Your team starts using AI. The output is technically fine - grammatically correct, reasonably structured. But it doesn't sound like you. It's generic. It uses phrases you'd never say. It misses the tone your best salespeople have nailed over five years of customer calls. So you edit heavily, stop using it, or quietly accept that your content now sounds like everyone else's. None of those are the right answer.
The context problem
Large language models are trained on the internet. The internet, by volume, is mostly average. When you ask an AI to write something without giving it context about who you are, who you're writing for, and what your voice sounds like - it defaults to the statistical average of all the content it has ever seen on the topic.
That's why B2B AI content so often reads like it was written by a committee that had never met a customer. There is an interactive demo of exactly this if you want to watch it happen. It's not wrong, exactly. It's just the most average possible version of the thing you asked for. And average is invisible.
"Most teams are prompting AI like they're Googling something. One sentence in, expect a finished result out. The model knows nothing about you unless you tell it."
What context actually means
Context isn't just a system prompt. It's everything the AI needs to know to produce output that could have come from your best writer on their best day - your brand voice, your ICP, your product's specific positioning, the phrases you use and the ones you deliberately avoid, and the format conventions you've established.
Most teams provide none of this. They open a chat window, type a one-liner, and wonder why the output needs six rounds of editing.
The brand voice document
A brand voice document is the single most leveraged AI setup asset a B2B company can build. Not a brand guidelines PDF with hex codes and font names - a working document that describes how your brand actually talks. Sentence length. Words you use. Words you ban. Tone adjectives with examples. Your point of view on the market. The one thing you'd never say.
When this document is fed into every AI session as context, the output quality gap closes significantly. The model isn't guessing at your voice anymore - it has a reference point.
// Paste at the top of any AI session or into your system prompt BRAND VOICE Tone: Direct, confident, a little dry. No hype. Sentence length: Short to medium. Max 2 clauses. We say: "founders", "operators", "pipeline", "signal" Never say: "leverage", "synergy", "robust", "seamlessly" We don't: Start sentences with "In today's..." We do: Use em dashes. Lead with the point. POV: Operators beat strategists. Show your work. Format: Subheads every 3-4 paragraphs. No bullet soup.
ICP and customer context
AI doesn't know who your customer is unless you tell it. And "B2B SaaS founders" isn't enough. What stage are they at? What's the problem they're actively aware of versus the one you're trying to make them aware of? What objections do they carry? What language do they use to describe the problem you solve?
The more specific your ICP context, the more the AI can calibrate - not just the words it uses, but the problems it leads with, the examples it reaches for, and the objections it addresses.
Product and positioning context
This is where most AI outputs go wrong in B2B specifically. The model knows your product category from the internet - which means it knows it the way a general observer would describe it, not the way your best sales rep would. It doesn't know your specific differentiators, your pricing model, or the exact use case your best customers hired you for.
A product context document - 300 to 500 words describing what you do, for whom, at what price point, and how you're different - transforms AI output from generic category description to something that could actually live on your website.
Building the system, not just the prompt
One good prompt is a hack. A system is what makes AI output consistently good across your whole team - not just when the one person who's good at prompting happens to be using it that day.
A shared prompt library
A prompt library is a structured collection of tested prompts for the recurring content tasks your team does: LinkedIn posts, cold email sequences, case study drafts, product update announcements, proposal sections, sales follow-ups. Each prompt includes the context preamble, task instruction, and format requirements.
Without a library, every team member reinvents the wheel every time. With one, the best prompt your team has ever used for a given task is available to everyone. The quality floor rises immediately.
Custom GPTs and Claude Projects
Most B2B teams are still using off-the-shelf ChatGPT or Claude with no customisation. The step change comes when you build a Custom GPT or Claude Project that has your brand voice, ICP, product context, and prompt templates baked in as persistent system-level context.
Your team opens the tool and it already knows who you are. Every session starts with a full context baseline. The output quality difference is immediate - and it removes the dependency on the one person who happens to know how to prompt well.
Feedback loops and iteration
AI setup isn't a one-time project. The brand voice document needs updating when the brand evolves. The prompt library needs new entries as new content types emerge. The ICP context needs refreshing when your target customer changes. A lightweight quarterly review - 30 minutes - keeps the system calibrated as the business grows.
The hallucination question
Most AI hallucinations in B2B content aren't the model inventing facts - they're the model filling in gaps you left open. When you don't tell it your pricing, it guesses. When you don't tell it your differentiators, it uses generic category differentiators. When you don't tell it your customer's specific pain points, it uses the most commonly discussed ones in content about your category.
This is why hallucination is mostly a context problem, not a model problem. The model isn't lying - it's doing its best with what you gave it. More context means fewer gaps. Fewer gaps means fewer errors and far less editing time.
- Pricing hallucinations: give it your actual pricing, or explicitly tell it not to mention pricing at all
- Feature hallucinations: include a product fact sheet with specific feature names and descriptions
- Competitor hallucinations: tell it which competitors you acknowledge and which you don't engage with
- Customer hallucinations: name your actual customer segments, verticals, and use cases explicitly
What good AI setup looks like in practice
A B2B company with a well-built AI foundation looks like this: a shared Notion or Google Doc workspace with the brand voice doc, ICP document, product context brief, and prompt library. A Custom GPT or Claude Project with those documents loaded as persistent context. A naming convention for saved prompts. A quarterly review cadence. A short onboarding doc so new team members can use the system from day one.
The build time is typically 2–3 weeks for a focused operator. The payoff is a team producing on-brand AI content in a fraction of the editing time, without depending on whoever happens to be good at prompting that week.
"Every hour spent building the AI context system saves three hours of editing per week, indefinitely. It's the highest-leverage marketing infrastructure investment most B2B companies aren't making."
Where to start
- Write the brand voice document: 400–600 words. What you sound like, what you don't. Use real examples from your best-performing content.
- Build the ICP brief: one page. Stage, role, problem awareness, language they use, objections they carry. Be specific - "Series A SaaS founders in fintech" beats "B2B buyers."
- Write the product context brief: what you do, for whom, at what price point, how you differ. Written the way your best sales rep explains it on a first call.
- Set up a Custom GPT or Claude Project: load all three documents as system context. Test with five common content tasks before rolling out to the team.
- Start the prompt library: begin with your five most common content tasks. Document the best prompt for each. Share it with the whole team.
- Book a quarterly review: 30 minutes every three months to update context docs and add new prompt entries as the business evolves.
Your AI outputs sound generic because you're giving it generic inputs.
Brand voice, ICP context, product positioning, and a shared prompt library are the four foundations. Build them once, and every AI session your team runs benefits. Skip them, and you're paying an editing tax on every piece of content indefinitely.
Fixing it properly. This is a one-off setup job, not an ongoing prompt-writing habit. Build the context once and every output after it improves. Here is how I run that.
Want to see how generic your current copy reads? The Waffle-o-Meter scores any page or block of text for buzzword density, sentence bloat and unbacked claims, then quotes your worst sentences back at you. Free, no sign-up.
