Idea Gallery
17 ideas to spark yours, or bring your own. Each card has the problem and how an agent could solve it.
note
Inspired by the industries of the people in the room. These are illustrative, synthetic ideas: not real projects, data or plans from the companies named. Pick one, remix one, or bring your own.
They are for inspiration, not assignments: no team owns a card, and you do not have to pick one to start. The guided path does not depend on this page.
Retail & e-commerce Store-ops copilot for till incidents
Inspired by Extenda Retail
The problem
- Till and POS errors reach support as vague tickets from busy stores
- Working out whether it is one store or a whole chain takes an expert
How an agent could solve it
- Keep a memory per store and line up its synthetic error log with recent config changes
- Write a short case file: likely cause, affected stores, the runbook page it used
- Propose a config rollback that a human approves before it is pushed
Retail & e-commerce Delivery promise watcher
Inspired by Ingrid
The problem
- Carrier delivery estimates drift, but checkout keeps promising the old window
- Customers find out before the merchant does
How an agent could solve it
- Check a synthetic carrier feed on a schedule and compare it with the promised windows
- Stay quiet when nothing changed; alert with evidence when a promise is at risk
- Draft new checkout delivery options for a human to approve
Retail & e-commerce Market-compliance checker for product pages
Inspired by Haypp Group
The problem
- Regulated products need different age gates and warnings in every market
- Checking every product page in every market by hand does not scale
How an agent could solve it
- Render product pages per market in a real browser
- Check them against a rule book indexed in AI Search and cite the rule
- Queue the fix for a human; never edit a live page on its own
Security & compliance Drop-day bot and fraud triage
Inspired by H&M Group
The problem
- Limited product drops attract bots and credential-stuffing in minutes
- Rules written under pressure are too broad, or never removed
How an agent could solve it
- Investigate a synthetic spike of logins and add-to-cart events
- Draft a time-boxed rate-limit or WAF rule scoped to the attacked path
- Apply it only after a human approves, and let it expire on its own
Security & compliance Phishing and lookalike-domain hunter
Inspired by Swedbank
The problem
- Lookalike domains imitate a bank to trick customers into logging in
- Reviewing each candidate by eye is slow and inconsistent
How an agent could solve it
- Screenshot candidate domains on a schedule with Browser Run
- Score how closely each one imitates the real site, with the evidence
- Draft takedown requests that a security analyst approves
Finance & insurance Claims intake assistant
Inspired by If P&C Insurance
The problem
- A first notice of loss arrives as free text and photos, missing half the details
- Handlers spend their time chasing information instead of deciding
How an agent could solve it
- Talk to the customer, ask for what is missing and read the photos with a vision model
- Store a structured claim and flag anything that looks inconsistent
- Route it to a handler; never promise a payout or decide the claim
Finance & insurance Donation and payout reconciliation
Inspired by Octany
The problem
- Non-profits reconcile one-off gifts, monthly donors and failed payments by hand
- Failed recurring payments quietly turn into lost donors
How an agent could solve it
- Watch synthetic payment events and match them to donors
- Spot failed or missing recurring payments and draft a friendly follow-up
- Send nothing until a person approves the batch
Finance & insurance Purchasing agent with a spend limit
Inspired by SolvaPay
The problem
- Agents that can pay for things need budgets, receipts and a way to say no
- Nobody wants to discover what an agent bought at the end of the month
How an agent could solve it
- Quote a purchase and check it against a budget kept in D1
- Run small purchases freely; stop for human approval above a threshold
- File every receipt in R2 with who approved it
Industry & mobility Connected-product field issue triage
Inspired by Husqvarna
The problem
- A fleet of connected products reports thousands of error codes a day
- Telling a firmware regression from bad weather takes days
How an agent could solve it
- Group synthetic error reports by model, firmware and region
- Cite the service manual in AI Search for each cluster it explains
- Propose pausing a firmware rollout; a human makes the call
Industry & mobility Flaky test investigator
Inspired by Volvo Cars
The problem
- Large software programmes drown in flaky and failing test runs
- Engineers lose hours working out which commit broke what
How an agent could solve it
- Line up failing tests with recent commits and earlier runs it has remembered
- Write a case file per failure with the likely commit and its confidence
- Ask a human before quarantining a test or reverting anything
Data & AI platforms Data pipeline freshness watcher
Inspired by Ericsson & Perspetivo
The problem
- Dashboards quietly show stale numbers when an upstream job is late
- By the time someone notices, decisions were made on old data
How an agent could solve it
- Check table freshness and row counts on a schedule
- Explain the anomaly in plain language with the job that caused it
- Start a backfill Workflow that pauses for a human to approve
Industry & mobility Building documents Q&A with citations
Inspired by Brickanta
The problem
- Drawings, specs and inspection reports for one building live in many places
- Answering "what was installed where, and when?" means digging through PDFs
How an agent could solve it
- Index synthetic building documents in AI Search
- Answer with the document and page it came from, and say what is missing
- Draft a change request that a project lead approves
Data & AI platforms Model deployment guardian
Inspired by Deeploy
The problem
- Models in production drift, and promoting a new version is a risky, manual call
- Regulators and customers ask why a model decided what it did
How an agent could solve it
- Watch synthetic drift and accuracy metrics for each deployed model
- Write a plain-language explanation of what changed and why it matters
- Promote or roll back only after a named human approves in a Workflow
Security & compliance Validation evidence pack builder
Inspired by TPP Validation
The problem
- Validating regulated software means collecting screenshots and test evidence for every requirement
- Building the evidence pack by hand is slow and easy to get wrong
How an agent could solve it
- Run scripted checks against a synthetic app and screenshot each step with Browser Run
- Trace every piece of evidence to its requirement and store it in R2
- Hand the pack to a human for sign-off; it never signs anything itself
Media & creative Show notes and music-rights checker
Inspired by Soundright Media
The problem
- Every episode needs chapters, show notes and a check that each music cue is cleared
- A missed licence is discovered after publishing
How an agent could solve it
- Draft chapters and show notes from a synthetic transcript
- Check the cue sheet against a licence register in D1
- Publish only after a producer approves
Media & creative Brand-safe image studio
Inspired by AiBrush
The problem
- Generated images are quick to make and easy to get off-brand
- Reviewing every image against brand rules is a bottleneck
How an agent could solve it
- Generate images with a Workers AI image model from a brief
- Check each one against written brand rules and explain any problem
- Store drafts in R2; a designer approves before anything is shared
Any industry Stale deploy and cache operator
Inspired by Niteco & Keystone
The problem
- Agencies and platform teams run many sites, and old preview deploys and Worker versions pile up
- Cache hit ratio drops and nobody knows which header is to blame
How an agent could solve it
- List stale deployments and sample responses for cache-busting headers
- Propose exactly what to delete and which cache rule to add
- Act only on the list a human approved
Your idea Bring your own idea
The problem
- A slow, manual or error-prone job in your own work
- Something you would be glad to hand to an agent
How an agent could solve it
- Start from your problem, not from these cards
- Pick one action worth automating, and one moment where a person approves
- Use synthetic data: your demo URL may be publicly reachable
Looking for the shapes agents take (Watcher, Operator, …)? They are in the Agentic Patterns section.