Why I built Brivvy

I woke up one morning, opened my inbox, and saw that a new email had gone out to our entire user base overnight. Title case on every heading. I had asked for sentence case over and over. A few scrolls down, and there they were: emojis scattered through the body copy. We had agreed on no emojis. Multiple times. In writing. If you can't tell, I'm not a big fan.

This was not a one-off problem. It was a systematic issue that kept happening, over and over, for six years. And it was this exact frustration, compounded by the arrival of AI writing tools that made it exponentially worse, that led me to build Brivvy: infrastructure that gives teams enforceable, machine-readable brand voice parameters at the point of generation. But let me start from the beginning.

The human problem

I spent six years building and growing Uxcel, an education platform focused on UX, product management, and AI. Over that time, we scaled from a two-person operation to multiple teams creating written content: marketing, education, product, leadership, support, and more. Every single team had its own interpretation of "the brand."

Marketing wrote with energy and flair. Product wrote with precision and technical detail. Support wrote with warmth and simplicity. On paper, those all sound fine. In practice, you end up with a company that sounds like five different companies depending on which email, blog post, or help article a user happened to read.

You know what the standard advice is? Create brand guidelines. We did. Most companies do. Here is the problem: 95% of organizations have brand guidelines, and only 25 – 30% actively use them. Guidelines are documents, and documents ultimately collect dust.

The real challenge is not defining your voice, it is enforcing it. Keeping every writer, across every team, aligned on the small mechanical decisions that collectively are your brand. Title case versus sentence case. Contractions or no contractions. Emoji or no emoji. These feel trivial in isolation, but they compound into chaos at scale.

I was far from the only founder dealing with this. When I started interviewing other operators, the same story came up again and again. Gene Kamenez, my co-founder and CEO of Uxcel, described how comms started manageable when it was just him posting on Instagram. The moment he expanded to multiple channels and brought other team members on board, consistency fell apart. Content ended up scattered across Google Docs, Figma, and social schedulers with no single source of truth. Six years in, he still considered it an unsolved problem.

His brand voice evolved organically, never through formal guidelines. And when you are shipping fast, there is simply no time to review everything before it goes out.

Then AI showed up

Let me be clear about something before I go further. Writing with AI is not the problem. It is a natural, inevitable progression that makes teams faster, more productive, and more capable of producing content at a scale that was simply not possible before. Every team should be writing with AI, and I believe that fully.

The problem is that AI, out of the box, does not write like you. It does not write like your brand. It writes like a statistical average of the entire internet, and no amount of prompting changes that in a durable way. What we need is the ability to configure these tools so they mimic how we actually speak, as individuals and as a brand, down to an exact science. That infrastructure did not exist, and so the consistency problem I described above did not get better with AI. It got exponentially worse.

In late 2022, ChatGPT launched and everything changed. Overnight, every team member had access to an AI writing assistant, and content volume went through the roof. Before AI, at least the inconsistency was human. Different writers had different voices, but they were real voices with real personality behind them. Now there was a new writer on the team, one that produced perfectly grammatical, confidently structured, utterly generic content that sounded like nobody and everybody at the same time.

The new nightmare became clear: how do you keep a team consistently creating high-quality output that adheres to your guidelines when the AI forgets those guidelines every single session? Where can those rules live so you do not have to re-upload a document each and every time?

Gene tried to solve this at Uxcel by creating guideline documents in ChatGPT, one for email, one for social, one for thought leadership, and uploading them to "projects" so the AI would reference them automatically. It did not work. The AI forgot its own instructions constantly, his team had no way to verify they were even using the right documents, and every session felt like starting from scratch. When I asked him what caused the most frustration, he did not hold back:

"The most pain for me is aligning this all the time, because it always breaks."

That line stuck with me, because it was not just Gene's problem. It was every founder's problem, and AI had made it structural.

The numbers back this up. As of 2026, roughly 90% of content marketers use AI writing tools in their workflows, and that number keeps climbing. But according to a 2024 Gartner survey of 418 marketing leaders, 77% have explored generative AI while only 44% report realizing significant benefits, with on-brand content generation remaining one of the biggest gaps. Content production has scaled dramatically across organizations, while consistency has moved in the opposite direction.

How AI actually writes

To understand why this problem exists, you need to understand why AI writes the way it does, not at a surface level, but the actual mechanics behind it.

Every LLM is a prediction machine. It reads everything before the cursor, calculates the probability of every possible next word, and picks from that distribution. This process inherently favors the most statistically likely continuation, which is the linguistic equivalent of always ordering the most popular dish on the menu. Safe, predictable, and never distinctive.

Two metrics capture why AI text feels different from human writing. Perplexity measures how predictable each word choice is, and AI text scores lower because it follows safer patterns. Burstiness measures variation in sentence length and rhythm, where humans ramble, pause, go on tangents, and vary naturally while AI stays smooth and even. Together, these explain why even grammatically flawless AI output feels flat.

Temperature does not fix brand voice. Temperature is the dial that controls how "creative" versus "predictable" AI output is. Low temperature gives you nearly the same output every time, while high temperature introduces more randomness. Most AI tools default to a moderate setting. Here is the critical insight: higher temperature introduces randomness, not brand specificity. Cranking it up gives you weirder word choices, not your brand's choices. Random variation is not purposeful.

Reinforcement Learning from Human Feedback (RLHF) is the averaging machine. After initial training, LLMs go through reinforcement learning from human feedback, where thousands of human annotators rank model outputs for the same prompt, those rankings train a "reward model," and the AI optimizes for the highest-scoring responses. When thousands of evaluators rank what is "good," their collective preferences converge on what is generally acceptable rather than what is distinctive. It is regression toward the mean, and the tradeoff is well documented: RLHF significantly reduces output diversity compared to supervised fine-tuning across a variety of measures.

The annotator demographics add another layer. For InstructGPT, more than 85% of evaluators had university degrees, and the largest ethnic group was Southeast Asian at 52.6%. Their linguistic preferences got encoded into the model, which is why words like "delve," "tapestry," and "nuanced" surged in AI output while the em dash became AI's most recognizable tell, with GPT-4o using roughly 10 times more em dashes than GPT-3.5.

Every AI tool converges. GPT, Claude, and Gemini produce similar-sounding output despite being built by different companies, because they share overlapping training data, similar RLHF objectives, the same transformer architecture, and a common set of performance benchmarks. The result is a kind of social dilemma: AI-assisted writing gets rated as more creative on an individual level, but the pieces end up more similar to each other than human-only writing. Writers are individually better off, while the collective scope of content narrows. During Italy's temporary ChatGPT ban in 2023, restaurant content became 15% more lexically diverse and received 3.5% more engagement.

The bottom line is simple. AI was never designed to write like your brand. It was designed to produce the statistical average of all writing, which is the opposite of distinctive.

Nothing else works

If you have tried to solve this problem, you already know the common suggestions do not hold up.

Prompt engineering cannot fix a structural problem. A 2025 peer-reviewed study was direct: enhancing creative diversity through parameter or prompt modifications does not close the diversity gap. You are still sampling from the same probability distribution, and a better prompt does not change what tokens the model considers likely.

Custom GPTs do not scale. System prompts get overridden, they do not share context across team members, and they forget preferences between sessions. Every founder I interviewed who tried this approach described the same cycle: uploaded guideline documents required constant manual re-uploading, had no version control, and the AI still ignored them regularly. They knew they needed more granular voice controls but had no tool to deliver them.

Specialized tools like Jasper offer brand voice features, but they are expensive at scale, require heavy setup, and most marketing teams are consolidating their tool stacks rather than expanding them.

Every one of these approaches tries to constrain AI after generation, treating the symptom rather than the cause.

So I built Brivvy

After six years at Uxcel and dozens of interviews with marketing leaders, founders, and content teams, the pattern was impossible to ignore. Everyone described some version of the same problem: they needed a system that ensures consistent AI output without slowing teams down, a place where brand voice rules live permanently and get enforced automatically.

Brivvy is infrastructure built for the AI era, for teams that want to maintain and adhere to brand consistency in the way they talk. This is absolutely fundamental.

The product decomposes brand voice into two enforceable layers:

• Tones are five adjustable dimensions that shape personality: Warmth, Confidence, Formality, Playfulness, and Technical Depth. Each sits on a five-point scale, and these govern how your brand sounds.

• Rules are hard constraints covering more than 30 parameters: pronoun usage, sentence voice, slang level, emoji frequency, Oxford commas, em dash usage (yes, specifically), contractions, heading case, sentence length, paragraph density, spelling variants, date formats, and more. These govern what your brand does mechanically.

Tones shape personality. Rules enforce mechanics. When they conflict, rules win. Your brand voice becomes machine-readable parameters, not a PDF that gets ignored or a prompt that gets forgotten.

Where Brivvy is headed goes beyond the infrastructure layer:

• Agents that systematically scan and correct content that has already been published.

• A/B testing for copy, because once you define exactly how a brand should write, you can test different parameter scenarios to optimize campaigns.

• Real-time writing, automation, and publishing workflows from creation to send.

• Living wherever you work, starting with AI clients and then extending to Notion, Google Docs, Figma, and anywhere written content gets created.

Full circle

I still think about that email with the title case headings and the emojis. Not because it was a disaster, but because it represented something I could never fully solve with documents, training sessions, or better prompts. The problem was structural. Every person who touched our content introduced drift, and AI amplified that drift by orders of magnitude.

AI was never designed to write like your brand. It was designed to write like everyone's average. The solution is not a better prompt or a more expensive tool. It is enforceable, persistent, machine-readable voice parameters delivered at the point of generation.