An AI system that finds creators, scores them and runs the first conversations

Before the system, every step was manual. A manager searched Instagram for bloggers, checked their numbers by eye, wrote to them from a personal phone and kept prices and agreements in spreadsheets.

That set a hard ceiling. One person could review a few dozen profiles a day, while a campaign in a new market needed hundreds of suitable creators within a week. Inflated accounts, media pages and creators with the wrong audience slipped into proposals, and good creators got lost. When a manager left, their contacts and negotiation history left with them.

We built four services over one shared database. Discovery starts from creators who already work well and walks their lists of similar accounts, region by region, until it reaches the target. It reads each profile's posts, engagement spread and contacts, and drops accounts from the wrong country or language before paying for a lookup.

Every profile is scored by a model against a four-part rubric: engagement relative to size, audience size, how well the niche converts for the brand, and signs the creator already takes paid work. Code checks sit on top of the model. Inflated engagement, media pages and inactive accounts are penalised, and any sign of an underage creator sets the score to zero with no override.

Outreach runs in three channels under one lead identity, so no one gets the same pitch twice. On Instagram the system first becomes a familiar face with likes, a follow and a story view, then sends one message written for that person. Email picks up creators who stayed silent and YouTube channels, from forty-odd mailboxes with full authentication and a human sending rhythm. LinkedIn works the other side of the market: marketing leads at brands that buy influencer campaigns.

A model classifies every reply. Routine answers go out automatically, anything about money goes to a person, and warm leads reach managers as a card with the translated thread, audience stats and contact. Managers work from one portal that follows each creator from first contact to a published placement, matches creators to a client brief and checks every proposal against stop lists and past refusals.

On top of the pipeline sit AI agents with their own portal accounts. AI managers take a warm creator through qualification, price negotiation on cost per thousand views, launch, proof of publication and the stats screenshot. An AI team lead assembles offers, keeps deals and payouts in the books, assigns briefs and reports every evening. Limits live in code, not in prompts: an agent never sees the client price, and payments are made only by a person.

Streamers on Twitch and Kick get their own track. Instead of trusting viewer counts, the system downloads recent broadcasts with chat and counts the live core of people who actually write, filtering out bot schedules and accounts registered in batches. Paid streams are watched by a machine that transcribes audio, reads brand placements from frames and checks the client's rules.

Every human decision becomes data. Operator edits to drafts are stored as pairs and fed back to the writer. Letters on a manager's own list are written in that manager's voice, learned from their edits. In LinkedIn, autoreplies for each kind of message unlock only after a run of approvals with no edits, and one edit resets the count. A strategist reviews recent conversations several times a day, finds the groups that answer least and proposes a prompt variant, which a person approves before it runs.

The system is in production and runs the agency's sourcing and outreach every day. Discovery and scoring need no one to review profiles by hand, cold email draws replies on substance at a rate manual outreach rarely reaches, and a third of incoming mail is answered without a person. Since May the client's proposals, approvals, placements and payouts have run through the portal, and new regions are added as new AI managers with their own language and persona.