LLMs Were Never the Bottleneck
Two years ago, AI engineers were obsessed with prompts. They spent their days writing and rewriting the perfect phrasing, hoping that if they asked the model the right way, it would do what they needed. But they were wrong.
They moved on to something called context engineering. Andrej Karpathy described it as filling the context window with the right information for the next step.
From X
Gartner declared in 2025 that prompt engineering was out and context engineering was in. Anthropic now defines it as choosing every token a model sees before it reasons.
Turns out the model wasn't the bottleneck. The context was.
Marketing is two years behind this realisation. The cost of that delay is sitting on every CMO's screen right now, in the form of a dashboard.
When AI teams stopped tuning prompts and started engineering context, the lift was exceptional. Workflows that kept breaking started holding together. Things that felt like science experiments started shipping. Nothing changed about the models themselves. Only what they could see.
It turned out that you can keep asking sharper questions of a system with no memory and no shared state. Or you can fix what's underneath so the system actually knows what it needs to know.
Marketing has been asking sharper questions of dashboards for twenty years.
It's time we admitted the dashboards have no memory and no shared state either.
Fourteen Dashboards, Zero Answers
Most marketing teams maintain somewhere between fourteen and twenty dashboards. Each one was built to answer a specific question someone in leadership asked at some point. Together, they answer almost nothing.
A marketing performance dashboard, a campaign attribution dashboard, a channel ROI dashboard, a pipeline contribution dashboard, a funnel velocity dashboard, a persona engagement dashboard, an ABM dashboard, a web analytics dashboard, a nurturing performance dashboard, an events dashboard, a content dashboard.
The Frozen Snapshot Problem
A dashboard is a frozen snapshot of a motion that's already moved. By the time the data lands, the channel weights have shifted, buyer behaviour has changed, last week's working campaign has saturated, and the dashboard is reporting on a world that doesn't exist anymore.
The CMO as Integration Layer
A CMO walks into the QBR with three windows open, mentally reconciling marketing's MQL number with sales' opportunity number with finance's pipeline number, hoping the maths she does in her head doesn't get challenged before she finishes the slide.
Then the meeting becomes a debate about whose dashboard is right instead of a conversation about what to do next.
AI Stack Inherited The Problems That Dashboards Had
Some of the CMOs or CROs killed these dashboards a year ago and replaced them with something more interesting - a stack of AI workflows. People are calling this GTM engineering. It's real, it works, and it's one of the most useful shifts in B2B marketing in years.
Clay Enrichment Chain
Automated data enrichment pulling signal from across the web into your CRM.
Custom n8n Agent
Stitched-together workflows that score accounts and route hot ones to Slack.
Agent on Sales Transcripts
A custom model/agent trained on your own sales conversations for in-context reasoning.
Signal Scraper → HubSpot
Pulls intent signal from the web and writes it back into your CRM automatically.
People are calling this GTM engineering. It's real, it works, and it's one of the most useful shifts in B2B marketing in years.
But that didn't solve the context problem. It just got moved a layer up, to another function/role.
You used to have this: | Now you have this: |
|---|---|
Fourteen dashboards that didn't share state. Each one a frozen slice of a motion that had already moved. | Fourteen workflows that don't share state either. The enrichment agent has no idea what the scoring agent decided last week. The Slack alert can't tell you which deal it actually helped close. |
You're doing prompt engineering, one workflow at a time. Which is exactly where AI teams were two years ago. The context layer underneath the workflows is what matters now.
Stateless to Stateful
What actually replaces the dashboard era, whether you're still on dashboards or you've moved to a workflow stack - is a living context graph that updates as the motion happens.
In a context graph, the question doesn't have to be predefined. The CMO doesn't have to know in advance which fourteen questions she's going to ask each quarter. She asks the question she actually needs answered, and the graph composes the answer from the underlying signal in real time.
Questions the graph answers in real time
How is the enterprise segment doing this quarter versus last?
Which campaigns are showing diminishing returns over the last thirty days?
Where is pipeline slipping that wasn't slipping a month ago?
What's our likely shortfall against plan if current conversion rates hold?
No pre-built dashboard required
None of these questions need a pre-built dashboard. None of them need a custom workflow you have to build, maintain, and explain to your team every time it breaks. They need a graph that already knows the data, the semantics, and the motion.
This is the same architectural shift the AI world has been making at the model layer. Systems that used to forget started remembering. Tools that used to work in isolation started coordinating.
Marketing measurement has been stateless for two decades. It's finally getting a memory.
Stop Reconciling. Start Compounding.
When marketing teams stop maintaining twenty dashboards or twenty workflows that each see only a slice of the picture, they get back the hours spent reconciling them, the time spent translating between teams, and the energy spent defending their numbers in meetings.
That time goes back into the work that actually moves pipeline. Building better campaigns, sharpening positioning, running more experiments, talking to sales about something other than dashboard discrepancies.
The CMOs who get there first will be running on infrastructure their competitors can't see yet. The gap will compounds quarter by quarter. Early movers will have real advantage here.
The Human Era Is Ending
The dashboard era was the human era of GTM. Humans were the integration layer, the reconciliation engine, the connective tissue carrying context between systems that couldn't talk to each other. The context graph era is the one starting now.
Whether you're still running QBRs out of fourteen dashboards or you've already moved on and you're running them out of fourteen AI workflows that don't talk to each other, the problem underneath is the same.
The teams that solve the context layer get the next decade. Everyone else feels the way prompt engineers feel today - late by a year, to a shift that already happened.
P.S. If you're deep in your own GTM engineering stack right now, running agents, stitching automations across five tools, training custom GPTs on your own data - you're closer to this shift than most. The hard part is making the workflows share a memory. That's the layer we've spent the last few years building at RevSure.
Reply directly or write to [email protected]. I want to know what you're stitching together and where it's breaking.