BuildSmarter Marketing AI is built around large language models. This page explains what that means in practice — what the models do, what they don't, and what stays your responsibility as the operator.
1. What the agents do
Each customer-facing agent has one job:
- Brand Brain reads your onboarding answers and voice samples and produces a structured Brand Profile that every other agent writes from.
- Content Studio generates LinkedIn posts, carousels, blog posts, and emails on your brand voice.
- SEO Engine drafts geo × vertical landing pages with meta tags, body sections, proof points, and FAQ.
- CMO Reporter writes a weekly digest of your marketing activity for your inbox.
All four agents currently call Anthropic Claude models. The system prompts, output schemas, and eval suites that constrain them live in our codebase.
2. What the agents don't do
They don't publish. Every piece of output lands in your review queue. Nothing reaches LinkedIn, your blog, your inbox, or any other public surface unless you approve it.
They don't verify facts. Voice and structure are constrained by our prompts and evals. Specific claims — project counts, dollar figures, dates, awards — come from the inputs you supply. The agent doesn't cross-check them against the world.
They don't replace counsel or expertise. The output is marketing copy, not legal advice, financial guidance, or engineering certification. Anything published under your name is your statement.
3. The honest failure modes
Large language models can:
- Fabricate specifics that sound plausible (a project that doesn't exist, a metric you never claimed) — known as “hallucination.”
- Use cliché phrasing that the eval suite tries to catch (“best-in-class,” “leverage”) but won't catch every time.
- Score their own output too generously — every customer-facing agent includes a self-rating, and we treat low scores as reliable signal and high scores as “worth a closer read.”
- Refuse to handle inputs that look like prompt injections, in ways that may surprise you. If you see refusals on legitimate content, contact us.
We measure these failure modes with a paid eval suite on every agent before shipping prompt or model changes. Current baselines (Brand Brain ≥98%, Content Studio ≥96%, SEO Engine ≥92%, CMO Reporter ≥76%) reflect the percentage of test cases that pass our hard gates and quality judges.
4. What we send to the model provider
When you trigger an agent, we send Anthropic the relevant Brand Profile context and the agent input (your topic, format, settings). Anthropic does not retain prompts or completions for training on our plan. We do not log the full prompt or completion outside of our database (and Sentry stack traces on failure).
5. Your responsibilities
Review every piece of output before it leaves the platform. Verify any specific claim (numbers, project names, awards, dates) against your own records. Apply your own judgment on tone and message — the agent matches your voice, but you know your relationships.
6. Disclosure when content is published
We do not require you to disclose AI involvement when you publish content drafted with the Service — common practice in marketing today is that AI assistance is implicit. Some jurisdictions or platforms have their own disclosure rules; complying with those is your responsibility.
Contact
Questions about how we use AI: hello@buildsmarter.ai.