Building the AI context layer that turns generic answers into brand-specific work
Most teams that bought AI tools this year came away disappointed. The 2026 State of AI for GTM report, where Kyle Poyar and Maja Voje interviewed thirty leaders and collected forty working plays, put a number on it. Fifty-three percent of go-to-market leaders said AI delivered little to no impact. Only 24 percent reported real returns.
Read that again. Three out of four teams spent money and saw nothing move.
Here is the part that should bother you. Everyone in that survey had access to the same models. Same ChatGPT, same Claude, same Gemini. The teams seeing returns and the teams seeing nothing were typing into identical tools. So the gap was never about who had the better software.
What “generic” actually means
Open your last AI-written email or landing page. Read it as if a competitor sent it. Would you know the difference?
For most teams, the honest answer is no. The output reads like it could have come from any company in the category, because it was built from the same raw material every other company is using. A request goes in, a general answer comes back, someone polishes the surface, and it ships.
That output is the average of the internet on your topic. It is competent. It is also indistinguishable from what the team across the street produced the same morning.
When you swap tools and tighten your prompts and still get bland, interchangeable copy, the tool is doing its job. You handed it generic inputs and it gave you generic results. The machine is a mirror.
The teams getting real work out of AI feed it something different
The 24 percent are not better at writing prompts. They have built a habit the other teams skipped. Before they ask AI to produce anything, they make sure it already knows the business.
Kieran Flanagan at HubSpot described his version of this. He built what he calls a customer digital twin, a model he tests campaigns against before launch. It runs on real sales call transcripts, real G2 reviews, real objection notes from the CRM. His warning was blunt. Synthetic data in, synthetic insights out. The twin is only as sharp as the voice-of-customer you train it on.
That principle holds for everything, not just customer simulations. The AI writing your outbound, your blog drafts, your positioning, all of it inherits the quality of what you put in front of it. Feed it your actual customer language and it writes in your customer’s voice. Feed it nothing and it writes in the voice of the average.
A context layer is a curation job, not a coding job
This is where teams talk themselves out of the fix. They assume building this kind of system means a technical project, a data engineer, a six-month integration.
It does not. The report’s own conclusion was that you do not need to be an AI expert or a software engineer to start seeing value. The highest-leverage move a functional team can make is a documentation project. You decide what the AI needs to know, gather it, and write it down in a form the model can read. Most of that work is judgment, and your team already has the judgment. They just have it scattered across people’s heads, old decks, and a hundred Slack threads.
Maja Voje described the simplest version of this in the report. She has the teams she works with build a content assistant inside a single AI project. They upload their messaging and brand guidelines, a knowledge base describing the company, and examples of their best-performing content. Then they write instructions that define the voice and spell out what to do and what to avoid. That is the whole setup. No special platform, no engineering ticket.
What goes into the layer
Build it from the intelligence you already own:
Your ICP, written down. Not the aspirational version on the website. The real one, including the segments that close fast and the ones that waste your time.
Real customer language. Pull exact phrases from call transcripts, reviews, and support tickets. The words your buyers use to describe the problem, copied verbatim.
Your best work. Three to five examples of copy, posts, or pages you were proud of. The model learns your voice from samples, not adjectives.
What has failed. The campaigns that flopped, the messaging that confused people, the angles that never landed in your market. Telling AI what to avoid is as valuable as telling it what to chase.
Your market position. How you differ from the obvious alternatives, in one honest paragraph.
Once it exists, the same prompt behaves completely differently. The request that produced generic filler last week now produces something that sounds like you, references your customers correctly, and reflects how you actually win. You did not get better at prompting. You gave the model a memory.
Start with one workflow
You do not need to document the entire company before you see a return. Pick the one task your team runs most often. Outbound email, social posts, first-draft blogs, whatever eats the most hours.
Joey Maddox at Verisoul makes the case for this with meeting prep. His team loads prior transcripts, email threads, and CRM history into a custom GPT before every call, and his takeaway was simple. The more relevant data the AI can reach, the more useful it gets. Start there. Gather the context for one workflow, load it into a project, and run your normal prompt against it. Compare the two outputs side by side. The difference is usually obvious enough to settle whether this is worth your time.
Then do the next workflow. The layer compounds. Every document you add makes every future task sharper, and you build it once.
The teams in that report who saw nothing are still treating AI like a search engine, asking strangers for answers and accepting whatever comes back. The teams seeing returns gave their AI a home, a memory, and a clear picture of the business it works for. The tool was never the variable. What it knew was.
. . .
Want to save hours each week by turning work into repeatable AI workflows?
The Fortune 100 AI Skills Library™ includes plug-and-play prompts built to save leaders time and money. Copy, paste, and edit in 60 seconds, then apply them across planning, execution, and reporting.
Top comments (0)