CG Common Ground | Common Ground Systems
What we did and what it producedActive

The work, decision by decision

What was built

  1. Agents trained on the firm's own record. Not a separate script, and not a general-purpose sales bot pointed at the firm's website. The agents draw on the actual people, companies and relationship history already in the firm's own systems, the same record the CRM migration described elsewhere in this room made possible.
  2. A test that decides whether output ships or goes back to training. The agent has to make a call that reflects what it actually knows about the specific relationship or business in front of it. When it holds the specific over the generic, it is doing real work. When it falls back to a generic script the moment a conversation gets particular, it goes back to training, not to a client.
  3. A human gate before anything public. This is the governance rule and the actual point of the system: nothing an agent drafts, a comment, an outreach message, a reply, reaches a client or a partner's public voice without a person approving it first. Automated systems that comment or engage on a partner's behalf in public settings do not run unreviewed; a governance line exists specifically for that before any such system touches a partner's public voice.

What it changed

The alternative most vendors sell is a scripted layer bolted onto a sales process, which produces generic output fast and cheap and reads as generic the moment a real prospect asks a real question. Running the agents through the firm's own record instead means the ceiling on what they can say is the same ceiling the firm's own knowledge has, which is a real ceiling, not a marketing one, and it is why the human gate matters as much as the training data does: an agent grounded in real data can still say something wrong, out of date, or wrong for a specific relationship, and the gate is what catches that before a client or a partner's audience ever sees it.

What we kept, replaced and installed

We kept the underlying relationship data as the single source the agents draw from, rather than letting them accumulate a separate memory of their own. We replaced the scripted, generic chatbot pattern that dominates the market for this kind of tool. We installed a hard governance rule as a permanent feature of the system rather than a temporary safeguard to be relaxed once the agents prove themselves: nothing reaches a client or a partner's public voice without a person's approval first, on every run, indefinitely.

What it costs, and what we would watch

A human approval step is real friction, on purpose, and it means the system is slower to scale than a fully autonomous one would be. That is the trade being made deliberately: a faster, unreviewed system is a bigger, cheaper way to say the wrong thing to a client at scale. What we would watch: any request, from inside the firm or from a client, to relax the review step for the sake of speed is exactly the pressure this governance rule was built to resist, and it should be treated that way every time it comes up, not granted as a one-time exception.

What it produced

Sales agents now run as infrastructure behind the firm's own sales motion, trained on the same relationship record the team works from, held to a standing test (specific knowledge over a generic script) before output is trusted, and gated by a mandatory human approval step before anything reaches a client or a partner's public voice.

A slice of the project list

A few related projects.