What UK Email Law Means for AI Outbound in HubSpot

Someone has put a proposal in front of you: use AI to research accounts, draft the approach and run outbound at a volume your team could not reach by hand. The commercial case is easy to follow. What nobody in the room can answer is whether it is legal in the UK, and what has to be true to keep it legal.

The rules are not new, and they do not change with volume. What changes when a model does the generating is that the checks a human sender ran without noticing stop happening: recognising the customer who complained last year, noticing this is the third email to the same person this month. Those go once drafting and sending are separated from the reader, and an assumption that produced one bad email now produces a segment of them. The law takes two minutes. The configuration is the rest of this post.

What PECR regulation 22 says, and who it covers

PECR, the Privacy and Electronic Communications Regulations 2003, carries the UK rules on marketing email. Regulation 22 sets its own scope in its opening line (retrieved from legislation.gov.uk on 16 Aug 2026):

"This regulation applies to the transmission of unsolicited communications by means of electronic mail to individual subscribers."

It does not extend to corporate subscribers, and that one word, individual, is where the confusion starts.

For individual subscribers, regulation 22(3) carries the soft opt-in, permitting unsolicited direct marketing email where all three of these hold:

• the sender obtained the contact details during the sale or negotiation of a product or service to that recipient
• the marketing concerns only similar products and services from the sender
• the recipient was given a simple, free means of refusing at the point the details were collected, and in every subsequent message

Why "B2B email is exempt" is not a safe summary

Sole traders and partnerships are generally treated as individual subscribers. A list labelled B2B in your CRM will therefore hold a mixture: limited companies, and people trading under a business name who sit inside regulation 22 as a private individual does. The position depends on who each recipient actually is, which is a question about your data.

SpotDev does not advise on regulatory classification and nothing here is legal advice. Which of your recipients are individual subscribers, and what your obligations are on a given list, are for your own advisers to rule on. Two things are worth taking into it. A corporate recipient is not a free pass, because a named person at a limited company is still personal data and UK GDPR applies regardless of where the PECR position lands. And what follows is engineering rather than interpretation, so you need it whichever way that goes.

What you configure so the rules hold

Consent and lawful basis on the record, not in a spreadsheet

Most organisations can say what their consent position is. Far fewer can say what it is for one named contact without opening a file, and state that lives in a spreadsheet cannot be read at the moment a message is about to leave.

It belongs on the contact record as properties the system can read: the lawful basis relied on, where and when the details were obtained, and which categories of marketing the person accepted or refused. If you rely on the soft opt-in, the sale or negotiation that produced the details has to be identifiable on the record, not asserted at the level of the list.

The test is a query test. If you cannot ask your CRM for every contact currently relying on the soft opt-in and get an answer in seconds, you do not have a consent record. You have a description of one.

Suppression checked at send time, not at list build

A list is assembled on Monday and the messages go out on Thursday. In between, someone unsubscribes, an account manager adds a name after a difficult call, and a deal is lost. If suppression was applied at list build, all three are sent to. The list was right when made and wrong when used.

So the check runs at the moment of sending, against live state, per message. The send path is designed to fail closed, so a check that cannot complete stops the message rather than being treated as a yes.

An instruction in the agent's prompt telling it not to message a suppressed contact is not a control. It is a preference expressed to a system that generates text. The check belongs in the sending path itself, outside the agent.

Making "similar products and services" enforceable

The aspirational version of this condition is a line in a policy document, interpreted by whoever writes the campaign that week.

The enforceable version has three parts. What the recipient bought or negotiated is recorded in a form the system can read, whether that is the line items on their deals or a category property kept for the purpose. Each outbound asset carries a category. The send is a match between the two rather than a judgement at a keyboard.

An agent may draft an approach outside the permitted category. It must not go out under a basis which does not cover it, so it routes to a person.

An unsubscribe route in every message, and what breaks it

The third condition requires a simple, free means of refusing at collection and in every subsequent message. What breaks it:

• A sending path that bypasses the usual one. Outbound meant to read as personal goes through a route separate from the one the rest of your marketing uses, which was never set up to carry a refusal link.
• The footer removed to make the email look human. Somebody strips it from the template because it reads like a mailshot, and the sequence inherits the change.
• A refusal route that takes effort. A link that asks the recipient to log in or explain themselves. Whether that satisfies the condition is for your advisers. As engineering it is broken, because people who cannot use it report the message as spam.
• An opt-out honoured in one place and not another. A refusal registered against marketing email but never applied to the sales sending path, or a reply saying "take me off your list" that is deleted while the record stays untouched.

Every sending path needs a working refusal route, and an opt-out captured anywhere has to take effect everywhere.

What the agent selects and drafts, and what a person authorises

The same capability can build an internal queue unasked and still be unable to release one message until a named person approves it. For outbound the steady state is usually draft, permanently.

Nothing there has an agent deciding anything consequential. It selects inside rules a person set, drafts for review and flags what needs a human.

Logging that can reconstruct a single message

Sooner or later somebody asks why one person received one message. Answering needs six things, recorded at the time of sending:

• which audience rule selected them, and the version in force
• which lawful basis applied, and what the record said at that moment
• which suppression checks ran, and what each returned
• which content version was sent, and through which sending path
• who approved the release, and when
• what happened next: delivery, reply, opt-out, complaint

The second is the one people get wrong. Consent state is not static, so a log that points at the record rather than copying the state as it stood will tell you only what is true today.

One system of record instead of three

Most tools sold for this arrive as a separate platform holding their own copy of your contacts, their own suppression list and their own audit log, so consent state sits in two places and reconciles on a schedule. A contact opts out in your CRM on Tuesday, the platform is working from last week's copy, and it sends on Wednesday. Every component behaved as designed.

Built inside the CRM the business already runs, consent state, the suppression list and the audit trail are one thing rather than three. That is the case for sales AI built inside your HubSpot estate rather than bought alongside it.

How you would know the configuration is holding

Measures rather than assurances. Four worth reporting monthly:

• messages released where the send-time suppression check did not complete. The target is zero, and it needs monitoring
• the split between opt-outs by link and opt-outs by reply. Mostly replies means the link route is broken
• elapsed time from an opt-out arriving to it taking effect on every sending path
• drafts rejected at approval, with reasons. A rising rejection rate is a design problem, not a compliance problem

Then run the sampling test. Take ten sent messages at random and reconstruct each from the log alone. If you cannot, do not increase the volume.

The next step

If AI outbound has been proposed and nobody can say which contacts rely on which lawful basis, that is the first piece of work, and it is a data question before it is an AI one.

Our diagnostics are short, fixed-scope assessments that show where to invest first and the roadmap to get there. No obligation. For outbound that means the state of your consent and suppression data, and where the approval boundaries would sit before anything runs at scale.

Where outbound is part of a wider departmental rollout, our 12-month AI transformation programme takes one department at a time, activates two agents per wave, and is committed a quarter at a time. Where outbound sits in a wider plan is covered in how a rollout is sequenced.