What it takes to hand an AI a budget

Why not connect a general AI assistant to your ad accounts and your email platform, and ask it to run your marketing? The connections exist, and the assistants are good and improving fast. A product that is only a chat window over marketing tools will be replaced by the general assistants, and it should be.

We aren’t building that. The AI model is a part we replace when a better one arrives. The product is everything a team needs before it can trust AI with a real budget and its customers’ inboxes, day after day.

The test

Would you trust it with $10,000 a month and your customer list while your attention is on something else? An assistant shows that it can do a task. Handing it a budget takes proof that it will keep doing it, inside your limits. Those are different products.

Where they differ

Once the question is whether you can trust it with a budget, seven things matter.

• It runs on a schedule. An assistant acts mostly when someone asks. Agents here work on a schedule against a standing brief, file reports, and are managed like a team.
• Its limits are code it can’t change. Spend caps, the limit on each budget change, contact caps, consent, banned phrases and the stop-everything switch are enforced by ordinary code that the AI can’t talk its way past. Agents propose, your rules decide, and the rules run again when you approve.
• No credentials in the agent. A connected assistant holds your account access and makes changes itself. Here, only the executor makes changes, and only after your rules pass.
• A record of everything. In a chat, the conversation is the only record of why a change was made. Here, every action records who proposed it, why, the checks it passed, who approved it and what happened.
• Measured on your numbers. Agents are judged on your own conversion events and what each is worth to you. Experiment statistics are calculated by software, and email sends keep a holdout group by default.
• Memory that lasts. Each agent keeps a notebook of its hypotheses, decisions, lessons and your guidance, and the notebook outlives any model.
• Briefs and reports read like a colleague’s, and decisions are framed in money.

What this means for the model

Because the product is everything around the model, the model can change. Each step of an agent’s work uses whichever model does that step best, from any provider, and your rules, history and library stay the same whichever is in use.