Applied AI · agents, governance and MCP

Sales agents with governance, not with luck

The hard part of AI in a product is not the model call. It is everything around it: permission, cost, context, limits and what happens when the model gets it wrong. That is what I designed and built.

My role

Agent system design and implementation

When

2026, in production

Outcome

AI the customer trusts, because it has limits, leaves a trail and can be switched off per conversation. No black box.

What I did

• A per tenant SDR agent, with reactivating the cold base as the core, because that is where the money already is, and prospecting as the add on.
• An MCP server with 44 tools that exposes the CRM to the customer's own AI, with an explicit boundary on what stays out, such as billing, team and sensitive data.
• A governance layer every call goes through: a permission and cost gate per tenant, a training screen, a governance screen and a record of what the AI did.
• An intent engine that scores each lead from 0 to 100 and returns the next action, instead of returning a lonely number nobody knows how to use.
• Anti ban humanization on WhatsApp: typing delay, a daily volume ramp and a cadence that postpones instead of burning the customer's number.