Collaborative Operations Automation

Many useful AI opportunities are not grand transformations. They are small, repeated decisions that consume expert attention every day: understanding a sales conversation, drafting a customer response, finding the next account to pursue, or resolving an exception without losing context.

I designed a portfolio of operational automations around a simple principle: automate the work that benefits from speed and structure, then preserve human control wherever an action has customer, commercial, or reputational consequences. Every automation was built with AI-assisted coding — I designed and reviewed each one; an AI agent wrote the implementation.

A Repeatable Operating Pattern

Each workflow follows the same basic architecture:

1. Gather the relevant context from the systems where work already happens.
2. Structure or recommend using AI and deterministic rules appropriate to the task.
3. Route ambiguity or consequential decisions to a named person or review queue.
4. Write narrowly and visibly to the system of record, with state, audit history, and safe re-run behavior.

This is a more durable pattern than simply connecting a model to an API and hoping for the best.

Commercial Intelligence From Sales Conversations

One workflow reads business-development meeting transcripts and turns them into a consistent six-part deal brief: summary, answered and open questions, engagement read, missing technical details, next-call prompts, and structured account facts. The output is written to the team’s existing deal workspace and linked to the source transcript, so the workflow can be safely re-run without creating duplicates. A failure on one transcript does not stop the rest of the batch.

AI Drafts, Human Approval, Controlled Action

Another workflow processes new customer feedback, generates a proposed response, sends it to a human review queue, and posts only after approval. It maintains workflow state and an audit record, and defaults to a non-posting mode until the operating team explicitly enables live action. The important design choice is not the generated text; it is the approval boundary around the customer-facing action.

Decision Support Where Rules Are Better Than Language Models

Not every workflow needs an LLM. For rejected orders, a rules-based automation finds viable substitute products using inventory, product attributes, category, and pricing constraints. It ranks candidates transparently and writes only the intended replacement fields back to the operations worksheet. Where the system cannot classify a case confidently, it leaves the record alone.

Key Design Decisions

Idempotency isn’t one-size-fits-all. Each workflow tracks its own state the way that fits how it already operates: the customer-response workflow keeps a ledger of posted/skipped ticket IDs so a re-run never double-posts; the deal-brief workflow checks what’s already in the destination database before generating a new one; the order-rejection workflow treats an empty spreadsheet column as the only state it needs. No shared framework imposed a generic mechanism across all three — each one gets the safest approach for how it actually operates.

Work around the platform, don’t fight it. Zendesk rejects an API request to close a ticket unless a specific custom field is set — a rule that has nothing to do with this workflow. Rather than force that field or route around it, the automation tags the ticket and leaves it open for a person to close. A production integration inherits the target system’s constraints; the design has to accommodate them, not assume them away.

A safety mode has to be tested for what it doesn’t do, not just what it does. An earlier version of the customer-response workflow wrote to its idempotency ledger even during a dry run, which silently blocked the real post once dry-run mode was turned off. The fix was narrow — only record state on an actual post — but it’s a real example of a safeguard almost causing the failure it existed to prevent.

What This Demonstrates

The individual automations are different, but together they show a practical way to deploy AI in operations: start with a valuable workflow, give people clearer and faster inputs, add controls before autonomy, and make the system safe to operate repeatedly. It’s the same thesis I lay out in Collaborative AI — applied here to production customer and commercial operations.

Tech Stack

Node.js, Claude, Notion, Zendesk, Google Sheets, BigQuery, marketplace APIs, workflow state management, and structured audit logging.