The Jagged Go-To-Market Frontier
Why AI is bad at GTM tasks, and how to fix it.
AI is proving what it can do, domain by domain. Engineering teams are compounding on it every quarter. GTM teams are running the same models and waiting. Before deciding what to do about that, look at where the capability actually is.
It's not the models. It's us.
If you understand what these models are good at and how they actually work, AI absolutely works for go-to-market. The teams proving it aren't running better models. They're running better systems.
To see what that system has to provide, start where AI already works.
What engineering got for free
Pick a harness: Cursor, Replit, Claude Code. The pattern is the same. A language model gets dropped into a codebase, handed meaningful work, and produces useful output within minutes. Two conditions make that possible, and neither is the model. Everything below turns on them.
Context Availability
What the AI can see and reason against. An engineering agent lands in a codebase: a complete, versioned description of the system it's about to work on. Everything it needs is already there, in a form it can read. Hold onto this one, it's where go-to-market breaks first.
Task Clarity
How well-defined the work is. Code compiles or it doesn't. Tests pass or fail. The work arrives in discrete, reviewable units, and the agent gets a verdict on its output in seconds. There's always a check at the end.
Engineering has both
Drop a model into that structure and it compounds. Engineering didn't redesign itself for AI; AI fit the shape engineering work already had. That's the upper-right corner, and it's where the results everyone quotes come from.
GTM has neither
Run go-to-market through the same two conditions and it lands in the opposite corner. Not because the people are worse. Because this structure was never built. Both axes have to move, and no one is going to hand you either.
Both conditions are buildable in go-to-market. Nobody hands them to you, and no tool purchase substitutes for them. The rest of this page walks the build, starting with the condition that fails first.
Context Systems for GTM
Time to peel back the curtain: AI wrote most of this article, and AI built most of the page you are reading. It could do that because this article sits on top of a pile of concentrated research, opinionated synthesis, and structured human thinking. That pile is live, beside this text. Every concept this piece leans on, every source it cites, every position it takes is a node the writing system can read.
That is a context system. AI cannot do useful work in GTM without one that is well designed and fit for purpose.
Run GTM through the first condition
GTM context lives in Google Docs, sales decks, Slack threads, half-finished playbooks, voice notes, and the heads of senior people. Even when the data is searchable (Glean ingests every file in your stack) search isn't structure. The team's ICP isn't a file the model can read. It's an interpretation the team holds, sometimes disagrees on, and propagates informally.
Here is the asymmetry that matters. For AI to work in go-to-market, treat context like a codebase. In coding, the codebase is the context. In go-to-market, the context is the codebase you have to write. The codebase is already there for the engineering agent, and richer context makes a capable agent better. Go-to-market has no equivalent. Strip out the customer, the product, the motion, and the brand, and there is nothing left for the model to be capable at. When you ask an LLM to help with a GTM task and it has none of that, it guesses. If you've gotten good GTM output, it's because you wrote that codebase yourself: the long prompt, the Claude project, the configured tooling, the strategy doc you fed it.
And the second
Task clarity fails next. What's the best headline for the homepage? Which feature should you demo first on this call? GTM roles cover whole functions with individuals owning fuzzy outcomes rather than discrete deliverables. There isn't a clean spec the model can read its way to, and there isn't an automated check at the end. The feedback arrives in pipeline numbers, weeks later, noisy. Until the work is decomposed into units an agent can be clear about, the AI is guessing and so is everyone reviewing it.
Your GTM org already runs an informal version of the codebase, holding and transmitting its strategy through people and documents. The sooner you make it explicit, the sooner AI starts producing work your team will ship. The graph beside this section is ours. The next section shows what it produces.
Taste, Tasks, & Tokenomics
This page is the demonstration. Everything above was produced by AI agents working inside the context system you just saw, against a brief a human marked up, under rules a human locked. Here is the functional process, exhibits included.
It starts with taste. The design brief for this page is an annotated PDF: a human deciding what the argument is, where the punchline lands, what the reader should feel at each scroll position. Nothing downstream works without that document being opinionated.
The brief
Page one of the annotated campaign brief: the marked-up intro spread with the scroll-build notes and the punchline annotation visible.
Export from the working PDF; keep the annotations raw. The mess is the point.
The brief lands in a build system that treats this page like a product. The layout is a toolkit of reusable row components; the words live in the source file as data; the charts are drawn from data in the same file. A change to the copy touches copy alone. Locked language stays locked because the system enforces it, not because everyone remembers.
The build session
A working agent session mid-build: the row components in the editor, the rendered page in the browser, the agent transcript composing them.
Capture during a real revision round, not a staged one. Component filenames legible.
The verification pass
The visual-verification contact sheet: this page at four widths and multiple scroll states, with the overflow tripwire output.
npm run shoot output. Shows the page has a pass/fail, like a test suite.
The art runs through the same discipline. The orange you scrubbed through at the top was generated in our image studio from brand seed plates, tuned across a session, then re-encoded frame-by-frame so it scrubs in reverse. Brand images here are produced by a routine, not commissioned per asset.
The art pipeline
The Studio session that produced the orange: seed plates, the shot type configuration, and the output grid with the selected render.
Show seeded generation from brand plates, not raw prompting. The point is repeatability.
Name what just happened to the two conditions. The context system supplied what the AI needed to see. The brief, the row toolkit, and the verification harness decomposed fuzzy creative work into tasks with acceptance criteria. Taste stayed human and became enforceable. That is task clarity, manufactured where it didn't exist.
And the economics hold up. Producing a flagship interactive piece this way costs a fraction of the agency equivalent, and the second piece costs less than the first, because the system is the asset that compounds.
What this page cost
A production ledger: agent hours, human hours, and token spend for this page, lined up against a comparable agency engagement.
Fill with real numbers from the session logs before publish. If the real numbers are unflattering anywhere, print them anyway and say why.
Really good briefs. Systems and tools that support people and agents in production. Clarity across brand, offering, and message. Treat GTM work as systems and product work, and you get product-quality outcomes.
It's hard to change how a complex GTM org runs
If you lead a GTM organization and this page reads like an indictment, it isn't. You're hitting a number this quarter. The product is shifting under you. The team is at capacity, the market moved twice since planning, and every vendor in your inbox is promising transformation by Tuesday. Where exactly is the time to rebuild the operating model supposed to come from?
Just because we've demonstrated what's possible doesn't mean every GTM org can run this way right now. Nobody serious is claiming otherwise. The honest version: the cone of uncertainty between "we ran a pilot" and "this is how we operate" only narrows by operating, and the first stretch of that road is the hardest to walk while carrying a quota.
But the turn is happening, in public. Kyle Norton's team at Owner.com books three times the revenue per AE of any team he has previously run, on AI-infused motions built exactly this way. Tom Wentworth's five-person team at incident.io runs a full enterprise marketing stack. Teams that solved the two conditions first are posting the engineering-style numbers, and none of them started with slack in the calendar.
...but we can help.
This is why Fresh Context exists. We help large go-to-market organizations build the two conditions this page demonstrates: the shared context your AI works against, and the task structure that makes the work compound. Diagnose, architect, operate, with your team, on your number.
Citations
Dell'Acqua, Mollick, et al. "Navigating the Jagged Technological Frontier." HBS Working Paper 24-013, Sept 2023; Organization Science, March 2026. 758 BCG consultants, GPT-4. Framing extended in Ethan Mollick's writing. The mechanism is durable; the specific statistics carry a 2023 vintage.
Nobel Prize in Chemistry 2024. Demis Hassabis and John Jumper, for AlphaFold. NobelPrize.org.
Project Glasswing. Anthropic, April 2026: 10,000+ high- and critical-severity vulnerabilities surfaced in the first month across core open-source infrastructure. anthropic.com/glasswing, initial update.
Disclosure. Sam Gong leads marketing and AI go-to-market transformation at WorkSpan. Fresh Context productizes that experience. AI drafted and built this page inside the systems it describes; a human set the argument, marked the brief, and locked the language.