Brand voice does not usually break because a message got too personal, it breaks from inconsistency between writers, campaigns, and message types that nobody sorted properly in the first place.
Introduction
There is a fear that sits underneath a lot of hesitation about scaling personalisation, and it rarely gets said out loud in quite these words: if every message is written individually, for one specific person, does the brand start to disappear. Does scale mean sounding like everyone and, eventually, sounding like no one in particular.
We think this fear has the mechanism backwards. Brand voice does not usually break because there was too much personality in the room. It breaks because there was too much inconsistency, and inconsistency and personalisation are not the same problem at all.
What Actually Erodes Brand Feel At Volume
Think about how brand voice actually degrades in a growing CRM programme, in practice rather than in theory. It rarely happens because someone wrote something too personal. It happens because five different people wrote five different campaigns this month, each reasonably on brand individually, but subtly different from each other in a way that adds up. One campaign is warm and casual. Another, written under deadline pressure, is clipped and formal. A third leans a little too hard on urgency because that person’s instinct runs that way.
None of these are wrong exactly. Collectively, they are the actual thing eroding brand feel, not the amount of personalisation in any single message. A customer who receives all three over a few months is quietly being introduced to three slightly different versions of your business, and that inconsistency is what reads as the brand losing its shape, not any individual message going too far.
The natural response to worrying about brand consistency is to add more human review, and that instinct is correct up to a point and then works against you. A small team, reviewing a small number of campaigns, can genuinely hold a consistent voice in their heads. As volume grows, that same team is reviewing more content, faster, and the quality of that review degrades exactly when you need it to hold steady.
This is the trap that manual scaling walks into. You cannot solve an inconsistency problem by asking increasingly stretched humans to catch increasingly frequent inconsistencies. The review capacity and the volume are moving in opposite directions.
So what practical steps can you take?
Expert help is only a call away. We are always happy to give advice, offer an impartial opinion and put you on the right track. Book a call with a member of our friendly team today.
Step One: Write The Voice Down Properly
Let’s be honest, most guidelines are 117 powerpoint slides on what colour blue to use with one page telling you to write something warm but informative. It was probably written 4 years ago by some swanky external agency, everyone agreed it was great, and no one has looked at it since.
Fix this first, before anything else on this list. A useful version is short and specific: the tone in one sentence, three things the brand would never say, and two or three real example messages that everyone agrees sound right. This step needs no tools and no budget, and most teams who do it for the first time are surprised by how much disagreement it surfaces about what “on brand” actually means.
Once the voice is written down, turn it into something you can actually check against. Pick fifteen or twenty realistic scenarios, a churn risk message, an upsell moment, a complaint follow up, and write what a genuinely on brand response looks like for each. This can be done entirely in a spreadsheet, scored by a human against the document from Step One. This is your eval set, and building it by hand, before any AI is involved, is what makes it genuinely useful later rather than an afterthought bolted onto a tool.
Step Two: Score Your Current, Human Written Campaigns Against It
Before worrying about whether a future agent will stay on brand, check whether your current human team already does, consistently, across different writers and different weeks. Take a sample of recent real campaigns and score them against your test cases and rubric from Step One. This usually reveals the actual, current state of your brand consistency, which is frequently less consistent than assumed, entirely without any automation being involved yet. It also gives you a genuine baseline to compare anything automated against later.
For performance led messages, quantitative funnel data is the right primary measure: open rate, click rate, conversion rate, revenue per send, with unsubscribe rate as a check against pushing too hard.
For brand led messages, you will need softer relationship-level signals over the following weeks. Whether engagement with this customer looks any different to a matched group who did not receive it, whether reply sentiment stays positive, whether the unsubscribe rate stays flat, which is itself the success.
Run this scoring against a sample of real, recent, human written campaigns. It usually reveals the actual state of your brand consistency, separately for each purpose, which is a baseline worth having before anything automated enters the picture.
Step Three: Only Then Introduce AI
Once you do start using AI to help draft messages, run its outputs through the exact same test cases and rubric before trusting them for real campaigns. This is the point where the work from the earlier steps pays off, because you are not judging a new tool against a vague feeling of whether it sounds right. You are judging it against the same, specific standard you have already established and already know your own team’s baseline against.
Make sure the purpose tag, performance led or brand led, travels with the message into whatever brief or prompt the drafting agent works from. This matters more than it sounds. An agent given access to engagement data, without being told a given message is brand led, will tend to optimise for whatever it can actually measure, and will quietly start adding a call to action to a message that was only ever supposed to say hello, because a click is visible to it and warmth is not. That is a textbook case of a measure becoming the target at the expense of the thing it was meant to represent, and the purpose tag is what prevents it.
Whenever the underlying tool, prompt, or model changes, run the same eval set again before trusting new output for real campaigns, the same discipline a software team applies before shipping a change. This is what catches drift early, a subtle shift in tone that crept in after an update, rather than noticing three months later that campaigns have slowly started to feel different.
The Data Worth Watching As Complexity Builds
Two different scorecards, watched on the same schedule, not one blended number. For performance led messages, the usual funnel metrics, tracked over time, with the consistency score from your rubric as a sense check alongside them, since a message can convert well and still be slightly off brand, and that gap is worth knowing about even when the numbers look fine. For brand led messages, the consistency score carries most of the weight, alongside the slower relationship signals, retention lift, reply sentiment, flat unsubscribe rate, that confirm it worked weeks later rather than the same afternoon.
Watch the variance in consistency scores across different writers or, later, different agent versions, since a wide spread is often more useful than the average alone. And keep an eye on customer facing language directly, complaint themes mentioning the message felt generic or off, which tends to surface a brand consistency problem before it shows up in any of the numbers above.
Start with the document, sort by purpose, and build both test sets by hand before you ever check a tool against them. By the time personalisation and automation genuinely scale up, you will have a specific, purpose aware standard to hold them to, rather than a feeling that something has started to sound slightly wrong.
Where Agentic Fits In
This is where a QA agent, checked explicitly against a written brand voice guide rather than a vague shared sense of “how we sound,” genuinely outperforms a stretched human team, not because it is cleverer, but because it does not get tired on the fifth review of the day the way a person does. It applies the same standard to message one and message five hundred, at the exact volume where a human team’s consistency naturally starts to slip.
Personalisation, in this framing, is not the threat to brand voice. It is the layer on top of a consistent foundation. The voice, the tone, the things the brand would never say, stay fixed. What changes message to message is which specific detail about this specific customer gets referenced, inside a voice that never wavers. Done properly, a customer should be able to tell it is your business from the tone alone, even before they notice the personal detail sitting inside it.
None of this removes the need for a person who genuinely understands the brand. It changes what they spend their time on. Instead of reviewing every individual message for tone, hoping to catch drift before it ships, they write and maintain the actual definition of the voice, the document a QA agent is checking every message against, and they spend their remaining time on the messages that are genuinely ambiguous or high stakes enough to deserve their full attention.
Get In Touch
Our friendly team are always on hand to answer questions, troubleshoot problems and point you in the right direction.
