What Is an AI Harness for Go-to-Market?

An AI harness for go-to-market gives AI a direct, codified pathway to the knowledge, tools, and ways of working it needs to operate inside your business. It connects the strategy leadership has decided to the work people and agents actually do, with checks on what they can change, say, and send out into the world.

I just spent a week with business leaders at Pavilion’s GTM Summit talking through their biggest AI challenges. A lot of them have moved past unsure, past inspired, and into activated. They have built something. Now they need other people to be able to use it.

That is where the conversation gets interesting. We need to be able to orchestrate. We need to be able to codify strategic thinking. We need to be able to add decision-making layers.

Our marketers, our leaders, our CMOs, and our CROs need to get very clear on the levers of control and strategic thinking they need to be putting into their AI systems.

The smartest person in the room still needs context

I use this example a lot. You could find the smartest person in the world and drop him into the front office of P&G. Homeboy won’t be able to find the bathroom.

He does not know who the people are, what the processes are, or who’s in charge of what. He’s still the smartest kid in the room. We still have to give him scaffolding.

We teach him how to look at problems. We give him access to the right tools, capabilities, and data. We give him awareness of the procedures, policies, and people he needs to work with.

AI needs that context, too. An intelligent model does not arrive knowing how your business competes, which claims it can make, or when a person needs to take over.

That is why I don’t think anyone in your organization should be operating business AI in an unharnessed environment.

What the harness actually does

In its most rudimentary form, you can start with a controlled file system and a clear instruction: when you come in here, you have to look at this knowledge.

When I say repo, think of a folder with files in it, with a way to track changes and share the work. We’re going to keep it super basic. We’re not going to freak anybody out.

The folder holds the business’s knowledge and instructions. The software running the AI has to use them and enforce the boundaries. As the work gets more complicated, that includes selecting the relevant information, making tool calls, keeping track of the task, checking the work, and stopping for a person when needed.

That distinction matters. A rule written in a file and a system that enforces the rule are different things. In the technical documentation, a harness is the software around the model that manages this execution, including tools, permissions, and the state of the work. OpenAI’s harness documentation describes these responsibilities.

For a GTM leader, the practical question is: what does that environment need to contain so the work represents our business?

Here is how I think about it, starting with marketing and connecting it to the wider buyer journey.

Explore the architecture below. Follow the system from top to bottom; the diagram fits the page on desktop and mobile. Open the full-size diagram.

Leadership owns the direction. People and agents work from the same foundation. The harness controls access, execution, and the checks along the way.

Give the system a foundation it can work from

Strategic knowledge with a clear hierarchy

In the past, we said: dump in everything your AI needs to know about you and have it figure it out. It’s smart, right?

That is the equivalent of giving someone the dictionary and saying, go be a marketer. You’ve got all the words. I don’t know what you need from me.

It’s funny, but it’s what we’re doing.

Not all ideas are created equal. Your core brand positioning, when executed properly, touches every touchpoint of your marketing. What you learned about one Facebook ad has a very different implication.

Those ideas don’t carry the same amount of strategic structural weight. Your knowledge has to be structured intentionally.

I split it into two sides of the same coin:

• Core strategic knowledge: positioning, ICP and audience, products, points of parity, points of differentiation, reasons to believe, pricing and offers, and brand voice.
• The reference library: the deeper material that supports those ideas, including case studies, customer conversations, research, and product detail.

The core should be distilled. As short as it can be while still being correct. It progressively discloses, meaning it points to other layers of thinking and evidence when the task needs them.

Now we’re talking reasons to believe. Now we’re talking case studies. This person’s in this industry. Here’s the relevant case study.

You have given the AI a path through the thinking. Leadership owns the core, decides who can edit it, and gives the rest of the team access to work from it. When the strategy changes, there is a shared place to change it and a defined way for subsequent work to pick it up.

A design system that makes people feel something

The visual design system is how your brand is visually and emotionally experienced by the world.

Most people think design systems are a set of colors and fonts. Design systems have stories. Design systems have emotion. Design systems have choices that make them different from something else.

You can codify those stories and choices into the same environment that holds your knowledge. Give it the visual idea behind the brand, the photography and illustration direction, typography, layout, reusable components, and examples of what belongs and what doesn’t.

Then, when AI creates assets on behalf of your company, it has a much more meaningful basis for how those assets should look and feel.

This is where I use the deck example. I don’t need the deck skill to say make it blue, make it purple, put this over there. The design system supplies that direction. The skill can focus on the success criteria: use visual storytelling and make the first, second, and third most important things easy to understand.

You start to get good at understanding what lever needs to control what.

The website as the place your thinking comes to life

A lot of people see AI building a website and think about how quickly they can post more content.

The nuance is that your website plays a very unique role. It is where your core strategy, your design system, and the way you frame your company’s value proposition come to life.

It is the database of that thinking.

When I say that, I mean it is a concrete, reusable expression of the brand: the arguments, the product explanations, the visuals, and the way you help a buyer understand the value.

Make that approved material accessible to the system. A proposal can draw on an existing explanation of the product. A deck can reuse the way you’ve made a complicated idea visible. A campaign can start from an argument you’ve already worked through.

You can keep the website’s source alongside the other files, or give the system access to the approved material another way. What matters is that the work can use it, and that published copy remains subordinate to current strategy and approved claims.

Production pipelines that carry the thinking into the work

Once that foundation is there, build the production pipelines around it.

This is how we as a business research, ideate, evaluate, and produce for a particular channel. Where does a person make a decision? What can AI do? What needs to be checked before it leaves?

For a campaign, that might mean agreeing the buyer problem and angle, finding the evidence, creating the assets, reviewing them together, and approving publication. For a buyer conversation, the playbook needs to define what to discover, what establishes fit, and when the person should reach a seller.

The methods differ. Both should work from the business’s approved knowledge.

This is also where skills belong. A skill packages a repeatable method and the resources needed for a particular job. Anthropic’s explanation of Agent Skills describes how those resources can be loaded when needed.

But every skill should earn its place. Give AI the same task with and without the skill. Have a separate reviewer compare the work against the same success criteria. Did that skill make the deliverable better?

If it did, keep it. If not, remove it or change it. When the model changes, retest. A library of instructions needs maintenance.

Connect the systems where the work already lives

The foundation needs access to the information and capabilities required to do the job.

For GTM, that might mean the CRM for account history, shared files for approved material, email for a draft or an authorized send, and the website’s publishing system for an approved update.

A connector gives the AI a way to work with one of those systems. You still need to decide what it is allowed to do. Reading a customer’s record is different from changing it. Drafting an email is different from sending it.

Start with a real workflow and map which systems it needs, which information it can read, and which actions need permission.

Then look at what you’re paying other software to do. If the harness can reliably handle a workflow that previously needed a separate subscription, test that replacement. Include the checks, the exceptions, and who maintains it before deciding to retire the tool.

Put the checks into the system

I call the combination of LLM reasoning and software rules Spanglish. Some parts need the model to think. Other parts need software to enforce a boundary.

Think about bowling. Those are the bumpers on the side that allow the LLM to proceed without relying entirely on another instruction asking it to stay in the lane.

Take pricing. If a proposal doesn’t articulate the current pricing package correctly, it should not proceed to publishing or sending. The system should stop it and point the person back to the approved offer.

You need to define who may change the underlying pricing, what evidence a claim needs, and what approval an outgoing action requires. Those controls belong at the relevant action, including tool calls during the work.

For the quality of the writing, give a separate reviewer a voice rubric and examples of your work. I call that adversarial QA. We never want the AI that does the work to be the only judge of whether the work is good.

A separate reviewer can still miss something. Software can enforce that approval exists without proving every sentence is correct. Decide which failures can go back for revision, which need a person, and when the work must stop.

The thing about loops is they’re expensive. You need a stopping point, too.

Follow a discovery call through the system

Here’s an illustrative version of the proposal example I showed people at the summit.

You finish a discovery call and ask the system to build the proposal.

It uses the call to understand what this buyer is trying to solve. It reads the core positioning and current offer. It follows the reference links to evidence relevant to the buyer’s situation.

The design system guides how the proposal looks and feels. The approved website material supplies product explanations and assets. The proposal skill defines what a useful proposal needs to communicate and how to make the priorities clear.

Then the work goes through the required checks. Does the evidence support the claims? Does the pricing match the approved offer? Did we actually address what the buyer said? Does it sound like us?

The seller reviews the proposal and the email. Only after the required approval can the connected email system send it. If something is missing or outside the rules, it comes back to a person.

That is what it means for these pieces to collaborate. You can see how the same foundation supports marketing, a buyer conversation, and the seller’s next move.

Welcome to multiplayer mode

When you build something for yourself, you carry a lot of mental scaffolding around it. You know what you meant. You know where the information lives. You know when something isn’t supposed to go there.

Other people don’t have all of that in their heads.

When you build across teams and functions, people who weren’t in the room need to understand the system. They need access to the right material, clarity about who can change it, and a way to see where the work stands.

Give the organization a shared library of knowledge and skills. Let people propose improvements. Test a new skill for whether it improves the work and whether its cost is justified before promoting it for everybody to use.

Leadership can maintain the core strategy. The team can improve how the work gets done. Results and corrections can inform those changes without automatically rewriting the business’s positioning.

That is also why I care about portability. Keep the business’s knowledge, design, and methods in a form the business owns and can maintain. Changing the model should not mean starting your strategic thinking over. It does mean checking the connections, permissions, and outputs again.

Full compatibility across providers is work. You don’t need to build every possible version on day one.

Where I would start

Pick one meaningful workflow, such as turning a discovery call into a proposal. Write down what the AI needs to know, which systems it needs to access, what good work looks like, who can change the rules, and who approves the result.

Build the smallest version that lets another person do that work without you standing next to them. Try it on real examples in a review environment. Inspect the output, the actions, and where the person still has to fill in missing context.

For a tool you use yourself, I’d often build it. As team size, complexity, and QA requirements go up, we get closer to bringing in specialists or buying maintained capabilities. These systems take maintenance. They take evolution.

The model supplies reasoning. An agent uses that reasoning and tools to pursue a task. A workflow supplies a defined process. The harness is what runs that work with the context and controls you’ve given it. You can start with one agent and a straightforward workflow; more agents are not a requirement. Anthropic’s architecture guide makes the distinction between predefined workflows and agent-directed work.

The art of go-to-market is orchestrating a buyer’s journey such that their problems line up with your unique solution. The system should help us do that work well.

I don’t want people stuck looking at the output and missing the system. The organizational structure in which humans and AI collaborate across the path to purchase is changing. Our job as leaders is to put the strategic thinking into it.

Sources

Adapted from my Pavilion GTM Summit walk-and-talk, the live roundtable, and my conversation with Nick Brunker about AI education for marketers. Technical references are linked where used.