AI vs humans at SocialAmp: the actual receipts

Last week I wrote that "the software is the easy part, the community is the moat." The week before that I wrote that "AI runs the decisions, humans create the content."

Both of those posts make a claim. This one is the receipt.

If SocialAmp is going to sit on the "real humans, no AI in the engagement loop" position, that claim has to survive an audit. Anyone — a member, a competitor, a journalist, LinkedIn's algorithm team — should be able to walk through the platform and verify the line between "AI did this" and "a human did this." Here's the full map.

What AI does at SocialAmp

These are the systems where AI is doing real work. Not theatrical "powered by AI" stickers — actual model calls in production, on every relevant request.

Moderation detection. When a member posts a group message or sends a DM, an automated pattern matcher scans for insults, slurs, spam shapes, and off-topic patterns. The matcher is rules-based plus model-assisted. It flags. A human reviews every flag against the published community rules before any action is taken.

Member support chat. The floating ⚡ help button on every authenticated page is powered by Claude Haiku. It loads your account state (channels, verified status, group, recent activity) and answers common questions — "why didn't I get my SMS code," "when does my group rotate," "how do I add an email channel" — instantly. Top common questions are answered via a deterministic rules layer first (faster, no hallucination risk); the LLM only kicks in for free-form follow-ups.

MCP server for member-side AI. SocialAmp ships an MCP server with 10 tools. If you connect Claude Desktop, Cursor, or any MCP client, your own AI can read your peer queue, surface engagement priorities, draft DMs (which you then send), submit pricing-research, mark links engaged. Critically: this is your AI working for you on your side, not ours acting on your behalf inside LinkedIn.

Channel routing and delivery jitter. When an event fires (new DM, new group message, new shared link), backend code uses deterministic per-user offset math (hash of your user-id + the date, mapped into your morning/afternoon delivery band) to schedule when each peer's notification arrives. No model involved — just code — but it counts as "machine-decided" and we surface it here for completeness.

Link preview generation. A cron walks recently-shared URLs every 15 minutes, fetches them with a Slackbot User-Agent (which LinkedIn whitelists for social-card previews), and caches the Open Graph title/description/image so notifications come with rich previews.

Identity-protection sanitization. When a peer's message is quoted in a notification, code (not a model) strips signatures, mobile auto-appends, and normalizes the display name to "First L." so we never expose more identifying detail than the recipient already has in-app.

Pricing-research analysis. Eventually, when there's enough data, the model behind the admin pricing-research panel will help summarize bucket distributions and free-text comments. (Not yet live — currently the data lives there, no analysis layer yet.)

What humans do at SocialAmp

These are the surfaces where the model is gone and a human is doing the work. Where the operational reality has to match the brand line.

Every engagement on LinkedIn. Every comment, like, save, and reshare that lands on a peer's post comes from a real human on their own LinkedIn account, writing their own words. No browser extension. No automation. No "draft this for me" button that posts as you. SocialAmp does not have your LinkedIn credentials. SocialAmp cannot post for you. SocialAmp cannot engage for you. This isn't a technical capability we've chosen not to build — it's the architectural boundary the whole product is organized around.

Every group placement. Members are hand-placed into groups based on industry, seniority, and (eventually, as scale grows) the audiences each member wants to reach. No clustering algorithm picks your group. No "we matched you because the cosine similarity score was 0.84" — it's a human deciding "you and these other people will benefit from showing up for each other."

Every moderation enforcement decision. AI flags, humans decide. A member warned or removed has been reviewed against the published community rules by an actual person. The bar is "would a reasonable moderator agree?" not "did the model score >0.7?"

Every group rotation. Groups rotate up to once a month. Who goes where, who pairs with whom, who's been in a group long enough to refresh — those are judgment calls. Not generated.

Every conversation in DMs and group chat. Real members typing to other real members. SocialAmp doesn't auto-draft, auto-suggest, or auto-reply in chat. (The MCP server lets your AI suggest things to you — but you type and send what actually goes in.)

Every support email and escalation. The chat widget answers common questions. Anything past common — billing disputes, account issues, content moderation appeals, weird edge cases — goes to a human inbox and gets a human reply.

Every word on the marketing site. Including this post. AI helps me draft and refine — same as a copy editor would — but the perspective, the receipts, the willingness to make specific operational claims, all of that comes from me being on the hook for them.

The boundary

The pattern is consistent: AI handles infrastructure, code, and the things customers never directly see. Humans handle every surface where the audience is supposed to be looking at the product.

Or more bluntly: an AI couldn't have made the choice to draw the line this way. The line is the product.

What this means if you're evaluating SocialAmp

If you're considering SocialAmp and you want to verify the "no AI in customer-facing engagement" claim before signing up, here's what to do:

1. Read this list and check it against what you experience. If you spot a place where AI is doing something I claimed a human does, tell me. Email russ@bintelligence.com.
2. Look at the platform itself. SocialAmp does not have a Chrome extension. Search the Chrome Web Store. It's not there because it doesn't exist. Compare that to other tools in this space.
3. Check the LinkedIn auth model. SocialAmp doesn't use LinkedIn OAuth. We never ask for your LinkedIn credentials. We can't post or engage on your behalf — the architecture doesn't permit it.
4. Notice what we don't promise. No "AI will write your comments." No "we'll engage on auto-pilot." No "set it and forget it." If we promised those, we wouldn't be SocialAmp — we'd be the thing LinkedIn's 2026 algorithm is actively burying.

Why we publish this list

A lot of the brands now claiming "human-made" or "anti-AI" haven't done the work to make those claims operationally true. The credibility damage when audiences spot the gap is worse than never making the claim — that was the warning from the State of Brand piece I wrote about earlier this month.

So: receipts. If we ever quietly slip something into the "AI does" column that belongs in the "humans do" column, this post will be updated and the change logged. Hold us to it.

---

See where SocialAmp lands in the broader comparison or start a 30-day free trial.