The situation
GTM had real AI usage. People were getting value in chat windows and then pasting the output somewhere by hand. Every win stayed local, and the effort of moving information between systems ate most of the gain.
The actual problem
The request was
“Roll out more AI tools to the GTM team.”
Discovery showed
More tools would have added more copy and paste. The friction was between the systems, not inside them. Context lived in Notion and Salesforce, conversation lived in Slack, and the model had access to neither at the moment it was needed.
The goal I set was radical frictionlessness: the work reaches the right person or agent with the context already attached, and the result writes back where the next step happens. That is an architecture problem, not a licensing one.
What I learned
- Mapped where GTM context actually lives and which systems can read and write it.
- Traced handoffs to find the points where information was retyped rather than passed.
- Tested MCP connections and skills against real tasks before committing to a pattern.
- Set autonomy levels per task class so write-back never outran trust.
How I framed it
Exhibit
The Operating Architecture
01Surfaces
Slack and Cowork, where the work is already being discussed
02Agents & skills
Reusable skills per task class, with defined autonomy
03Context (MCP)
Notion, CRM, and data sources connected as first-class context
04Write-back
Results land in the system of record, not in a thread
05Notifications
The next human is told, with the reasoning attached
Write-back is what turns a helpful answer into a completed step.
Exhibit
Autonomy by Task Class
- 01
Draft and summarize
- Context needed
- Thread + docs
- Write access
- None
- Human role
- Reviews output
- 02
Assemble and enrich
- Context needed
- CRM + Notion
- Write access
- Scoped write
- Human role
- Spot checks
- 03
Route and notify
- Context needed
- Ownership map
- Write access
- Write + notify
- Human role
- Handles exceptions
- 04
Commit a change
- Context needed
- Full record
- Write access
- Audited write
- Human role
- Approves first
| Context needed | Write access | Human role | |
|---|---|---|---|
| 01Draft and summarize | Thread + docs | None | Reviews output |
| 02Assemble and enrich | CRM + Notion | Scoped write | Spot checks |
| 03Route and notify | Ownership map | Write + notify | Handles exceptions |
| 04Commit a change | Full record | Audited write | Approves first |
What we built
- 01Skills I wrote for recurring GTM task classes, so the same work runs the same way instead of living in one-off prompts.
- 02MCP connections into Notion, the CRM, and the data sources, so agents read the real systems of record.
- 03Persistent project context and agent memory, so a request does not start from a blank page every time.
- 04Write-back and Slack notifications orchestrated through n8n and MCP, so a result lands in the system of record and the next person is told.
- 05Concrete Marketing workflows: campaign briefs assembled from live context, content drafted against the current positioning, and results written back.
- 06Concrete Recruiting workflows: role intake, candidate summaries against the scorecard, and status updates pushed into Slack for the hiring team.
- 07Autonomy levels tied to task class, with evaluation before any level increase.
My role
- Architecture
- Designed the connected system and chose the integration patterns.
- Build
- Wrote the skills, connected the MCPs, set up project context and memory, and wired n8n and MCP write-back with Slack notifications myself.
- Workflows
- Built the Marketing and Recruiting workflows end to end with the people who run them.
- Adoption
- Worked inside GTM workflows rather than shipping a tool and leaving.
- Guardrails
- Defined autonomy levels and the evidence required to raise them.
The outcome
- AI moved from isolated chat sessions into the GTM workflow itself.
- Handoffs that used to require a human copying context now carry it automatically.
- New use cases plug into existing context and write-back instead of starting from zero.
- Autonomy became a decision with criteria rather than a preference.
What I'd do differently
The interesting engineering was not the agents. It was the plumbing that let them see the truth and change it safely. Nobody demos plumbing, and it is the entire difference between a pilot and an operating system.