← Projects

Jiyana AI Agent Operations

Jiyana · 2025 – 2026 · Co-Owner & Operations Lead

The three agents, each with one job, all reached the same way: a plain-language message in Slack.

The problem

Jiyana runs with a small team, and small teams generate more work than hands to do it: inventory needs logging every time stock moves, content needs planning every month, decisions and follow-ups pile up across sales, finance, and operations. The traditional answer is to hire, bringing in headcount for the administrative and recurring work so the core team can move faster on what actually needs their judgment.

We took a different path. Instead of adding headcount and the fixed cost that comes with it, we built a layer of automation and AI agents to absorb the recurring, administrative parts of running the business (inventory logging, content planning, and cross-functional tracking) while keeping the judgment calls with us.

What I built

Every agent follows the same pattern. One function each, all operated through Slack, all in plain language with no command syntax to learn. Whoever is using it describes what happened or what they need, and the right agent handles the structure, the recordkeeping, and the recurring effort behind it. The judgment calls stay with us.

How an agent works end to end: a plain-language message in, a structured record out.

Inventory. Stock movements get logged by describing them in Slack, and the agent records them into the system covered in Inventory Management System. It was the first agent we built and the template for the ones that followed.

The inventory agent. Describe what moved, and it records itself.

Content Planning. Monthly social content, captions and on-image copy, gets drafted by the agent from the visual itself, then reviewed and approved directly in Slack. Nothing needs a separate content tool or a spreadsheet someone has to remember to open.

The content planning agent. Drafts from the visual itself, then waits for approval.

Project Memory. Decisions, open threads, and follow-ups across sales, accounting, finance, strategy, product, and operations get logged with a simple tag: [Decision], [Parked], [Recall], [Close], instead of getting lost in scattered conversation history. The agent tracks and flags; it never makes the call itself.

The project assistant agent. Decisions and follow-ups tagged so nothing gets lost in the thread.

The decisions behind it

Slack, not a new tool. The agents came to the team rather than the other way round. A system that asks people to go somewhere else has to earn that trip every single time, and most days it will not.

Plain language, not commands. A command format is one more thing to learn, and one more thing to forget. Describing what happened asks nothing new of anyone, which is what makes the agents usable by the whole team from day one.

The agent asks when it is not sure. We decided early that an agent guessing is worse than an agent pausing. Anything ambiguous stops and waits for a person to confirm. It costs a few seconds and saves the cleanup.

Judgment stays with people. Every agent records, drafts, tracks, and flags. None of them decide. Recording and drafting are safe to hand over. Deciding is not.

The impact

Absorbed recurring administrative work across inventory, content, and cross-functional tracking without adding headcount or fixed payroll cost
Cut monthly content planning time by roughly 70 to 80%, from 1 to 2 hours of full effort (plan, draft, review, check, finalize) to about 20 minutes of human review, revision requests, and approval
Built to ask rather than assume. When an entry is ambiguous, the agent stops and offers a simple choice instead of guessing, so incorrect records never enter the system in the first place
One consistent interface (Slack, plain language) across every function, no tool-specific training needed for the team