Most marketers are still treating AI like a computer.
Type a prompt. Get an answer.
That is why so many marketing teams are frustrated. The output is generic, the tone is wrong, the data is wrong, and the back and forth takes longer than just doing the thing yourself.
The teams pulling ahead are building context systems.
They've organized what their business knows, stored it somewhere their tools can reach, and put a system in place to keep it current as things change.
Prompt skill is the small part. Context architecture is the game.
Dave Shanley and I have been comparing notes in building systems once a week for the past couple of years.
We're both technical founders, we both own marketing technology companies, and we're both building AI systems for marketing and sales. And we have arrived at almost exactly the same conclusion from different directions.
Here is the on-ramp we walked through on the webinar. Crawl. Walk. Run. No matter where you are starting, there is a step in front of you that doesn't require a Kubernetes cluster or an engineering hire.
Co-presented with Dave Shanley, ContentCamel, on May 27, 2026.
If you hire a new person, you do not expect them to know your customers, your pricing, your tone, your sales motion, or the name of the deal you closed last quarter. You teach them. You give them the documents. You sit them down with a few transcripts. You point them at the website. You tell them what you do not do.
AI is the same. The quirks are just different.
It has no sense of time. It does not know who you are. It is not deterministic, so the same prompt at 9am and 11am can give you different answers. It knows a lot about the public internet and almost nothing about your business unless you tell it.
Dave said it on the webinar:
"It's like hiring an extremely intelligent person that just graduated. The best junior you've ever hired, that can type really, really fast, and then putting whole teams of those together."
A team of brilliant juniors who can type fast and never sleep is a real asset.
But only if you have onboarded them. The companies stuck in the frustration phase are the ones who keep asking a brand new hire to write the proposal without showing them the price list.
The first real step beyond "open ChatGPT and type a prompt" is the project.
Every major AI tool has them. Claude and ChatGPT call them projects. Microsoft Copilot and Gemini have a version, but mechanic is the same across all of them. A project is a container that holds two things your prompts cannot: instructions and files.
Most smart people I talk to are already using projects. But they are using them empty. They are using projects as a way to group chats by topic. That is a fine first move, but it is not what makes the difference.
A loaded project has the instructions written down. Who is the AI in this context. What is it supposed to do. What is the format. What are the rules. What are the things it should never do. The same kind of brief you would write for a new contractor.
And it has the files. Brand and tone document. Price list. Personas. Examples of your best writing. Examples of the format you want the output in. Anything you would hand a new hire on day one belongs in the project.
The single biggest step-change at this stage is the brand and tone document.
Most teams do not have one. They have a brand book somewhere in Figma. They have a tone of voice slide deck that the agency made in 2022. They have a writing sample folder that nobody updates. None of those things are accessible to an AI tool.
The fix is a single markdown file. Not a PDF, not a Figma board, not a slide deck. A text file with headers. You can ask Claude or ChatGPT to read your three best pieces of writing and extract the tone for you. Audience. Voice. Signature phrases. Things you never say. Edit the output, save it as a markdown file, drop it in every project. You will get a step change in output quality the next time you ask for anything that needs to sound like your company.
Here is what that gets you, with a real example from last week.
A lead came in from a major referral source. The kind of referral you do not want to fumble. He asked for a one-pager. I did not have a one-pager.
A year ago that is a half day of work. Pull together what we do, who we serve, the pricing, the testimonials, the messaging, the format, the design, then back and forth with someone on the team to get it right.
What I did: screenshot of the email, dropped it into our VERA project, asked for a tailored one-pager. Five minutes for the first draft. About thirty minutes including the back and forth to tighten it. Sent.
It worked because VERA already knew everything about us. Our positioning. Our pricing. Our testimonials. Who we serve. What we do not do. That is the project doing the work. The prompt was almost nothing. The context was everything.
That is the crawl phase. You will know you are out of it when you stop being frustrated by AI output and start being surprised by it.
Once one person on your team has a loaded project that produces good work, the next problem hits fast.
How do you share it? How do you keep it current? Where does the context live?
This was the biggest question that came out of our first webinar.
People left thinking, alright, I get it, context is the thing. But where do I store it? How does my team get to it? What about the things that change every week?
The answer is that not all context is the same.
Some context almost never changes. Brand and tone. Positioning. The high-level "who we are" file. Drop it in a project once, leave it alone for six months.
Some context changes a few times a quarter. Personas. Pricing. Product messaging. Sales decks. You will refresh these, but not constantly.
Some context changes every single day. Sales call transcripts. Support tickets. New marketing data. The latest deal that closed and why. This is the context that has the most signal and the hardest storage problem, because if it lives in a static project file someone has to remember to update it.
We mapped this out on the webinar across eight buckets that cover most of what a sales and marketing org runs on. Customer and buyer intelligence. Products and pricing. Brand and tone. Sales materials. Marketing data. Content assets. Competitive intelligence. Internal conversations and decisions.
For each bucket the questions are the same. Where does it live today. How often does it change. How does the AI tool get to it. If you cannot answer those three questions for a given bucket, you are not ready to ask the AI to act on it yet.
If you only have time to fix one bucket this quarter, fix sales call transcripts.
Dave was the one who hammered this home on the webinar and he is right:
"If you focus on just one thing, get sales transcripts nailed, because that is the single best source of information that will completely transform what's possible with these tools overnight."
Sales calls are a live stream of voice of customer. Real objections, in real words, from real buyers. Most of your marketing problems get easier if your tools can read your sales calls. Messaging tightens. Persona work stops being a guess. Content gets specific. The objections section in your sales enablement deck writes itself.
The hard part is rarely the technology. The hard part is org silos. Marketing often cannot access sales call recordings because sales owns the tool and nobody has connected the pipes. Naming that gap is half the work. Closing it might be your single highest-impact AI project this year.
The mechanical answer for storing changing context is some combination of three things. Team-level accounts on your AI tool, so you can share a project across the org instead of everyone running their own. MCP servers (Model Context Protocol, the new standard that lets AI tools connect to systems like Google Drive, Slack, ClickUp, HubSpot, and the rest of your stack) so the AI can read your live data without you copy and pasting. And eventually, when the volume gets big enough, your own storage layer.
But you do not need to start with the storage layer. You need to start with the team account and the shared projects, because that is free and it works today.
This is where it gets interesting and where the public conversation is the most disconnected from what is happening on the ground.
Once you have shared projects, shared context, and MCP servers connecting your tools, you can move from "the team uses AI" to "the team has AI systems that run while they sleep."
What that looks like in practice for us, for the inbound marketing side: VERA. She started as a Claude project for content development. Personas, brand intelligence, competitive analysis, product messaging, SEO strategy, content strategy, all loaded in.
Then we started building MCP servers to all the tools we have in the the PureSEM Software so she could see the live content calendar, keyword data, traffic, and conversions.
VERA now runs 22 agents. She is our account manager, our analyst, and the engine that produces our client content. She knows the personas, the content strategy, the internal link map, the keyword targets, and the pipeline data for every client we serve. When we develop content, the brief sounds like this: develop this piece in line with the client's content strategy, the buyer's stage, the persona, the internal link plan, and the historical performance data. Closer to onboarding a senior writer than to typing "write me a blog post."
What that looks like on Dave's side, for sales enablement:
ContentCamel runs an always-on social listening system that scans LinkedIn, Reddit, and Slack communities for buyer signals. About 130 signals a day at his current scale. A second system filters those against ICP and persona definitions extracted from sales transcripts and internal documents, classifies them, and pushes warm cold leads into Slack with a draft message ready to go.
A third system runs his newsletter engine. It scans the market, scores about 80 candidate posts a week, runs them through a BS-detector agent, assembles a through line, drafts in brand voice, and hands the final to a human for the last edit.
That is the description of three different teams of five people. Running automatically. Costing approximately the price of a few paid AI subscriptions.
Both stacks were built the same way. One useful slice, then the next, then the next. Compounding.
You should be suspicious of anyone telling you they have all of this nailed. We do not.
A few of the things that are still not solved:
The content production floor inside VERA is 13 agents just for content creation. Writer, editor, editor of the editor, fact checker, link checker, voice checker, and so on. We have it that way because a single agent will miss things. The editor catches the em dashes that the writer left in. The QA catches the missing internal link that the editor missed.
It works incredibly well.
But still misses things and gets things wrong. Vera needs expert human oversight (and expect always will), and it still needs yet more context. We're continuing to add and expand context layers, agents, and micro agents and refine the inputs outputs.
The size of the team matters less than it used to.
A small, agile marketing and engineering team with the discipline to consolidate its context and build always-on systems can now produce work that two years ago would have required either a multinational's infrastructure or a much larger headcount.
Most multinationals cannot move this fast. Their stacks are locked up, their data is in different silos owned by different VPs, and any rollout takes a year.
Smaller companies have an opening here and the teams that take it will pull ahead.
The teams that keep copying and pasting prompts prompted on LinkedIn while their context sits in eleven different Google Drives will not.
If you want to see where you stand on the AI side of inbound, request the Free AI Visibility Assessment.
It is not instant, and that's on purpose, for now.
We'll run our brand intelligence and competitive analysis, build personas, generate persona-driven prompts, check them across the four major AI engines, and generate a massive amount of data. We then have VERA analyze the data and the produce a report with prioritized recommendations. Then we walk you through it. No cost or obligation.
We are running our next webinar in this series in late June. Dave and I will be back, same format. We'll share what we've shipped between now and then, what we have learned, and we're still building, with an aim to continue to continue provide value. If you have questions, we'd love to have them and we'll aim to address them in our next call.
This webinar was co-presented with Dave Shanley, ContentCamel, on May 27, 2026. Watch the full replay at the top of this post.