Skip to main content

Build vs Buy AI Agents for Product Marketing: What It Takes

Product marketing agents built in Claude or ChatGPT start fast and get hard later. Here is what each stage of the build asks of you, and when buying wins.

·11 min read

Short on time?

Summarize it with your favorite AI tool.

AI agents for product marketing are easy to demo. A battle card skill in Claude turns out a decent first draft in an afternoon, and LinkedIn is full of people who built a PMM agent in 15 minutes. So the build vs buy question comes up fast for anyone who leads product marketing: build your own agents in Claude or ChatGPT, or buy a product marketing platform that already has them?

You can build them. What’s hard to see from the demo is what it takes to make them work for a whole team, week after week. This guide follows a build in Claude through its three stages and shows what each one asks of your time and your team, then sets that against buying. Calven sells the ready-made option, so where a number isn’t sourced, it’s marked as Calven’s estimate.

If you’d rather watch than read, there’s a 12-minute video version of this guide. It builds the same skills library in Claude, shows how to set the skills up so the output is right, and tallies the hours it takes to keep them working.

Claude Skills for Product Marketing: The Unseen Maintenance Cost · Watch on YouTube

To share it with your team, the PDF below has the same material.

What does building AI agents for product marketing involve?

Building can mean two things, at very different levels of complexity.

No code: skills and agents in Claude or ChatGPT

You write custom skills and agents inside the AI tool you already pay for: skills and Projects in Claude, Projects in ChatGPT, scheduled tasks in both. Per Anthropic’s Agent Skills docs, a skill packages instructions and optional scripts that Claude loads when a request matches its description. This is where PMM teams start, and it’s what this guide covers. The examples use Claude, but the same skills and agents can be built in ChatGPT.

Code: vibe coding your own app

The second level is writing a real app, in JavaScript or similar, with AI coding tools and deploying it somewhere the team can reach. It can do more, but it means programming your own tool: even when AI writes the code fast, someone with an engineering background has to design, host, secure and patch it. Calven estimates a production version at 6 to 12 months for two engineers. In theory, a marketing team could take this route. In practice, the complexity and the engineering it requires put it out of reach: a standard marketing team has neither the engineers nor the time to build and run its own software. That’s why this guide leaves it out.

The three stages of a no-code build

A no-code build goes through three stages, and its effort curve looks nothing like the one LinkedIn promises.

The prototype, in weeks one to four, is the exciting part: the first battle card comes out well and everything looks possible. The build, from month two to about six, is where projects stall, on output quality, on the AI skills it takes and on time nobody has. Running it starts around month seven and never drops back to zero. Everything the 15-minute demo leaves out is the unseen work.

Stage 1: how long does a skills library take?

Weeks of a PMM’s time, by Calven’s estimate, across at least 21 skills. Finding skills isn’t the hard part. Plenty are published online, and Calven’s open-source marketing-as-code repo alone ships 99, among them battle card, competitor watch, ICP, persona, positioning, messaging and win/loss skills.

The hard part is quality. A skill encodes how one team works, with its own examples, sources and definitions, so one that worked for another company may not work for yours. You can’t download a library and start using it. Each skill has to be tested on your own competitors and deals, and rewritten where it falls short, before anyone relies on its output.

Covering the core of the job takes eight skills for competitive intelligence, four for win/loss, four for ICP and personas and five for positioning and messaging. On top sit two interfaces no skill can build, a Slack or Microsoft Teams app that answers people’s questions and an MCP server that lets every AI tool reach the approved work, and both are software.

In Claude, the finished library is a list under Customize.

Once the library exists, the skills wait. Custom skills on claude.ai are individual to each user, and a skill runs when someone asks for it. Scheduling closes part of that gap: Claude’s scheduled tasks run a saved prompt hourly, daily or weekly, skills included. But each schedule belongs to one person’s account. Five PMMs with five schedules produce five battle cards for the same competitor, and nobody has approved any of them.

A finished library buys a faster PMM, with quicker research and more consistent output. It doesn’t transform the team: no agent flags what needs doing, output and quality take no step change, and the hiring plan stays the same.

Stage 2: where does the build stall?

In the prototype, nobody reads closely. The battle card looks right at first sight, and the excitement is that it works at all. The build is where you go into the details: check every claim against its source, run the same skill on five competitors, compare this week’s output with last week’s. That’s where the problems show up: output changes from run to run, claims have no source, and the AI can’t see your approved material. Most of them trace back to one cause: what you built is a skill, not an agent.

Claude skills vs agents: what’s the difference?

A Claude skill is packaged instructions, sometimes with scripts. An agent is that skill plus the machinery around the model, the harness, that lets it work without being asked and catch its own mistakes.

Claude Cowork gives you part of the harness. Claude runs the loop itself, planning, acting and checking; connectors reach the CRM, calls and docs; plugins bundle skills, connectors and sub-agents; and scheduled tasks run without being asked. What it leaves to you is the part that makes the output dependable for a team: evals that catch bad output before a rep sees it, guardrails on sources and cost, permissions on who sees what, and monitoring that alerts when a run breaks. Anthropic’s engineers make the same point about evaluating agents: “we’re evaluating the harness and the model working together”. Building the rest takes engineers, which turns a no-code build into the app this guide set aside.

Why doesn’t the AI know your approved assets?

A common assumption is that once the skills write the battle cards, ICP and messaging into a shared folder, everyone in the company has them. That’s only true for people who reference those documents in their prompt. Ask Claude about your competitive position and it says it has no access to your competitive materials. Each person can connect Drive or add files to a Project, but those are separate copies, none of them the approved version. Unless one approved source reaches AI tools through MCP, it gets used only when someone remembers to paste it in.

HubSpot, Salesforce and Gong all offer MCP servers that let Claude read deals and calls once each person connects them. None serves your approved battle card or messaging. That takes an MCP server of your own, which is code.

Can you trust data the AI collected itself?

Another big issue with custom skills is that you rely on the AI to collect the data, not just to analyze it. You can’t trust that data without checking every line: the skill searches whatever it can reach and fills the gaps with guesses.

The AI may list the pages it used, but the links often lead to pages that don’t exist. Nor does it say which source backs which section, so you can’t check a single claim.

What else turns up as you build?

As the library grows, more gaps appear. Some need workarounds, and some no skill can close.

  • Monitoring runs on one person’s schedule. A scheduled task checks competitors only as often as its owner set it to, and only on that owner’s account.
  • Win/loss interviews still need a person. A skill can draft the questionnaire, but someone has to reach the buyer, run the interview and write it up after every closed deal. That’s why win/loss is its own software category.
  • Not every tool has an MCP connector. MCP is becoming the standard for CRMs and call recorders, but some don’t offer one yet, or expose only part of their data. Without it, a skill can’t read or summarize your call transcripts, so check the tools you use.
  • Updates rewrite the whole document. Ask the AI to update a battle card and it writes a new one, so you can’t see what changed, and the meaning drifts a little with every rewrite.
  • Nothing answers in Slack or Microsoft Teams. A skill runs only when someone opens Claude and asks. Answering a rep’s competitive question in the channel, on its own, takes a chat app, and that’s software.

Using ChatGPT or Claude for CI shows the same ceiling for competitive intelligence.

Stage 3: what does it take to keep it running?

By Calven’s estimate, 15 to 28 hours a month, about 10 to 20% of one PMM’s time, for as long as the system exists. That’s roughly half a day to a day every week spent keeping the skills working instead of using them.

WorkHours a monthWhy it never stops
Quality drift4 to 8Output slides as the market and product move
Model updates4 to 6Every skill is re-tested when the model changes
Broken sources2 to 4Sites redesign, links break, APIs change
Sales feedback2 to 4Fix the skill, not just the one answer
Security patches2 to 4The chat app and MCP server need updates
Usage costs1 to 2Someone watches the model bill
Total15 to 28About 10 to 20% of one PMM

In a no-code build, that upkeep falls on a PMM, out of the hours meant for positioning.

When should you buy instead?

If there’s no budget at all, building is the only option left: one PMM writes the skills and uses them for their own work. Set the expectations clearly, though. It’s a local optimization, not a revolution: one person gets faster, but the team works the way it always has, so it won’t make a material difference. It’s a patch, and it won’t fix the underlying problem. When the team needs agents that work without being asked, one approved version every AI tool reads and a source behind every claim, buy.

What you get with Calven

Calven is an AI product marketing platform with nine agents across research, audience, and positioning and messaging. The agents are autonomous: they start working on day one, with product marketing best practices built in, and keep going without anyone asking.

How Calven works. Click the diagram to see it in full.

Each gap from the build has an answer already in place.

The gap in a no-code buildWith Calven
Writing and testing every skill114 skills, written and tested, with the PMM frameworks built in
Monitoring on one person’s scheduleThe Competitive Intelligence Agent watches competitor websites and docs for the whole team
Win/loss interviews by handThe Win/Loss Agent runs AI voice interviews or surveys after each closed deal
Data the AI collected itselfData from paid providers and defined sources; the AI only analyzes it, and each claim cites its origin
Whole-document rewritesEach change lands in the right section; its owner approves it or lets routine updates publish
Approved assets the AI can’t seeCalven MCP serves the approved ICP, positioning and messaging to Claude, ChatGPT and any MCP client, and answers arrive in Slack
Half a day to a day a week of upkeepModel updates, broken sources and security are Calven’s job

It connects natively to Salesforce, Gong, Chorus and Zoom, and Calven Professional costs $9,600 a year for the whole team.

Build the skills that express your positioning, and buy the system that keeps it current. A 14-day trial needs no credit card and no sales call. Run it on the three competitors you lose to most, and compare its battle cards and the hours they took with what your skills produce.

Frequently asked questions

What's the difference between Claude skills and agents?

A Claude skill packages instructions and optional scripts that Claude loads when a request matches. In Claude Cowork, a skill can work much like an agent: Claude plans and checks its own steps, connectors reach your CRM and docs, and a scheduled task runs it without being asked. What Cowork leaves to you is what a team needs to rely on it: evals that catch bad output, guardrails on sources and cost, shared permissions and monitoring, and building those takes engineers.

How long does it take to build AI agents for product marketing?

By Calven's estimate, a prototype in Claude or ChatGPT takes a few weeks, and the build that makes it reliable runs from about month two to month six, where quality, data and agent tooling take the time. After launch it needs 15 to 28 PMM hours a month to keep running.

Should we build our own competitive intelligence tool with Claude?

Build Claude skills when one person needs faster battle cards; a skill drafts them well. A tool the whole company relies on, with continuous monitoring, sourced cards and one approved version, needs agents, an MCP server and a chat app around the skills, which is engineering work. Calven's comparison of ChatGPT and Claude for competitive intelligence covers the narrower question.

Is it cheaper to buy or build?

Count the maintenance hours and compare them with the price. A no-code build takes 15 to 28 PMM hours a month to keep running, by Calven's estimate, while Calven Professional costs $9,600 a year for the whole team. A trial on your own competitors lets you check that against your own hours.

Can we keep our Claude skills if we buy Calven?

Yes. The skills a team already wrote keep working through Calven MCP, which serves the approved ICP, positioning and messaging, with sources, to Claude, ChatGPT and any MCP client. Each skill then reads the current approved version instead of whatever someone pasted in.

David Kolinek

David Kolinek is the co-founder and CEO of Calven. He spent nearly a decade at Ataccama, a B2B data management company, rising from product design to VP of Product and then VP of Product Marketing, where he lived the gap between the strategic work PMMs sign up for and the tactical grind that replaces it. He writes about product marketing, competitive intelligence, and how small teams put AI to work without the busywork.

Connect on LinkedIn