Helping a business put AI to work

A produce business of about 15 people had Claude accounts and a few people using them in different ways, with no shared setup and no easy way to bring the company's own information into the work. Over one engagement I took the company from scattered experiments to a working setup: six role-based Claude workspaces, written guidance, staff training, the CEO's daily report automated, and a controlled path to the company's ERP data: a dedicated server now captures data from the ERP for an AI vendor's integration, without handing anyone a live write path.

People were trying AI in different ways, without a shared setup or clear guidance. The CEO had a working daily analysis in Claude. One salesperson was feeding it institutional knowledge in chat, where nothing was kept. Others had not started, and one of them was sitting on years of customer notes in a spreadsheet. The ERP had no ready connector, so the company's own data could not reach any of this. The CEO had already tried two automation tools for the daily report and found them too complicated.

• Six workspaces. I built six Claude workspaces around the real org: a desk each for leadership, sales, operations, finance and design, plus a shared prompt library. Each desk has its own instructions and knowledge files, backed by a library of 32 prompts written for those roles and per-role cheat sheets. Every desk carries the same written rule: Claude drafts, a human sends.
• A getting-started site. A short, client-branded guide covering the basics, how to use Claude, how to set up a Project, how to work securely, a prompt library and builder, an English-Spanish helper, and a page on which AI to use for what. I pulled the pages about their own produce because we did not know their business well enough to stand behind them.
• Training. A live group session with the staff, a working session on the Microsoft 365 connector, and a session with the CEO on Cowork.
• The daily report. The CEO's report moved into Claude Cowork, the tool already on their side, instead of the heavier pipeline the company had been considering. It runs automatically: it reads the day's spreadsheet from email, runs the CEO's own tested analysis script, reconciles the totals, and emails a full report to the people who need it on a schedule, with a guard against sending twice. If the totals do not balance, the report is held and never sent. The report follows the company's season, so it pauses when the season does.
• The data question. The ERP data model was obtained and analysed. I designed a self-built bridge, then set it aside as reference when a managed read-only data service and an AI vendor's integration came into view. I brought the vendors in and set the terms: read-only first, writes only with human review. The company now runs its own dedicated server instance that captures data from the ERP for that integration, and it has been running for a while. On top of it, the vendor's pilot is deliberately narrow: creating produce lots from grower-manifest emails. I wrote the due-diligence brief and a client FAQ along the way.