The real cost of building your own AI marketing agent with Claude Code
The honest total cost of ownership for building a marketing agent with Claude Code: engineering hours per integration, drift, model churn and security — versus the flat pricing of buying the harness. Numbers-first, estimates clearly labelled.
Every build-vs-buy conversation has the same shape. The build side shows you a two-minute demo: Claude Code, a couple of MCP servers, and the AI is suddenly reading real ad-platform data and drafting recommendations. It looks free, because the cost is your own time.
The buy side is where the uncomfortable question lives: what does that prototype cost to run, not to build? This page puts a number on it. The numbers below are clearly-labelled estimates — your mileage will differ, and the point is the shape of the cost, not the precision of a single row. The build is cheap exactly once, and expensive every quarter after.
This is a companion to our build-vs-buy decision framework, which is where the 90/10 rule and the full checklist live. This page goes deeper on the one thing that framework leaves to you: the honest total cost of ownership.
The demo is not the cost
Let's be precise about what "building your own AI marketing agent" actually means. The demo captures the easy third: a model, a prompt, a readable connector. The cost hides in everything around it — and that surrounding work is recurring, not one-time.
- Live data connections. Connecting to Google Ads, Meta Ads, LinkedIn, GA4, Search Console, your warehouse. Each one is its own OAuth flow, scopes, consent screen, token format, rate limits and failure modes — and its own schema that drifts over time. One integration is a genuine afternoon of careful work on a good day.
- A scheduler that survives. Analyses that run on a clock, not when somebody remembers to ask. That means handling long-running jobs, resuming when a worker restarts, and never 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 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 into a system that gets sharper about your account instead of guessing.
None of that is in the hype. All of it is in the build.
What the build actually costs (estimates, clearly labelled)
The table below is a transparent model, not a promise. Think of each row as a first-year, one-time estimate plus a standing yearly upkeep — the two columns that matter for a total-cost-of-ownership decision:
| Component | First build (est.) | Ongoing upkeep per year (est.) |
|---|---|---|
| One platform connection (auth, rate limits, base schema) | 2–6 hours each | 4–10 hours to keep it alive through API changes |
| Schema drift handling (per platform) | part of the original build | 6–12 hours per platform per year |
| Scheduler (cron, resumability, leases) | 8–20 hours | 4–8 hours of babysitting and recovery |
| Approval gate + audit trail | 6–12 hours | 3–6 hours to keep rules current |
| Outcome measurement / verdict logic | 8–16 hours | ongoing as you adjust what you grade |
| Security (token storage, least-privilege scopes, review) | 4–10 hours | 4–8 hours a year, plus scare-driven patches |
| Model-churn migration (fold in the new model each time one ships) | — | 4–16 hours per model switch, on top of the rest |
Read those right to left and the picture inverts. The one-time build is real but bounded — call it a few days to a couple of weeks of concentrated time. The upkeep is where the bill compounds, and it never stops. The demo you saw on Tuesday is a recurring subscription you pay with hours, not dollars — and it doesn't come with a statement.
The model-churn row deserves a special note. Marketing agents named for a single model are obsolete on a schedule you don't control. The moment a better model ships, someone has to re-test the prompts, re-validate the tools against it, and re-check the output quality against a live account. That's real, repeated, unglamorous work — and it's the reason "built on the latest model" is a maintenance burden, not a feature.
The costs nobody itemises
Three more costs don't show up in a time sheet, but they are the ones that decide whether the build succeeds or quietly dies.
Drift is a tax, not an event
Marketing APIs change on someone else's schedule. Platforms retire endpoint versions, change response shapes, and deprecate fields — often without warning. A connector that worked in January is silently wrong in July. The uncomfortable truth is that a hand-built agent is only as current as the last person who touched it, and the person who built it is usually the person too busy to maintain it. Drift is what turns a working system into a job.
Security is a standing expense
An agent that can read — and sometimes write — your ad accounts holds real credentials. Doing that responsibly means least-privilege scopes, careful token storage, and a review process before anything touches a live account. It's not hard to describe; it's just more time, plus the occasional scare that forces you to drop everything and patch. The downside of getting it wrong is not a headline — it's a silent, expensive trust failure.
The model decides when you work
Every meaningful model release resets some of your testing. The prompts are tuned against one behavior; the next model behaves differently in the edges. Someone has to re-run the loops, re-check the tool calls, and re-baseline the quality. That someone is you, and it happens on the model's schedule, not yours.
What buying the harness buys you
Against that cost model, the flat pricing of buying a harness looks different. The point of the harness is that the recurring rows — connections, drift, scheduler, approval, measurement, security, model churn — are someone else's job, spread across everyone who uses it.
That is exactly the 90/10 rule from the framework: roughly 90% of a marketing harness is a commodity identical for every builder — live connections, scheduling, approval, measurement, memory. The remaining 10% — how your team words recommendations, which reports it reads, which channels it prioritises — is what is genuinely yours, and that's what you keep control of.
Buying the harness means buying that commodity core as a working system, not a black box: you bring your own AI, your people decide what ships, every change gets a verdict with the misses included, and the differentiated 10% that is yours stays portable. The flat, unmetered-MCP price is the honest cost of not paying for the recurring rows yourself.
For the mechanics of that "bring your own agent" shape, see our guide on your own MCP server vs a pre-built marketing MCP hub — the connector-vs-hub fork is the same build-vs-buy decision one level down.
The flat price, stated plainly
Cogny's paid plans are flat and unmetered for MCP under fair use: Solo ($9/mo), Cogny AI ($18/mo), and Cogny Pro ($90/mo) have the same features — only the credits differ (150 and 1,500 credits a month for AI and Pro respectively). Cloud ($499/mo) is for teams. Start on the Free plan — 15 MCP calls a month and 50 chat credits once, no card required — and every change you ship gets a verdict. There is no free trial on the paid plans: as the site puts it, free is the trial.
Against the upkeep column of the table above, that's the offer: one flat figure that doesn't grow when a platform changes its API, when a new model ships, or when you add a connection. The 10% that is yours — your wording, your priorities, your model — stays under your control on top of it.
The bottom line
The prototype is free. The production system is a subscription you pay with hours — and the invoice keeps coming. Drift, model churn, security and upkeep are not edge cases; they are the product you accidentally bought when you decided to build.
The honest decision rule: build the 10% that is yours, and buy the 90% that is the same for everyone. Buy the live connections, the scheduler, the approval gate and the verdict ledger as a working harness. Spend your own building where it actually differentiates you.
See the flat pricing and the free start on our pricing page, and ship one measured change before you decide.
Start free — no card required.