/notes / 11 Nov 2026

Stop Prompting, Start Automating: AI Workflows That Save You Hours Every Week

Part 7 of AI From Zero to Agents. Why most of your work needs an AI workflow rather than an agent, the five patterns behind almost every real automation, five examples you can copy, and a quick test for what to automate first.

· 5 min read

Series · AI From Zero to AgentsPart 7 of 9
  1. 1 AI for Newbies: The 10-Minute Guide That Finally Makes It Click
  2. 2 ChatGPT Isn't Magic, It's Autocomplete on Steroids
  3. 3 Not All AI Is the Same Brain: Meet the Model Family
  4. 4 ChatGPT vs Claude vs Gemini vs Grok vs Llama vs Perplexity: Which AI Should You Actually Use?
  5. 5 Free, Open and on Your Laptop: Open-Source Models and the Secret Currency Called Tokens
  6. 6 From Chatbot to Coworker: What Agentic AI Is and How to Build Your First Agent
  7. 7 Stop Prompting, Start Automating: AI Workflows That Save You Hours Every Week
  8. 8 The USB-C Moment for AI: MCP vs A2A vs ACP, and How Agents Talk to Each Other
  9. 9 Let's Build It: A Hands-On Agent-to-Agent Walkthrough
Documents and envelopes travelling along glowing rails through a series of light stations, with an hourglass in the foreground

A confession from someone who designs AI systems for a living: most of the time, you don't need an agent.

In Part 6 we built agents that decide their own next step. That freedom is great when every case is different, but you pay for it in speed, cost and predictability. The repetitive stuff that eats most of our week is better served by a plain, well-built AI workflow. It does the same thing every time, costs a fraction of an agent, and when something goes wrong you can actually see where.

The train and the taxi

A workflow is like a train on fixed rails: the path is designed in advance, predictable, cheap and easy to audit; an agent is like a taxi: it chooses its own route to the destination, which is flexible but less predictable and more expensive
A workflow follows the track you laid. An agent picks its own route.

The way I explain it to customers: a workflow is a train. You lay the track, and the AI does the clever bits at each station (reading a document, sorting an email, writing a draft). An agent is a taxi. You give it the destination and it works out the route as it goes.

My rule of thumb is simple. If you can draw the process on a whiteboard, build a workflow. If you can't, because every case comes out different, that's when an agent starts to earn its keep.

Five patterns you'll start seeing everywhere

People building with AI keep rediscovering the same handful of building blocks. Anthropic's guide Building Effective Agents gave them names, and I find them a useful vocabulary even for non-engineers.

Five workflow patterns drawn as small flow diagrams: chaining passes work through steps in sequence; routing sends each item to the right specialist path; parallel splits work to run at the same time and merges the results; orchestrator and workers has one AI break a job into subtasks for helper AIs; evaluator and optimiser has one AI draft and another critique until it passes
The five patterns. Real automations mix and match them.

The simplest is chaining, which is just an assembly line: the output of one step feeds the next. Transcribe the meeting, summarise it, pull out the action items, email them round. Then there's routing, which works like a good receptionist. The AI reads what came in and sends it down the right path, so complaints go to support, invoices to finance and new leads to the CRM.

Parallel is a team splitting the work. You might have one AI check a contract for legal risk while two others look at the financial and security angles, and you merge the three reviews at the end. Orchestrator and workers is the project manager version: one AI breaks a big job into pieces and hands them out, say planning the sections of a research report and letting helpers write each one.

The last one is my favourite, evaluator and optimiser. One AI writes, another plays editor and scores the draft against your criteria (tone, length, compliance wording), and they go back and forth until it passes. It's surprisingly good at marketing copy that has to tick a lot of boxes.

Five workflows worth stealing

Five example workflows drawn as pipelines: inbox triage, meeting to actions, invoice capture, content repurposing and customer feedback digest, each showing the trigger, the AI steps and the output
Five workflows that pay for themselves quickly

WorkflowTriggerAI doesOutputPattern
Inbox triageNew emailClassify urgency and topic, draft a reply for routine onesLabelled inbox, drafts waitingRouting
Meeting → actionsRecording finishedTranscribe, summarise, extract owners and datesSummary in the team chat, tasks in the trackerChaining
Invoice capturePDF arrives in a folderRead vendor, amount, due date, VATRow in a spreadsheet, flag anomaliesChaining
Content repurposingBlog publishedWrite a LinkedIn post, a newsletter blurb and 3 short postsDrafts for approvalParallel + evaluator
Feedback digestEvery FridayRead the week's reviews and tickets, cluster themes, score sentimentOne-page digest for the teamOrchestrator

You'll notice none of these sends anything out of the building without a person looking at it first. Same rule as the agents in Part 6, and I'd keep it that way until a workflow has earned your trust over a few weeks.

Where to build them

You don't need to be a developer for any of this. The automation platforms have all bolted AI steps on, and they're good.

The n8n home page showing a visual workflow canvas with an AI agent node connected to tools such as order lookup, refund policy and a Slack approval step
n8n: build an AI workflow by connecting boxes. Notice the human approval step on the right. Screenshot: n8n.io, October 2026.

Zapier and Make are the gentlest place to start, with thousands of app connectors and AI steps built in. n8n is a visual builder you can also host yourself, which is why technical teams who care about where their data lives tend to like it. If your company runs on Microsoft 365, Power Automate and Copilot Studio are the obvious choice. And for small personal routines you may not need any of them: ChatGPT, Gemini and Claude can all run scheduled tasks now ("every Monday at 8, summarise…").

What to automate first

Not everything deserves a workflow. I ask two questions about every task: how often do I do it, and how much judgement does it need?

A two-by-two grid: frequent tasks needing little judgement should be automated with a workflow now; frequent tasks needing high judgement get AI assistance with human approval; rare low-judgement tasks are not worth automating yet; rare high-judgement tasks stay human
Start in the top-left corner: frequent and routine

Then do the back-of-the-envelope maths before building anything:

A simple formula: minutes per task times times per week equals minutes saved per week; for example 15 minutes times 20 times a week is 300 minutes, which is 5 hours a week or about 250 hours a year
The back-of-the-envelope test. If the number is small, don’t automate it.

Five hours a week adds up to more than six working weeks a year, and that's from a single workflow. In my experience that one number does more to get an automation budget approved than any slide about the future of work.

Mistakes I keep running into

The most common one is automating a broken process. AI will make a messy process faster, not better, so fix it on paper first. Close behind is skipping the human checkpoint on anything that touches customers or money, and forgetting to log what each step did (it will go wrong at some point, and you'll want to know which step did what).

Two more are about cost and ambition. People reach for the biggest model for every step, when classifying and extracting work perfectly well on the small, cheap models from Part 3; save the flagship for the one step that's genuinely hard. And they build an agent when a workflow would have done. For repetitive work, predictable beats clever every time.

Before Friday

Write down every task you did more than three times last week. Next to each one, put how many minutes it takes and whether it needs much judgement. Circle the frequent, low-judgement ones and run the formula above. Whichever circled task gives the biggest number is your first workflow.

Next time we look at how all these agents and workflows plug into tools, and into each other. That's where MCP and A2A come in, along with a confusing extra acronym or two.

AIAI workflowsAutomationAgentic AIProductivity

More notes