$33 in incremental AI Spend
That's what it costs to run a full AI B2B marketing engine for a client right now. Content for LinkedIn. Email. Blog. SEO. Targeting and Pipeline prep for the sales team.
First quarter live the result was 25% more organic traffic year over year with stable lead performance.
The staff count went from 12 to 2.5 and other subscriptions for image and video generations stayed stable.
What's Actually Running
This isn't "we use AI to write content." That's noisy not measurable and most of it is garbage.
This is a coordinated AI system. Haiku handles the high-frequency, lower-complexity work. Sonnet takes the pieces that need judgment. A custom harness orchestrates the handoffs and build the context for each call.
The agents don't operate in a vacuum. They pull live business data through custom MCP connectors. They reach into Apollo for targeting intelligence. HubSpot for contact and deal context. A custom Outlook connector queues personalized 1:1 messages for the sales team to review and send so outreach looks like it came from a person, because a real person does review and send it. They just didn't write it from scratch at 7am.
The result is a system that knows who to reach, what to say, where to say it, and when.
What the Numbers Mean
$33.10 is a token cost. Infrastructure, oversight, and the build are separate. But the operating cost of content and outreach at this volume is genuinely $33 a month in model spend. The team had already subscribed to two tools for image and video generation those stayed the same, they just got integrated into a process.
The token efficiency comes from a specific design decisions: First if we could do it deterministic we did, code is cheap to run. Then we sized each AI task to the lowest cost model to be effective. Haiku doesn't do what Sonnet should. Sonnet doesn't do what Haiku can. And nothe in this needed Opus. That discipline compounds across tens of thousands of calls a month.
What it took to get here
The models were not the hard part, they never are, it is all about the data. We built custom MCP connectors to give the agents access to live business data, the kind of context that turns generic output into something that actually reflects the client's market position, ICP, and pipeline state. A custom Outlook connector bridges the system into the sales team's existing workflow. Multiple users. Real data. No copy-paste handoffs.
The harness keeps as much of the process deterministic as possible. Routing, formatting, sequencing, conditional logic — defined, not inferred. The agents handle judgment. The harness handles everything else. That's what keeps output consistent and token spend where it should be.
Most of this work happens before the first agent runs. That's the point.
One Quarter In
25% traffic lift os real, leads are converting at the same page as before all on a $33 in model spend.
The system is tuned with content flowing and tokens processing. The sales team is sending messages they didn't have to write. Marketing is pushing text, images, and video they didn’t have to draft.