Workflows

Examples of practical agent workflows

OEJ builds custom agents around one repeating task at a time. The agent prepares the work, keeps access limited, and waits for a person to confirm important actions.

Discuss your first workflowSee how a pilot works
One clear workflowEach example starts with one repeated task, not a broad AI project.
Scoped accessThe agent gets only the permissions needed for that workflow.
Human approvalThe agent prepares the work. A person confirms the important step.
Built workflows

Three workflows we have built

These are practical examples of how an agent can prepare work without taking control away from the person responsible for the result.

Scheduled decision support

Prediction entries prepared before the deadline

A football prediction competition needed picks entered on time for every match round. The repetitive part was collecting the fixtures, modelling the likely results, logging in and preparing entries.

  1. The agent models the fixtures and prepares suggested results with short reasoning.
  2. It opens the prediction website and prepares the round entries.
  3. It sends a summary of the suggested picks.
  4. A person reviews the picks, changes anything they want, and only then submits.
Everyday office work

A Google Workspace assistant in Estonian

The assistant helps with everyday work in Gmail, Calendar, Drive, Docs, Sheets, Slides and Forms. It summarises, writes drafts, prepares calendar entries and prepares documents. It does not silently send important messages.

With explicit permission, it can use selected or recent sent emails as examples, so drafts are closer to the person’s normal Estonian writing style. The person still reviews, edits and sends.

  • Email — summaries and reply drafts.
  • Calendar — event drafts and reminders.
  • Documents — Docs, Sheets and Slides drafts in Drive.
  • Forms — questionnaires prepared on request.
Small internal tool

A local Excel-to-SQL tool

A repeating manual job became a small reusable tool: take an Excel file and generate SQL INSERT statements. The tool can run on your own machine, and the generated SQL is reviewed before it touches a real database.

Sometimes the best agent output is not a chat reply. It is a small utility that removes a repetitive task from the daily routine.

Starting points

Other good first workflows

A first agent workflow should be narrow, useful and easy to approve. These are typical starting points for local businesses.

Inbound requests

Summarise a customer email or form submission, pull the background and draft the response for approval.

Daily briefing

One morning summary: unread mail, today’s calendar and anything important that changed overnight.

Quiet monitoring

Watch a webpage, price list or public register and ping only when something material changes.

Scheduled content drafts

Prepare posts or product copy into a review queue. Nothing publishes without sign-off.

Document summaries

Turn long documents or meeting notes into a short summary and a clear next-action list.

Small internal tools

Create small reusable utilities for repetitive manual jobs, like the Excel-to-SQL example.

Pilot

How a pilot works

The safest way to start is one useful workflow with clear access and approval points.

Pick one workflow

Usually a task already happening in email, calendar, documents, Telegram or spreadsheets.

Agree access and confirmations

We define what the agent may read, where it may create drafts and which actions need approval.

Check whether it is worth expanding

If it saves time and feels safe, we build on it. If not, we stop there.

Next step

Have one task that keeps repeating?

Tell us about one workflow that takes too much time or keeps slipping. We will say honestly whether it is a good first candidate for an agent.

Send a workflow description