stand” before the contact can be written to, so that the employer counts as a body corporate only through an active Companies House entry for a United Kingdom company or a recognised legal form ending the applicant’s name on the FDA’s record or, for a company known from an award, the organisation’s name in the award record, whose web address is stored as the evidence; and holds a Danish company back for the manual register check. A company held back, whether Danish or one whose status is not established, is recorded without the name or address of any contact there and receives no email. The daily report the software writes for the director lists companies, FDA record numbers, role groups, job titles and email domains and no person’s name or address, and it removes any part of the contact’s name found inside a job title. The software does not check that the person named in a summary still works at the company. From 13 October 2026 the software also reads a shared company inbox from a company’s own website, and since 14 October 2026 it does so for every company on the two FDA lists whose newest clearance was decided in the last two years and for which we hold no contact, unless our exclusion rules rule the company out. That includes a company whose summary gave no contact it could keep, because it names no one, names a consultant, or gives a job title that fits none of our role groups, and a company whose summary the software does not read, because its newest clearance is more than twelve months old or because it has only one entry on the two lists. The twelve months above applies to a named person read from a summary, who may have left the company, and not to a shared inbox, which does not go stale in that way. The other limits above apply to a shared inbox as they do to a contact from a summary. The software keeps only a shared inbox, such as info@ or regulatory@, and never a person’s address or an inbox for sales, careers, press, investors or privacy. It finds the site from the web domain of the customer contact address the company gives in the FDA’s medical device identifier database (GUDID, read through openFDA), from a web address or email domain printed in the summary that is not a consultant’s, or, only if neither gives one, from the company’s name; in each case only a domain that resembles the company’s name. It stores no address from the identifier database. It treats the site as the company’s only when a page on it shows the company’s name together with the FDA record number, the device’s name or, for a domain that did not come from the company’s name, the city on the FDA record. It reads at most six pages of the site, obeys the site’s robots.txt, identifies itself as ProofPack with the web address of the outreach privacy notice, keeps the address only on the company’s own domain or its www host, stores the web address of the page it was read on as the source, tries a company’s website at most once in 30 days, and tries the websites of at most 60 companies a day. A shared inbox is stored with no name, and the first message to it opens “Hello,”. Companies known from an award. From 14 October 2026, for a company that holds a federal small-business research award and is on neither FDA list, the software finds a contact in one of two ways, and keeps at most one contact for a company. The NSF’s award record publishes the email address of the award’s principal investigator. The software keeps it only when it is on the company’s own web domain (not a free-mail provider, a university or a government domain, and a domain that matches the company’s name); it then stores the person’s first name, the address and the web address of the award page as the source, records the role as principal investigator, puts the person in the engineering role group and writes to them as the company’s technical lead. NIH publishes no email address, and no named person comes from NIH. Otherwise the software reads the company’s own website for a shared company inbox. It finds the website from the domain of the principal investigator’s address, where that address passed the checks above, from a web address printed in the award’s abstract that is not a university, government or platform site, or, only if neither gives one, from a web address guessed from the company’s name. It guesses only from the company’s distinctive name words run together, when there are two or more of them, or from its whole name run together, never from the first word of a name alone, and from a name of one word only if it has at least seven letters. It accepts a website only if a page shows the company’s name together with the award’s city or state or, for a guessed web address, together with a phrase of the award’s title or the award’s number; for a guessed address it does not accept the city and the state together. The other limits on reading a website apply as for a company on an FDA list: at most six pages, robots.txt obeyed, an address only on the company’s own domain or its www host, a company tried at most once in 30 days, and at most 60 companies a day, a number shared with the companies on the FDA lists. The software does not check that a principal investigator still works at the company; it reads only the awards of the last two years. Its daily report lists companies, domains, award numbers and reasons, and no address or name. The volume caps above, at most 20 first-contact emails and 50 messages a day, apply to these contacts as to every other.
• Suppression wins over everything. When the software assembles an email for sending, it checks the address, its domain, the company and its corporate group against the suppression list, and a match on any of them stops that email. There is no override. The sending connection (item 1 below) sends only emails assembled this way.
• No re-approach. A company is not approached again within 90 days of any contact, or within 180 days of a completed sequence that drew no reply.
• Identity in every message. When the software assembles an email for sending, it refuses one that does not end in the fixed footer carrying that email’s own stop link. The footer names the sender, the company, its registration number and registered office, why the recipient is receiving it, the source of their data, a link to the outreach privacy notice and the stop link.
• A stop link and unsubscribe headers in every message. Every message carries the standard one-click unsubscribe headers (RFC 8058) and a stop link in the footer. The link is signed for the one address it was sent to and needs no login. The footer link opens a page on this website with one button: opening the page records nothing, and pressing the button asks for the record. A mail client that supports the headers can send the request itself, without opening the page. The code that answers the web address the link points to has been in the source of this website since 2 October 2026. While its signing secret is unset it answers an error; while this website is not connected to our database, it answers an error; while the table it writes is missing from our database, its button answers an error. In each case it records nothing. On 10 October 2026 it recorded a request made with a real link, one our software had signed for an address of our own, and our software read the record back from this website (item 3 below). Our sending software reads the stop-link records from this website before each sending run and does not send to an address recorded there; if it cannot read them, it does not send.
• Opt-outs recognised in replies. The part of the software that reads replies treats a reply of “unsubscribe”, “Opt out please”, “Stop emailing me”, “Please delete my details”, “No thanks.” or “I object.” as an opt-out and adds the address to the suppression list permanently. It does not recognise every wording: “I object to this.” and “Not interested, thanks” are not treated as opt-outs. A person therefore reads every reply and adds the address to the suppression list by hand when a reply objects in words the software did not recognise, and once someone has replied we send no further automated message in that thread. A reply written by someone else at the company in a thread of a message we sent to a colleague is treated the same way: if it asks to be left alone, the software adds the sender’s address to the suppression list and stops the sequence for the person we wrote to, and tells the director. A person can add a single address to the suppression list by hand. The reply reader reads the outreach mailbox through the connection in item 1 below, on the schedule in item 4 below, and item 4 says which messages it reads.
• No invented facts. The one sentence in each message about the recipient’s company is read word for word from the stored public record. No part of the software accepts free text for it, and no language model writes any part of a message.
• Choosing between fixed wordings. From 13 October 2026 the first message exists in three fixed wordings for each role group (regulatory, engineering or leadership), written in advance and not at the time of sending, any of which the director can remove at any time. From 23 October 2026 the three wordings for a role group differ only in the subject and have the same body, which says the same facts, with the same links, the same footer and the same sentence read from the public record; until then they differed in the subject and the opening paragraph. From 23 October 2026 every wording also says that, for a small number of teams, the first full pack is free, in return for the team’s regulatory affairs lead reviewing the pack wording and completing the installation and operational qualification (IQ/OQ) record, which is the offer the pricing page of this website calls “First full pack free”. A first message to a shared inbox opens “Hello,”, and from 23 October 2026 it carries one added line asking the reader to pass the message to their regulatory lead if it is theirs. The software chooses a wording for each new first message by a fixed rule. It counts, for each role group and wording, how many contacts were sent that wording and how many replied, were interested, asked us to stop or bounced; from 23 October 2026 it counts only first messages sent with the wordings of that date, and the counts for the earlier wordings are not used. Until each wording has been sent to 20 contacts in a role group, the least used one goes next. After that, a wording leads only if it has at least three interested replies, meaning replies that show interest or ask a question from a person who did not also ask us to stop, and at least two more of them than another wording would have had at the leader’s number of sends; an opt-out, a negative reply or a neutral one does not count towards a lead, and otherwise the first of the three wordings leads. Eight in ten first messages go to the leading wording and two in ten are spread over the others, so that the counts keep being compared. A wording whose subject would not be between four and eight words with the device’s name in it is not used for that contact. The counts are totals kept in the same database file and carry no name or address; a reply is counted in the total for the wording its sender was sent. The choice depends on the role group the person is in and on those totals, not on anything else about the person; it has no effect on the person beyond which of the fixed wordings of the same message they read, and it is not profiling and not an automated decision with any effect on the person. For a company known from an award there are three further fixed wordings for each role group, which from 23 October 2026 likewise differ only in the subject; they say that the award record is what we know about the company, what ProofPack does and, from 23 October 2026, the offer of a free first full pack on the terms above, and they do not say or imply that the company has or plans a submission; the software chooses between them by the same rule and counts them apart from the first three. No language model is involved; the software uses no new source of data.
• Alerts and summaries go only to the director. The software emails the director, from the outreach mailbox to the director’s own address, joshua.sandhu@globalphoenix.co.uk, and to no other address, because the recipient is fixed in the software and it refuses any other. It sends one alert for each Gmail thread in which a person replies with something that is not an out-of-office reply, a bounce or a stop request from the person we wrote to: the company, the role group, the step replied to, the date, a link to the Gmail thread the reply is in and the first 60 words of the reply with quoted earlier messages removed, which can include what the sender wrote about themselves. When someone other than the person we wrote to, such as a colleague the message was forwarded to, writes in a thread of that person’s sequence and asks to be left alone, it also sends one alert for that thread saying so, with the same fields and the first 60 words of the message. No alert includes the sender’s name or address. It sends an alert when a scheduled task fails or the daily check at 07:00 finds a problem, and, at 08:00 on Mondays, a summary of totals for the last seven days and to date, the counts for each wording and the number of messages due in the next seven days, with no name or address. Alerts are not counted in the volume caps. Google, the provider of the mailbox, is the only other party involved.
• Records. Every message is stored with the version of the template it used, the wording variant when it is a first message, and a fingerprint of the footer it carried, and every opt-out is stored with its time and how it arrived.
• The send switch is off by default. Sending is refused unless it is switched on explicitly for that run; without it the software only prepares messages for a person to read. The weekday task in item 4 below, which first runs on 12 October 2026, switches it on for its own run, and only after that run’s reading of replies has succeeded.

Ten safeguards we set before the first message, and where each stands on 10 October 2026

We do not send any outreach email until each of these is in place. Items 1, 2, 3, 5, 6, 8 and 9 are in place in our software. Item 10 is in place with one figure typed in by hand each day, and item 7 is done by hand. Item 4 is in place: the director registered the scheduled tasks it describes on 10 October 2026, and the weekday task first runs on 12 October 2026.

1. The sending connection. In place since 8 October 2026, when the director signed the outreach software in to the outreach mailbox, joshua.sandhu@globalphoenix.co.uk, through Google’s Gmail interface. The software refuses to send from any other mailbox. It hands the mailbox only emails that have passed the assembly step described above, with its suppression and footer checks, and only on a sending run the check in item 2 has allowed.
2. A live check of this page. In place. Before each sending run, the software fetches this assessment and the outreach privacy notice from this website and refuses to send if either cannot be fetched or is not marked live.
3. The unsubscribe endpoint. In place since 10 October 2026. Since 8 October 2026 the software reads the stop-link records from this website before each sending run, and does not send when it cannot, as the stop-link item above says. On 10 October 2026 we sent this website the one-click request that a mail client sends from a message’s unsubscribe headers, using a stop link our software had signed for an address of our own; this website recorded it, and our software’s next reading of the stop-link records returned it.
4. Reading replies. In place since 10 October 2026, when the director’s setup registered two scheduled tasks on the computer that runs our sending software. One runs the reply reader every hour. The other runs at 10:30 each weekday, first on 12 October 2026: it runs the reply reader once more and starts the sending run only if that reading succeeded. Both run only while the director is logged on to that computer and it is awake. The software reads the outreach mailbox’s Inbox and Spam folders and covers every message that arrived since its last successful reading, going back at least 26 hours, so a reply that arrives while the computer is off is read at the next reading and a reply that Gmail files as spam is read as well. It reads a message in full only when the message is in a thread our software sent into, or comes from an address in the outreach database, or has the subject “unsubscribe”, which is the subject a mail client uses for the unsubscribe address in the headers of our emails. A message with that subject is recorded as a request to stop for the address it came from. Because a mail client that offers the unsubscribe address in the headers sends a new message rather than a reply, that subject rule is how such a request is recorded.
5. The corporate-subscriber rules on every contact. In place. The exclusions described under “Other laws” are applied to every contact when it is loaded and again before any message to it is built, with an organisation whose status is unknown treated as excluded. The free-mail rule covers the 16 domains on the software’s list, not every free-mail provider.
6. The source of every address. In place since 8 October 2026. The software refuses to store a contact without the https web address it was read from, and the database file the setup created on 10 October 2026 refuses such a contact as well.
7. Deletion at 12 months. By hand. Contact data is deleted 12 months after the last activity on the record unless the company becomes a customer. The software records the date of the last activity on each contact (the day it was loaded, each message sent to it and each reply from it) and lists the contacts that are 12 months or more past it; it deletes nothing itself. The director runs that list at least once a month and deletes the contacts on it by hand.
8. The footer’s sources. In place since 8 October 2026. From 14 October 2026 the footer has two forms, and in both it names only the public records we collect from. In an email to a company on an FDA list it names the two public records we collect from for that list, the FDA’s 510(k) database read through openFDA and the FDA’s list of AI-enabled medical devices, and says that the recipient’s work address is published on their company’s website or in a public regulatory record. It no longer names MHRA or EUDAMED records. In an email to a company known from an award it reads: “Why you are receiving this: your company holds a US federal small-business research award (the public NIH RePORTER or NSF Awards records), and your work address is published on your company website or in a public award record.”
9. The spacing of the sequence. In place since 10 October 2026, and from 13 October 2026 the software sends all four steps. It times them for days 0, 5, 11 and 18 after the first message, and a follow-up more than three days past its day is not sent and ends the sequence. Steps two to four are sent as replies in the same thread as the first message, at most three of them, and the last says that we stop there. From 23 October 2026 step two links to the public sample pack and not to the demo, and states the offer of a free first full pack described under “Choosing between fixed wordings”; until then it linked to the public sample pack and the public demo. A contact whose first message went out before 23 October 2026 and whose step two is sent on or after that date reads the offer stated in full in step two rather than as a reminder. Step three describes how the tool compares two versions of a model, and step four says that it is the last message, points to the demo and to the stop link; none of them uses a further item of data about the recipient. The software sends no follow-up to a person after a reply, a stop request, a bounce, a suppression or a hold on the contact, and none while a deliverability pause (item 10) is on. A follow-up counts against the 50 messages a day in the volume caps above, not against the 20 first contacts, and when the 50 are short the follow-ups are sent before new first contacts.
10. Deliverability pauses. In place since 10 October 2026, with one figure typed in by hand. Before each sending run the software decides from the sending domain’s own figures whether to pause, and a pause stops first contacts and follow-ups alike. It pauses when the spam-complaint rate Google Postmaster Tools reports for the domain reaches 0.2 percent on a day, or when bounces reach 3 percent of the messages sent over the most recent days, today included, that together hold at least 100 messages. The bounces are counted from the software’s own records of what it sent and the bounce reports its reply reader records. The spam-complaint rate is not read automatically: the director reads it in Postmaster Tools each day and types it in. A day with no figure counts as not reported, which neither starts nor ends a pause, and a pause ends only after three consecutive days with a reported rate below 0.1 percent.

Weighing it up

For the processing: a specific, lawful commercial purpose; a small and precisely targeted group; professional data from public sources; a short sequence with hard daily caps; and a stop mechanism our software reads before every sending run.

Against it: the recipient did not ask to hear from us, and the first message is unsolicited.

We conclude that the recipient’s interests, rights and freedoms do not override our legitimate interest, because the impact is limited to receiving a small number of short, clearly identified business emails that can be stopped from the stop link in each of them, and because no message is sent until each of the ten safeguards listed above is in place.

Decision

Legitimate interests, Article 6(1)(f), applies to this processing. A data protection impact assessment is not required: the processing is not likely to result in a high risk to anyone’s rights and freedoms, on the ICO’s screening criteria.

We will review this assessment on 24 December 2026 and whenever the purpose, the data, the sources or the sending method changes. If the balance changes, the processing stops until a lawful basis is found.

The right to object

Anyone who receives a message from us can object at any time: by the stop link, by a reply saying so, or by an email to joshua.sandhu@globalphoenix.co.uk. We do not send the first message until all three work: the link has worked since 10 October 2026 (item 3 under the safeguards above), and replies are read every hour by a task registered on 10 October 2026 (item 4). The outreach privacy notice sets out this right and the others.

Changes

On 10 October 2026 this assessment changed from version lia-2026-10-08. “Where things stand” now gives the contact details we collected on 8 October 2026: how many, how we collected them, where they and their copies are held, and what our software checks when it loads them. The ten safeguards we set before the first message now say where each stands: which are in place in our software, which are done by hand, and which start on 12 October 2026. The sources now name both FDA sources, the 510(k) summaries the FDA publishes and the official company registers we read. “Other laws” now states each exclusion rule as our software applies it, and the exclusion of Danish companies that ask not to receive marketing. The sequence is described as our software times it, and today it ends after the first message. The stop link is described as it works: a page with one button, and the one-click form in the email’s headers. The scheduled tasks are described as running only while the director is logged on. Item 3 is now in place: on 10 October 2026 a request made with a stop link our software signed for an address of our own was recorded by this website and read back by our software.

On 11 October 2026 this assessment changed from version lia-2026-10-10. “Where things stand” now says that the outreach database exists: the director created it on 10 October 2026 by running the software’s setup, with 22 contacts and 7 suppression rows for the two Danish companies, where the earlier version said it did not exist yet and would be created on 12 October 2026. Item 4 now says the two scheduled tasks were registered on 10 October 2026 and that the weekday task first runs on 12 October 2026, and it describes the reading of replies as the software does it: the Inbox and Spam folders, every message since the last successful reading and at least the last 26 hours, in full only for a message in a thread we sent into, from an address in the outreach database, or with the subject “unsubscribe”, which is recorded as a request to stop for its sender. The opt-out item says that a person can add a single address to the suppression list by hand.

On 12 October 2026 this assessment changed from version lia-2026-10-11. From 12 October 2026 our software also collects new contacts by itself, once a day, from the 510(k) summaries the FDA publishes for a clearance, decided in the last twelve months, of a company on the two FDA lists, which name a contact person with a work address. “Where things stand” says so and dates it. The necessity test adds the hand collection as an alternative we considered. The nature of the data names the new source. A new safeguard, “Limits on the collection by software”, lists the limits: companies on the two FDA lists only, a newest clearance decided in the last twelve months, at most one contact per company, never a submission consultant, an address only on the company’s own domain, the PDF’s web address as the source, the same checks as every contact, a Danish company held for the manual register check, an address that appears in full in the PDF’s text, body-corporate evidence from Companies House for a United Kingdom company or from a recognised legal form ending the applicant’s name on the FDA’s record, with no contact stored for any other company, and a daily report that names no person. The 8 October collection by the director with Claude is unchanged. The review date is unchanged.

On 13 October 2026 this assessment changed from version lia-2026-10-12. From 13 October 2026 our software sends the three follow-ups as well as the first message: days 5, 11 and 18 after the first, as replies in the same thread, at most three, and the last says that we stop there. Item 9 says what each step links to, that the software sends no follow-up after a reply, a stop request, a bounce, a suppression or a hold or during a deliverability pause, and that follow-ups count against the 25 messages a day and are sent before new first contacts. Where the earlier version said that the software sent only the first message and held steps two and three, the necessity test, the sequence safeguard and item 9 now say that it sends all four. Two safeguards are added: “Choosing between fixed wordings”, which says that the first message exists in three fixed wordings for each role group and that the software chooses one by counts of interested replies for each wording (an opt-out, a negative or a neutral reply does not count), with no effect on a person beyond which wording they read and no profiling, and “Alerts and summaries go only to the director”, which says what the software emails to the director’s own address, that it sends one alert for each thread, that it alerts when someone other than the person we wrote to asks in a sequence thread to be left alone, and that it emails no one else. “Where things stand” is now dated 13 October 2026 and says that outreach email has been sent since 12 October 2026; the version of 12 October said that none had. The safeguard “Opt-outs recognised in replies” says that a stop request from a colleague in the thread also stops the sequence. The records safeguard adds the wording variant. The reasonable-expectations paragraph now says that a recipient might expect up to three follow-ups, and the likely-impact paragraph points to the wording safeguard. The same version also adds the reading of a company’s own website by our software for a shared company inbox when a summary gives no contact it can keep: “Where things stand”, the nature of the data and the safeguard “Limits on the collection by software” say how the software finds and checks the site, that it keeps only a shared inbox and never a person’s address, and that it uses the FDA’s device identifier database only to find the web domain. The review date is unchanged.

On 14 October 2026 this assessment changed from version lia-2026-10-13. The volume caps are now at most 20 first-contact emails a day, where the earlier version said 15, and at most 50 a day of the messages the outreach software counts, where it said 25. Follow-ups still count against the total and are still sent before new first contacts, and item 9 says so with the new figures. The limits for the first fourteen days of a new mailbox and the deliverability pauses are unchanged. The reading of a company’s own website for a shared company inbox, which the version of 13 October 2026 limited to a company whose summary gave no contact the software could keep, now covers every company on the two FDA lists whose newest clearance was decided in the last two years and for which we hold no contact, including a company whose summary the software does not read because its newest clearance is more than twelve months old or because it has only one entry on the two lists, and the software tries the websites of at most 60 companies a day. “Where things stand”, the nature of the data and the safeguard “Limits on the collection by software” say so. The twelve months still applies to a named person read from a 510(k) summary, because a named person may have left the company; a company’s shared inbox does not go stale in that way. Every other limit on the reading of websites is unchanged. The same version adds a source of companies to write to: from 14 October 2026 our software also reads the National Institutes of Health’s RePORTER database and the National Science Foundation’s Awards database, which list the small-business research awards (SBIR and STTR) made to companies in the United States, and keeps an award only if its title or abstract names artificial intelligence or machine learning (an index term in NIH’s record is not enough) with a medical product and a clinical setting, and does not describe drug discovery or development, a therapeutic, a vaccine, an antibody, gene or cell therapy, agriculture, veterinary or defence work. “Processing assessed”, “Where things stand”, the purpose test, the necessity test and the nature of the data now cover a company known from an award. For such a company the contact is the NSF award’s principal investigator, whose business email address the NSF publishes, taken only on the company’s own web domain, or a shared inbox from the company’s own website; no named person comes from NIH. A web address the software guessed from the company’s name is accepted only when a page shows the company’s name together with a phrase of the award’s title or the award’s number, and the software guesses only from the company’s distinctive name words run together or its whole name, never from the first word of a name alone. The safeguard “Limits on the collection by software” sets out these limits, the evidence that the employer is a body corporate (a legal-form suffix on the organisation’s name in the award record, with the award’s public web address as the evidence address) and the daily caps, which apply to these contacts as to every other. The balancing test adds the expectation of a principal investigator who lists a business email address on a federal award. “Choosing between fixed wordings” says that a company known from an award has three further fixed wordings of the first message, which say only that the award record is what we know about the company and do not say that it has or plans a submission. Item 8 now gives the footer in two forms and quotes the second, which names the NIH and NSF records. The same version also adds that, when the FDA’s list of AI-enabled medical devices gives no country for a company, our software finds it from the FDA’s own records read through openFDA: the 510(k) record of the company’s clearance, the premarket approval record (a US state gives the United States), or the establishment registration record when every establishment under the company’s name gives the same country. It stores the country with the web address of the record it came from, and when the records disagree, or a record under the same company name shows Germany, Austria, Denmark, Switzerland or Liechtenstein beside another country, the country stays unknown and the company is excluded. The nature of the data and “Other laws” say so. “Where things stand” is now dated 14 October 2026. The review date is unchanged.

On 23 October 2026 this assessment changed from version lia-2026-10-14. The wordings of the first message and of step two changed on that date. “Choosing between fixed wordings” now says that the three fixed wordings of the first message for a role group differ only in the subject and have the same body, where the earlier version said that they differed in the subject and the opening paragraph, and that the three further wordings for a company known from an award likewise differ only in the subject. It says that every wording of the first message now states that, for a small number of teams, the first full pack is free, in return for the team’s regulatory affairs lead reviewing the pack wording and completing the installation and operational qualification (IQ/OQ) record, the offer the pricing page calls “First full pack free”. The wordings for a company known from an award therefore no longer say only that the award record is what we know about the company and what ProofPack does: they also make that offer, and they still do not say or imply that the company has or plans a submission. A first message to a shared inbox, which opens “Hello,”, carries one added line asking the reader to pass the message to their regulatory lead if it is theirs. The rule that chooses a wording now counts only first messages sent with the wordings of 23 October 2026, and the counts for the earlier wordings are not used; the first of the three wordings leads until another earns the lead. Item 9 now says that step two links to the public sample pack and not to the demo, and states the same offer, in full for a contact whose first message went out before 23 October 2026 and whose step two is sent on or after that date, where the earlier version said that it linked to the public sample pack and the public demo. The sources, the volume caps, the number of steps and their timing, the footer and the records are unchanged. The review date is unchanged.