Build your own AI marketing harness with Claude — or buy one? A decision framework (why buying the pre-built core wins)
A clear build-vs-buy framework for your AI marketing stack: what real builders spend, the 90/10 rule, and why buying the pre-built harness wins.
The short answer
Build the part of your AI marketing stack that only you can build. Buy the part that is identical for everyone.
The part that is identical for everyone is the core layer: live connections to Google Ads, Meta Ads, GA4, Search Console and the rest; the scheduler that runs analyses on a clock; the approval gate before anything touches a live account; and the ledger that records, for every change, what was predicted and what actually happened. Every serious AI marketing setup needs that core. Very few budgets should re-build it from scratch.
The part that is genuinely yours is the differentiated 10%: the specific way you phrase a recommendation, the reports your team actually reads on a Monday, the channels and campaigns that matter to your business. That is where building — with Claude Code and your own tools — pays off.
This page is the anchor of our build-vs-buy series. If you read nothing else, read the 90/10 rule and the one-sitting checklist.
Why this question is suddenly everywhere
A couple of years ago, "build an AI marketing agent" meant either wiring together a dozen SaaS tools or waiting for a vendor to ship it for you. Neither felt like a viable option for a small team.
That changed. Today, you can open a terminal, install Claude Code and a few open-source MCP servers, and in an afternoon have Claude reading real ad-platform data and drafting recommendations. For the first time, "we'll just build our own" is a serious option — so serious that it is now the default question in every growth-ops conversation: should we build this, or buy it?
The risk is that "build it" quietly, week by week, becomes "maintain it." The incremental work that turns a working prototype into a production system is large and undramatic: OAuth token refreshes, rate-limit handling, schema drift, retry semantics, a scheduler that survives pod restarts, an audit trail. None of it is hard to describe. All of it is time. The honest cost of that time is what this decision is really about.
The honest cost of building it yourself
Let's be specific about what "building your own AI marketing harness" actually means in practice, because the gap between the demo and the production system is where the cost hides.
The core layer is a lot of unglamorous work:
- Live data connections. Connecting to Google Ads, Meta Ads, LinkedIn, GA4, Search Console, your warehouse — each one is its own auth flow, its own rate limits, its own failure modes, its own schema that drifts over time. One integration is an afternoon. Wiring up the set your marketing actually depends on, and keeping it alive, is a standing expense.
- A trustworthy scheduler. Analyses that run on a clock, not when somebody remembers to ask. This means handling long-running jobs, resumability when a worker restarts, and not silently stalling.
- An approval gate and an audit trail. Nobody wants an agent changing budgets at 03:00. Changes need a human yes, a log of what was approved and why, and the ability to see, months later, why a change was made.
- Outcome measurement. Every shipped change followed up: did the metric actually move? This is what turns recommendations from vibes into a system that gets sharper about your account over time.
The conventional framing treats this as a weekend project. The realistic framing is that the connections, scheduler, approval flow and measurement are most of the engineering — and they are the part that is identical for every team that builds one. Re-building that core means spending your engineering on the 90% that is the same for everyone, before you ever get to the 10% that is different for you.
The 90/10 rule
Here is the heuristic worth taking to your team:
The 90/10 rule: about 90% of an AI marketing harness is a commodity — live connections, scheduling, approval, measurement, memory. That core is the same for every builder. The remaining 10% — how your team words recommendations, which reports it reads, which channels it prioritises — is what is actually about your business.
Spend on the first 90% by buying or renting it, and spend your own building on the last 10%.
The reason the split matters is compounding. A commodity core does not get worse if you don't maintain it — someone else maintains it, everyone benefits from the shared improvements, and the integration set grows. A bespoke core gets more expensive every quarter: every platform upgrade, every API sunset, every auth refresh is on you. That is the difference between a system that compounds and one that slowly becomes a job.
The "buy the core, build the edges" shape is not controversial once you name it — the argument over where the line sits is the productive part. And the line has moved. Historical automation platforms were the "buy" because everything was hard to build. Now that an LLM hands you the reasoning layer almost for free, the commodity that is worth renting has narrowed to the pre-built MCP and agent harness around it.
What "buy the harness" should mean
Buying a harness should mean buying the core layer as a working system — not buying a "black box" that hides the reasoning from you and makes promises you can't check.
The practical definition of a harness you would want to buy:
- You bring your own model, or the harness runs one you can change. The harness should not lock you out of the 10% that is yours. You want to keep control of the reasoning and the wording.
- The connections are already wired. Not "connect it yourself, here's a tutorial" — the auth, rate limits and schema handling done, for the platforms your marketing actually depends on.
- Every change gets a verdict, misses included. The harness must measure what was shipped and record a result — worked, didn't work, or inconclusive — rather than celebrating only the wins. That is what makes it possible to trust the system with real budget.
- Humans approve ship decisions. The harness surfaces recommendation and drafts; your people decide. Nothing touches a live account without a human yes, and the approval is logged.
- It is yours to leave. If buying the core doesn't work, you should be able to take the differentiated 10% you built and move it anywhere.
That last point is why a harness that only works if you stay a customer is a trap. The reason to buy the core is not lock-in — it is that the core is not what creates the value for your business. You want the parts that are yours to stay portable.
This is the shape Cogny is built around: pre-built MCP connections for the channels you run, scheduled analyses, an approval gate, and a verdict ledger that records what was shipped and what happened — all wrapped around an agent you can drive yourself. Your people decide. The 10% that is yours stays yours.
The one-sitting checklist
If you are making this decision today, run through this in one sitting:
- Name your channels. List the platforms your marketing actually depends on (Google Ads, Meta, GA4, Search Console, your CRM or warehouse). Not "someday" — the ones in your weekly workflow.
- Estimate the core. For each, write down the effort to connect it live, keep it running, and handle auth and rate limits. Then add scheduling, an approval gate and outcome measurement on top. This is your true "build" number.
- Apply the 90/10 rule. Ask: is most of that list identical to what any competent builder would produce for another business? If yes, it is candidate-for-buying, not candidate-for-building.
- Protect the 10%. Write down the few things that are genuinely yours — your report reading habits, your recommendation style, your priorities. Keep those in your control.
- Check who measures. Whatever you build or buy must tell you, for each shipped change, whether it worked — a verdict, misses included. If a system cannot show you its misses, you cannot trust its wins.
- Decide on a month of data. Commit to a small, time-boxed experiment — a free start, a low-tier plan, or a single hand-built script — and let the measured outcome, not the demo, make the call.
The point of the checklist is that the decision should be made on the cost and the measurement, not on enthusiasm for the demo. A demo shows you what is possible. The ledger shows you what actually happened.
The verdict-ledger proof
Every harness needs to be able to show its own record — wins and misses — because that is the only way to audit whether "the harness works" is a claim or a fact. A system you cannot audit is a system you cannot trust with real budget.
That is why the record being public is the right test. At Cogny, every change is measured and logged: for the changes tracked in the verdict ledger, the outcome is recorded and published — worked, didn't work, or inconclusive — with the misses included. The principle behind it is simple: no system gets it right every time, and the one that tells you when it didn't is the one you can learn to calibrate. The verdict-ledger proof is not a guarantee of a specific result for your business; it is a guarantee of transparency about the results that did happen.
If you are deciding between building or buying, ask the same of whichever you choose: show me the record, misses included. A build you audit yourself and a harness that publishes its ledger both clear that bar; a tool that only shows you the wins does not.
The bottom line
Building your own AI marketing harness with Claude is a legitimate path — and the right one, for the part that is yours. The mistake is building the 90% that is a commodity and identical for everyone, and letting maintenance become the job.
Buy the core: live connections, scheduling, the approval gate, and a verdict ledger that records what happened, misses included. Build the differentiated 10% on top of it, the part that is actually about your business. That is the decision framework, and it is the one we hold ourselves to.
Cogny exists to be the pre-built core you can drive with your own AI — connect your channels, see the plans on our pricing page, and ship a change to see it measured.
Start free — no card required.