How To Review An AI-Generated Campaign Before Sending

Generative AI can produce an email campaign in minutes. That does not make it ready to send.

The real risk is rarely that a message merely “sounds like AI”. The greater problem is that automation can multiply ordinary campaign faults: an invented product claim, an offer that has expired, a personalisation field used in the wrong context, a tone that does not fit the relationship, or a large send to people who have not recently shown that they want it.

A useful review treats the generated draft as an input, rather than an approved asset. Someone who understands the offer, the audience and the sending set-up should check the campaign from the recipient’s point of view and from the mailbox provider’s point of view. The two overlap more than teams sometimes assume: unclear, irrelevant or unexpected email is more likely to be ignored, unsubscribed from or reported as spam.

Start with the campaign decision, not the copy

Before scrutinising a subject line, establish what the campaign is trying to do and whether email is justified for this audience at this time. AI tools are good at filling an empty page; they do not know whether the planned contact is appropriate unless the organisation supplies reliable constraints.

The campaign owner should be able to state, in plain language:

• the intended recipient segment and why those people should expect this message;
• the relationship or permission basis that supports the send;
• the single main action being requested;
• the offer, dates, prices, availability and geographic limits that have been verified by the responsible commercial team; and
• what should happen if a recipient does nothing, opts out, complains or replies with a question.

This is also the point to challenge volume. A persuasive-looking campaign is not a reason to add dormant records, old event leads or every customer ever acquired. If the campaign represents a material change in frequency, message type or audience, start with the people most likely to recognise and value it. That is sound audience practice as well as a practical way to limit the effect of an error.

Review the brief supplied to the AI

The quality of the first draft depends on the information and instructions that went into it. Retain the approved campaign brief and the final prompt or source inputs where practical. This creates a straightforward audit trail and makes it easier to understand why a questionable line appeared.

Check especially for data that should not have been supplied. A copywriting task usually does not need a full customer export, free-text support history or sensitive personal information. Use only the minimum data needed for the job, and make sure the AI supplier, workspace settings and internal access controls have been assessed for the intended use. The ICO’s guidance on AI and data protection highlights data minimisation and security as central considerations; its guidance is being kept under review, so teams should use the current version when shaping their internal process.

Where personalisation is needed, give the model approved fields and rules rather than asking it to infer personal circumstances. “Use the customer’s purchased product name, only if it is present and current” is a safer instruction than “make the email feel highly personal”. It reduces the chance of an email making an unwelcome or inaccurate inference about a recipient.

Check claims as though every sentence needs evidence

Generated prose can be fluent while being wrong. It may join two real facts into a false conclusion, use a feature name that does not exist, invent a customer quote, imply a guarantee or misstate a date. A confident tone is not evidence.

Give a named reviewer responsibility for validating each material assertion against a current primary source: the approved product information, promotion terms, price list, stock system, legal wording or a subject-matter expert. Do not rely on the AI conversation itself as the source.

Be particularly careful with:

• pricing, savings claims, finance, health, safety, environmental and performance claims;
• availability, delivery times, eligibility and deadline wording;
• comparisons with competitors;
• testimonials, reviews, awards and statistics; and
• claims that an account, subscription or benefit will be affected unless the recipient acts.

Review the message as a whole, including the subject line, preview text, hero image, call to action, footer and landing page. A carefully qualified body paragraph does not correct an over-claiming subject line. Equally, an image can carry a claim that the text review missed.

Separate persuasion from impersonation

AI often produces urgency because urgency is a common marketing pattern. The reviewer should ask whether the message creates a misleading impression of scarcity, account risk or personal attention. Phrases such as “final notice”, “we noticed you have not…” or “your access may be removed” need a genuine operational basis, not merely an expected lift in clicks.

Do not manufacture familiarity either. If a message refers to a previous purchase, an abandoned basket, an advisor or a location, make sure the underlying data is accurate and that the recipient would reasonably expect it to be used in this way. If the organisation cannot explain the reference comfortably to a recipient who replies, remove it.

Read it for relevance, clarity and brand judgement

A human editor should read the email aloud, or at least read it outside the generation interface. Repetition, generic superlatives and oddly formal phrasing become easier to spot when viewed as a recipient would see them.

Ask simple questions:

• Does the opening quickly explain who is writing and why this recipient is receiving it?
• Would a subscriber who has not seen the campaign brief understand the offer and the next step?
• Is there one primary call to action, with a destination that fulfils the promise made in the email?
• Does the tone suit the brand and the audience, including customers in a service or support-sensitive moment?
• Could any sentence be read as patronising, exclusionary, threatening or more personal than the relationship warrants?

This does not require stripping out personality. It means making deliberate choices. A short, plain email can be excellent if it is useful and expected. A long, highly designed email can work too, but only when its content earns the attention it asks for.

Review the audience and permissions independently of the draft

Copy approval and list approval should be separate decisions. An AI tool cannot establish whether a particular address has valid permission, has withdrawn it, has bounced repeatedly, has complained, or has received too many recent messages.

For UK marketing email, the rules depend on the recipient and circumstances, but organisations normally need consent to send unsolicited electronic marketing to individual subscribers unless an exception such as the products-and-services soft opt-in applies. The ICO also makes clear that recipients must be able to opt out and that organisations must respect that choice. Review the applicable rules with privacy or legal colleagues where the campaign, territory or data use creates uncertainty.

Before exporting the final segment, apply the current suppression list and exclude unsubscribes, complaints, hard bounces and records subject to any internal contact restriction. Confirm that the segment logic does not accidentally include people simply because data is missing. For example, “has not purchased in 12 months” is not the same as “has been engaged in the last 12 months”.

Pressure matters too. A recipient who has been sent three promotional messages in four days may regard a fourth, perfectly well-written promotion as unwanted. Review recent sends by audience and by individual where your platform supports it. If the campaign is not essential, suppress people who have recently received comparable messages or who have not shown recent signs of engagement.

Make unsubscribing and identity part of the content review

The unsubscribe link, sender name and footer are not decorative compliance items. They tell the recipient whether the message is trustworthy and give them a lower-friction alternative to marking it as spam.

Check that the visible sender identity is recognisable, the physical or contact details are correct where required, and the preference and unsubscribe links work in the rendered email. The route should be understandable: a recipient should not have to sign in, solve a puzzle or make several choices just to stop promotional email.

Google’s guidance treats marketing and subscription messages as a distinct category. For bulk senders to personal Gmail accounts, it requires one-click unsubscribe support for relevant traffic and a clearly visible unsubscribe link in the message body. Google’s subscription guidance also says senders should honour unsubscribe requests within 48 hours. These are operational requirements, so test the actual headers and unsubscribe endpoint rather than assuming the template setting is correct.

Test the assembled email, not just the approved words

Many pre-send failures happen after copy approval: a merge field breaks, a link adds the wrong campaign parameter, an image does not load, a dark-mode rendering makes text unreadable or a mobile layout hides the call to action. Send a proof through the exact platform, template and sending domain that will be used for the live campaign.

Test the offer path as well. A technically valid email can still fail the recipient if the landing page is slow, the voucher does not apply, stock is unavailable or the terms differ from the email. If someone can purchase or provide personal data after clicking, involve the team responsible for that experience in the approval.

Check delivery readiness before scheduling

AI-written content does not automatically cause spam placement, and changing a few supposedly “spammy” words is not a deliverability strategy. Mailbox providers use many signals, including authentication, recipient feedback, sending patterns and message characteristics. The practical question is whether this campaign is being sent through a healthy, appropriately configured programme to recipients who are likely to expect it.

At a minimum, confirm that the sending domain and provider configuration are the approved ones, and that SPF, DKIM and DMARC are operating as intended. Gmail requires all senders to meet baseline technical requirements and requires SPF, DKIM and DMARC for senders sending more than 5,000 messages a day to personal Gmail accounts, alongside alignment and unsubscribe requirements. Its Postmaster Tools can report on authentication, spam rate, reputation and delivery errors, although the data is not real time.

Review recent performance before increasing volume or reviving an inactive segment. Look for complaint signals, bounces, deferrals, authentication failures and changes in inbox placement. If the campaign has a different cadence, a much broader audience or a new creative approach, use a measured rollout and watch results before releasing the remainder. Avoid treating a warm-up or pacing plan as a way to force unwanted mail through; it should reduce operational risk while protecting recipients and existing sender reputation.

Do not use a normal employee mailbox as an improvised bulk-email platform. Microsoft says bulk sending through Microsoft 365 is supported only on a best-effort basis and advises customers to use appropriate bulk-email infrastructure, authenticated subdomains and current mailing lists. That distinction is important when a team wants to send an “urgent” AI-generated announcement quickly.

Give a human reviewer authority to stop the send

The final approval should be more than a tick-box. The reviewer needs authority to hold the campaign when the proof exposes a material concern: unclear permission, unverified claims, a misleading subject line, a faulty unsubscribe route, unexpected volume or an unhealthy delivery signal.

Keep a brief record of what was approved, by whom, the final audience count, version of the creative, proof send and scheduled time. This is useful when investigating a complaint or a delivery problem, and it prevents a last-minute “small” edit from bypassing the review.

For programmes already experiencing high complaints, poor inbox placement or domain-authentication problems, the immediate issue is bigger than the AI draft. Email Foundry’s managed recovery work is relevant in that situation because it combines authentication and infrastructure checks, audience suppression, contact-pressure controls, provider-level pacing and monitoring. The purpose is to restore a permission-based sending programme, not to make questionable campaigns deliverable.

AI can make campaign production faster. A disciplined pre-send review makes sure that speed does not become a faster route to recipient distrust, complaints and damaged deliverability.