Workflow examples

5 practical AI workflows for normal work

These are not abstract prompt ideas. Each workflow below is tied to an AIProductivity.guru evidence pack or live test log, so you can see what we are testing, what still blocks reliable use, and when setup friction is not the same as output quality.

Updated 2026-06-09

How to use this page

Pick a recurring task, start with sanitized inputs, keep AI in draft mode, and verify the result manually before you send, publish, schedule, or connect anything. Where our public tests only show signup gates, prompt echoes, or missing exports, we label that as setup friction and leave quality scores blank.

Example 1

Turn a messy meeting into a follow-up plan

Job: You have a transcript, notes, or a rough recording summary and need decisions, owners, deadlines, and unresolved questions without inventing promises.

Safe input boundary: Use a synthetic or consented transcript. Remove customer names, private compensation details, legal facts, and anything that should not leave the company.

  1. Ask AI to separate decisions, action items, open questions, and risks into labeled sections.
  2. Require every action item to include an owner, verb, due date or “date missing”, and quote/evidence from the notes.
  3. Review against the transcript before sending; do not let the tool email or schedule anything automatically.

Evidence link: The meeting-notes evidence log tracks setup friction, import limits, privacy/export notes, and action-item accuracy requirements before recommendations.

Open the related evidence log
Example 2

Build a Monday admin triage from messy notes

Job: You have inbox snippets, voice notes, reminders, and half-finished ideas and need a realistic week plan plus reply drafts.

Safe input boundary: Paste a cleaned bundle rather than connecting Gmail, Outlook, Slack, Notion, CRM, or calendar accounts unless the account and automation scope are project-safe.

  1. Ask for exactly six sections: urgent replies, calendar/reminder plan, decisions needed, delegated tasks, content reuse ideas, and follow-up questions.
  2. Require the assistant to mark missing facts instead of pretending it attached files, updated checklists, sent messages, or changed calendar events.
  3. Copy useful drafts manually after checking dates, names, money amounts, and action claims.

Evidence link: The admin-routine evidence log includes scored prompt-only baselines for tools that produced full answers and setup-friction records for tools that gated or echoed prompts.

Open the related evidence log
Example 3

Convert a rough strategy memo into client-ready slides

Job: You have a messy client memo and need a clear deck outline, talking points, and exportable slides without losing the business logic.

Safe input boundary: Use a fictional or sanitized memo first. Avoid uploading confidential client data to a tool whose retention/export rules you have not checked.

  1. Ask for an 8-slide structure with one core message per slide, speaker notes, and a list of claims that need human proof.
  2. Check whether the tool can produce a private draft and export to PPTX/PDF before investing time in design polish.
  3. Treat sign-in gates, stalled generation, and missing exports as setup friction, not output quality.

Evidence link: The presentation-tools evidence log currently records no-login/signup/security/export friction, including a SlideGen no-login composer run where no generated deck appeared.

Open the related evidence log
Example 4

Draft a customer-support FAQ bot safely

Job: You want a chatbot or internal assistant to answer repetitive support questions from a known FAQ without making up policy details.

Safe input boundary: Start with a small public FAQ or synthetic knowledge base. Do not connect live helpdesk, customer data, or billing systems until fallback and deletion controls are verified.

  1. Upload or paste only the allowed knowledge base, then ask test questions with exact expected answers and “should escalate” cases.
  2. Check grounding, citation/source behavior, fallback wording, handoff controls, transcript export, and deletion/training settings.
  3. Publish or recommend only after private-bot transcripts and answer-key scoring exist.

Evidence link: The chatbot evidence log maps public-doc availability, signup blockers, upload/export controls, and no-score setup-friction evidence before quality rankings.

Open the related evidence log
Example 5

Create a small-business website draft in one afternoon

Job: You need a first website draft for a local service business: homepage copy, services, local SEO basics, contact sections, and editability checks.

Safe input boundary: Use a fictional business brief first, and do not connect a real domain, payment plan, or customer analytics until export/lock-in and ownership rules are clear.

  1. Give the same bounded business brief to each builder so you can compare like with like.
  2. Check mobile layout, invented testimonials/certifications, local SEO basics, editing time, export/migration limits, and unpublish/delete controls.
  3. Keep draft-site screenshots and raw notes before assigning quality scores.

Evidence link: The website-builder evidence log uses one fictional cleaning-business brief and separates generated-site quality from signup, payment, domain, and export friction.

Open the related evidence log
Reader-safe checklist

Before you trust an AI workflow

CheckWhy it matters
Evidence sourcePrefer tools and workflows that can be tested with repeatable fixtures, screenshots, exports, and raw outputs rather than marketing claims.
Private data boundaryPaste sanitized examples first; avoid real inboxes, calendars, customer lists, or file drives until account, retention, deletion, and automation settings are understood.
Human review stepAI drafts can sound confident while inventing attachments, decisions, customer policies, certifications, or completed actions.
Export and lock-inA tool is less useful if you cannot copy, download, edit, delete, or migrate the work after generation.
Setup friction versus qualityA login gate, CAPTCHA, OAuth-only signup, or stalled composer is operational friction; it does not prove the AI output is bad, but it matters for readers trying to reproduce the workflow.