One email a week - something from which I hope you'll get real value. We talk about things we can build, and how to defend them. That can apply to cybersecurity, physical buildings, digital products, and .... just about anything. It gives me a lot of latitude in what I can write about, but the two concepts are important for progress - as individuals, and as society.
Apologies this one is late, I’m traveling in the US at the moment.
Today's topic is: identifying automating workflows with AI.
Most people get an AI agent and immediately ask the wrong question.
“What can this thing do?”
That sounds reasonable, but it usually leads to party tricks. Draft this. Summarize that. Make me a logo. Research a thing I may or may not read later.
Useful, sure. But not transformative.
The better question is:
“Where in my life or business do I already have a workflow that is repetitive, annoying, error-prone, or too easy to ignore?”
That is where agents start to matter and get interesting.
It's not because they are magical. Not because they replace judgment. Because they can sit inside a workflow and keep doing the boring part, consistently, while you stay focused on the part that actually needs a human.
That distinction matters.
A chatbot answers a question. An agent can watch, remember, fetch, compare, draft, file, remind, escalate, and hand you something closer to finished work.
But only if you give it the right job.
🔨 BUILD: Workflows
If you want an agent to become useful, do not begin with the agent.
Begin with your week.
Look for the places where work already happens in a pattern.
A decent agent workflow usually has four pieces:
1. A trigger
2. A source of information
3. A decision or transformation
4. An output
That is it. Four steps! It's like connect four but without little plastic discs.
For example:
Every morning → check weather, markets, calendar, news → summarize what matters → send me a briefing.
New customer intake form → read the submission → create a setup checklist → notify me if anything is missing.
New invoice email → extract the details → put it into a tracker → flag anything weird.
New security alert → check the asset, severity, and history → summarize the risk → tell me whether I should care.
Weekly content planning → review saved links and notes → propose newsletter angles → draft the best one.
This is much better than saying, “I want AI in my business.”
AI in your business doing what?
If you cannot describe the workflow, the agent cannot reliably help with it.
My rule of thumb:
If you have done the same thing three times and disliked it each time, it is a candidate.
Not automatically. But candidate.
The best early workflows are usually boring:
Morning briefings
Inbox triage
Meeting prep
Follow-up drafts
Research summaries
Checklist generation
CRM or spreadsheet updates
Status monitoring
Document cleanup
“Tell me if this changed” jobs
System administration
(This last one is one that I use AI for quite a bit now - it builds systems easily and can install anything into those systems you want.)
Don't fall into the trap of trying to give the agent a giant vague mission like “Grow my business.”
Instead, do something like:
“Every weekday at 8am, review yesterday’s inbound leads, categorize them by likely value, flag anything urgent, and draft my follow-up messages without sending them. Let me know once the drafts are complete in my draft folder and I'll review and send them."
That is a job.
🛡️ DEFEND: Your Permission Surface
Of course, an agent is most useful when it has access.
That is also when it becomes dangerous.
The moment you connect an agent to your email, files, calendar, codebase, CRM, bank data, cloud account, or customer records, you are no longer playing with a toy. You have created a new actor inside your environment. We talked about the permission surface before in newsletter 16. In the above example, it's the email inbox permissions - the agent has access to create emails, so it has to know at least how to connect to your inbox, with a username and password. You need to have it write, but not send.
So the defensive question is not “can I trust AI?” That's too broad.
The question is:
“What can this agent touch, what can it change, and how would I know if it got something wrong?”
A good agent workflow has a blast radius.
For early workflows, I like this pattern:
Read many things
Write drafts
Suggest actions
Require approval before anything external happens
That gives you leverage without letting the agent run around with scissors, or worse, a loaded gun.
Good first permissions:
Read-only calendar access
Read-only email search
Read-only notes/docs
Draft creation
Local file generation
Telegram or Slack notifications
Issue/checklist creation
Systems it builds itself
Be more careful with:
Sending emails
Posting publicly
Editing production systems
Moving money
Deleting files
Changing DNS
Deploying code
Modifying customer records
That does not mean “never.” It means the AI has to earn it. If it does earn it, but screws up, then it has to re-earn it. That usually takes longer.
Start with read-only. Then drafts. Then supervised actions. Then narrow automation where the failure mode is acceptable. (Of course, it you isolate the agents as I've demonstrated in past newsletters, setting out these permissions are easier.)
And logs matter. If an agent summarises something, you should be able to ask where it came from.
If it changes something, you should know what changed.
If it runs every morning, you should know when it failed.
The agent should not be a mysterious little gremlin living in your stack, occasionally producing nice messages and occasionally doing who-knows-what.
You can ask them to explain themselves. They will, though you might still have to verify. As Reagan said, "Trust but verify."
💰 STACK: Workflows
Here is the practical exercise I use for identifying workflows.
Make a list of recurring tasks in your life or business. Do not overthink it. Just write down the stuff that keeps coming back.
Then score each one from 1 to 5 on these four questions:
1. Frequency
How often does this happen?
Daily and weekly tasks are better candidates than once-a-quarter tasks.
2. Annoyance
How much do you dislike doing it?
This matters more than people admit. If a task is annoying, you will avoid it. If you avoid it, the agent may create value simply by making sure it happens.
3. Structure
Does the task follow a pattern?
Agents are better when the inputs and expected outputs are clear. “Summarize these five known sources every morning” is much easier than “figure out what I should do with my life.”
Sadly.
4. Risk
What happens if the agent gets it wrong?
Low-risk workflows are the best place to start. You want repetitions. You want practice. You want to build confidence before handing the agent sharp objects.
Then look for tasks with:
High frequency
High annoyance
Medium to high structure
Low to medium risk
That is the sweet spot.
A few examples:
Great agent workflow
Daily briefing from known sources
High frequency, useful, low risk
Output is informational
Easy to review
Good agent workflow
Drafting replies to common inbound emails
Frequent, annoying, structured
Human approves before sending
Riskier agent workflow
Automatically sending customer responses
Useful, but external-facing
Needs tight rules, logging, and review
Bad first workflow
“Manage my whole business and make no mistakes”
Vague, high risk, hard to verify
This is how you get chaos with a friendly interface
The boring scoring exercise saves a lot of time.
It also stops you from buying or building some elaborate AI setup and then wandering around looking for a problem to justify it.
(If you run a business, obviously this is a bit more complicated, because you're likely to have employees who might do a lot of these things and you can't have them stop to figure all this out, even assuming they'd get it correct. I'm working with a business that's going to solve this for people - in a really unique way - so if you're interested in finding out more or want to potentially be an early investor alongside me, email [email protected].)
🔗 LINKS
Probably the best single technical piece on when to use simple workflows vs more autonomous agents.
Useful business/operator framing for turning an agent from a demo into something with a defined job.
A more engineering-minded checklist for building agents that survive contact with real users and real workflows.
Worth watching because useful agents need safe, standard ways to connect to tools, files, and business systems.
The defensive side: once agents can touch email, files, code, or customer data, security stops being theoretical.
Interested in your own agent for yourself or your own business? Head over to workerbee.bot and register your interest. I'll get it set up for you and you can start to identify workflows and get back some time!
💬 ONE THING: Start Now
If you already have an agent in your life, give it one real job this week.
Not ten. One.
Pick something that happens often, takes attention, and does not need full autonomy.
Write the workflow down in plain English:
When should it run?
What should it look at?
What should it ignore?
What should it produce?
What should it never do without asking?
How will you know if it worked?
That last question is the one people skip. Don't be one of those people.
An agent that “helps with things” is a novelty. An agent that reliably handles a defined workflow is infrastructure.
And once you have one reliable workflow, the next one gets easier. You learn what to delegate, where the edges are, what needs approval, and what should stay human.
That is the real path.
Not replacing yourself overnight. Though that might be nice to think you could be sipping margaritas on the beach tomorrow - we're a bit away from replacing everyone just yet.
Just taking one recurring annoyance, giving it a small job description, and making it happen without needing to remember it every time.
That is where the useful stuff starts.
Thanks for reading this newsletter! Feel free to respond any time.
Thomas
Was this forwarded to you? Subscribe at builddefend.fyi.
Had enough? [Unsubscribe] - no hard feelings.*
* Well, a little bit.