Saddlebags
and the Ridewhen a chatbot starts doing a job
A horse that talks is useful. A horse that can open the gate, fetch the price list and carry the result back is a different animal — and it is the same horse.
Four exercises, all in a browser tab. By the end of today you will know which of your weekly jobs are worth handing over, what has to be true before you hand one over, and you will have picked the first one.
Presented by Emile du Toit
brainitconsulting.com
You can follow all of today without it. Here is an overview of last time in five lines.
- The horse is the model — ChatGPT, Claude, and the rest. It reads what is in front of it and works out what should come next, one word at a time. That is all the trick amounts to.
- The bridle and reins are the chat box. You steer with words, and a vague instruction gets you a wandering horse.
- The saddle is your standing instructions — who you are, what you always do, and what it must never do — written once so every conversation starts properly.
- It bolts sometimes. It will state a wrong thing as confidently as a right one, and it does not know what day it is.
- You stay on. Nothing has happened yet that a person did not approve.
Everything so far has been a horse that talks. Today we put it to work.
Contents
- Before you startWhat you need from #001, and one file to have ready.
- What's in the saddlebagsTools: how it reaches the real world.
- From a trick to a rideThe agentic loop, step by step.
- When a mistake repeatsThe failure that only appears once it can act.
- A day's workThree small-business jobs, fully harnessed.
- Your first rideAn eight-step checklist for Monday morning.
Then the questions this session always raises, the glossary so far, and the picture to photograph before you go.
Before you start
- Workshop #001 done, or at least an account open at claude.ai or chatgpt.com. Everything here builds on it.
- The standing instructions you wrote in #001, in a project. If you skipped that, do it now — today makes no sense without it.
- A file from your business you would not mind uploading. A price list, a rate card, a past quote. A spreadsheet or a PDF is fine.
- One task you do most weeks that is mostly reading and writing, and that annoys you. Have it in mind before you start.
Still nothing to install and nothing to pay for. This one runs entirely in the browser tab you already have open. The setup page is there if you want to look ahead.
What's in the saddlebags
A saddled horse still can't open the gate. Tools are how it reaches the real world.
Up to now everything the model produces is words on a screen. You still copy them somewhere. Tools change that. A tool is a specific thing the model is allowed to reach for — like handing your rider a set of keys, a price list and the mail.
Typical saddlebags for a small business:
That split matters more than it looks. The model doesn't secretly reach into your bank. It says, in effect, "I would like to use the send-email tool, addressed to this person, with this text." Software you control then decides whether that happens — straight away, or only after you press approve.
Model Context Protocol. It's an agreed standard for handing a model a set of tools, so your accounting software and your AI assistant can be introduced once and get on with it. You don't need to understand the plumbing. You only need to know that connecting an assistant to your existing systems is now a normal, boring thing to do.
Lab 1 — Hang the bags on it 15 min
You already have two tools sitting in front of you and probably have not used either. Both are in the chat box you opened in #001.
First, prove it cannot reach
Ask it something that depends on a fact only you hold. Do not attach anything yet.
It will tell you it has no way of knowing, or worse, it will guess. Either way, that is a horse with no bags: it cannot go and look.
Now give it the bag
- Click the + beside the message box. On Claude choose Add files or photos; on ChatGPT, Add photos & files.
- Attach your price list, rate card, or an old quote.
- Ask exactly the same question again.
Same question, completely different answer — because this time it could read something. That is a tool, and there is nothing more mysterious to it than that.
The second bag: going and looking
On Claude, click the same + and choose Web search until a tick appears beside it. If your Claude has no such switch, it searches by itself when it helps. On ChatGPT, type / in the message box and pick Search. Then ask something it cannot possibly have been trained on:
Ask for the link every time. An answer with a source you can click is a different thing from an answer you have to trust.
Ask the horse which bag it just reached into
You do not have to take any of this on faith. Ask it directly, straight after it has searched:
On claude.ai in August 2026, it answered plainly:
“When I search the web, I use a tool called web_search — it queries a search engine and returns the top results (typically around 10), which I can then read through and, if I need the full page content rather than just a snippet, follow up with a second tool called web_fetch to pull the actual page.”
“That’s what I used in this conversation: web_search to find the pages, then web_fetch several times to pull the actual content.”
Two bags, not one. One to find out where a page is, another to go and read it. That is all a tool is: a named thing the model can reach for, that does something the model cannot do by itself.
Worth doing once, because it turns this workshop from a metaphor into something you checked. It will tell you what it has and what it just used, and the answer is neither mysterious nor impressive — which is exactly the point.
Try the same question on whatever you use. The names differ, the idea does not. And if it tells you it has no such tool, that is worth knowing too — it means the answer you just got came from memory, not from looking.
What to notice
- Nothing was installed. The tools were already there and you had not turned them on.
- The model did not get smarter. It got reach.
- Everything in the next ten workshops is this same move, made bigger: giving it something real to read and somewhere real to write.
Upload the same rule from #001: do not attach anything you would not email to a stranger. A price list, yes. A customer list with phone numbers, think first.
From a trick to a ride
One instruction, one answer, is a trick. Real work is a loop.
Ask a question, get an answer, done — that is a horse doing one trick. Watch a working horse cut a calf out of a herd and you see something else entirely: it looks, it moves, it checks what happened, it adjusts, and it keeps going until the job is finished.
That is the agentic loop, and that is what people mean by an AI agent. Not a robot. A loop.
The same loop, in a real yard
Goal: "Follow up on every invoice more than 14 days past due."
- Look. Opens the accounts tool. Finds eleven overdue invoices.
- Decide. Notices two are from the same customer — those should be one letter, not two.
- Act. Drafts ten reminders, each with the right amount, date and job reference.
- Check. One customer has no email address on file. It can't send that one.
- Round again. Puts that one on a short list for you, marked "needs a phone call", and finishes the other nine.
- Stop. Leaves nine drafts in your outbox for approval, and a note about the tenth.
Nothing in there is clever. It is what a good office assistant does. The difference is that it happens at seven in the morning and takes forty seconds.
The whole rig
Lab 2 — Watch it check its own work 10 min
One instruction and one answer is a trick. The loop is what makes it useful, and you can watch it happen in a single message.
Give it a job with a checkable answer
Paste a real, messy list — a few inquiry emails, a handful of jobs, whatever you actually have. Then:
If nothing is to hand, the three inquiry emails in #001, chapter three work perfectly well for this — paste all three in as one list.
Read the second half first
The table is the easy part. The value is in what it says after the table — the gaps, the guesses, the odd one out. That is the horse looking back at the ground it just covered.
Then close the loop yourself
Answer one of its questions and ask it to redo the table. That exchange — go, check, go again — is all anybody means by an agent. You have just done it by hand.
Notice you did not have to know the word “agent” to do this, and knowing the word would not have helped.
When a mistake repeats
The one kind of failure that only shows up once it can act.
Last session covered the ways a talking horse goes wrong: it makes things up, it goes where it thinks you meant, and it has no idea what today is. All of those are still true and none of them have got worse. But giving it saddlebags adds one more, and it is the one worth taking seriously.
It does the wrong thing quickly, and repeatedly. A loop with tools attached amplifies whatever you gave it. When it could only talk, a bad instruction produced one bad paragraph that you read and threw away. When it can act, the same bad instruction produces two hundred emails, and you find out afterwards.
Nothing about the horse has changed. What has changed is that a mistake now leaves the yard.
- Start in the paddock. Test on copied data, not the live system.
- Check the girth before every ride. Anything that spends money, leaves the building, or cannot be undone gets a human approval step. Keep it that way until you're bored of approving.
- Keep a hand on the reins. Drafts for you to send beat messages sent for you.
- Keep a log. Every tool it used, every time. If you can't see what it did, you can't trust it.
- Don't hand it the deeds. Deleting records, changing bank details, publishing to the world — those stay with a person.
Those five are not a beginner's training wheels to be discarded later. Businesses running this properly still work that way, because the rules cost almost nothing and the alternative costs a customer.
The question is never "can I trust it?" It is "what happens if it's wrong here, and can I undo that?"
Lab 3 — One bad instruction, twenty times 10 min
This is the failure that only appears once it can act, and it is worth meeting on purpose while nothing is at stake.
Give it something genuinely ambiguous
Paste twenty rows with some deliberate duplicates, some near-duplicates that are actually different people, and one with a missing phone number. Use your own if you have them — there is a ready-made set below if you would rather get straight to the point. Then let it get on with it.
Twenty rows, if you would rather not invent them
Dale’s customer list, in the state customer lists are actually in. The traps are already in here — do not go looking for them first.
Afterwards, check its work against these. There are four kinds of trouble in the list:
- Plain duplicates. Alvarez, Pike and Cathey are each one person entered twice, with the punctuation moved about. Merging those is right.
- Two people with the same name. The two Ken Whitfields are different men at different addresses. Merging those loses a customer.
- Two people at the same address. Jim and Jean Hollifield share a house and a job but not a phone number. This is the one almost everything gets wrong, and it is the one that ends with a call to the wrong person.
- A row with a hole in it. Tanner Boyd has no phone number. Watch whether it tells you, quietly drops him, or fills something in.
Do not read that list before you run the exercise. The whole value is in seeing what it decided on its own, and you cannot unsee the answers.
Look at what “tidy” meant to it
It will have made a decision — probably merged the near-duplicates. Consistently. Twenty times. Without asking. It was not wrong so much as it was never told, and a human would have stopped at row three to check.
Now say what you actually meant
What to take away
- The danger is not that it makes a mistake. It is that it makes the same mistake at speed, evenly, across everything.
- “Never delete” and “stop and ask” are the two most valuable sentences you will ever write into an instruction.
- Add both to the standing instructions from #001. They cost nothing and they hold for every job.
A day's work
Three ordinary jobs, harnessed the same way each time.
The pattern never changes. Something arrives or some hour comes round; standing instructions say what to make of it; tools fetch what's needed and put the result somewhere; a person checks the end of it. Read the table across, one row at a time.
| The job | Saddle (standing instructions) | Saddlebags (tools) | Who checks |
|---|---|---|---|
| Turn inquiry emails into estimates | Rate card, sales tax rule, tone, the jobs we don't take | Read inbox · read price list · create draft estimate | You approve every estimate before it goes |
| Monday morning numbers | Which figures matter, what "unusual" means for us | Read accounts · read bookings · write a one-page note | Nobody — it only writes a summary for you to read |
| Answering the same five questions | The approved answers, in our words, and when to escalate | Read inbox · search our FAQ · draft a reply | You send. After a month of good drafts, reconsider |
Where it earns its oats
- Work that is repetitive, written, and low-stakes on its own — but adds up to hours.
- Reading long things and pulling out the three lines that matter.
- First drafts. Always first drafts. Blank pages are expensive.
- Sorting: which of these forty emails is actually a new customer?
What not to hand it first
- Anything where being wrong is expensive, legally binding, or hard to spot.
- Work you don't already understand yourself. You can't check what you can't judge.
- The one job that's messy because your process is messy. Fix the process; then automate it.
Your first ride
Eight steps. Nobody starts by galloping.
You do not need to buy anything, hire anyone, or understand a single acronym to start. Here is the order that works.
- Pick one task you do every week that is mostly reading and writing, and that annoys you.
- Write down how you do it now — in steps, the way you'd teach a new employee on their first morning.
- Do it by hand in a chat window, three times. No automation. Paste in a real example, ask for the result, correct it. Notice what you keep having to explain.
- Turn what you kept explaining into a saddle. Write those standing instructions down in a note, then save them somewhere the chatbot reads automatically, so you stop retyping them. Session one showed you where that setting lives.
- Use it manually for two weeks. If it isn't saving you time as a copy-and-paste job, tools won't save it either.
- Only now ask what tools it needs. Usually one or two. This is the point to bring in someone technical, and you'll brief them well because you did steps 1–5.
- Keep a person on the last step for the first month. Approve, don't automate.
- Then take the second task. Not before.
Write down one weekly task — on paper, on your phone, anywhere. That is all the homework there is. Everything in this book is easier once there's a real job in front of it.
Lab 4 — Pick your first real job 15 min
The eight steps above are the method. This is you doing steps one and two, now, before you close the page. Everything in the remaining ten workshops is easier with a real job in front of it.
Write the task down
One task. Weekly. Mostly reading and writing. Annoying. Do not choose the biggest problem in your business — choose a small dull one you will actually still care about next Tuesday.
Then write down how you do it now
In steps, the way you would teach somebody on their first morning. If that is hard, let it interview you:
Then do it by hand three times
No automation, no clever setup. Paste in a real example, ask for the result, correct what is wrong. Do that three times over the next fortnight and pay attention to what you keep having to explain. That repeated explanation is your saddle, and it is what you will write into your standing instructions.
If it is not saving you time as a copy-and-paste job, no amount of tooling will save it either. Better to learn that in week one than after buying something.
Questions people actually ask
The ones this session always raises.
Can it look at my spreadsheets, my email, my invoicing system?
Only what you put in its saddlebags.
Out of the box it sees nothing but what you paste in. You can upload a file to a conversation, and you can connect it to systems properly with tools — which is what this whole session is about. Nothing connects itself. If you didn't set it up or approve it, it isn't there.
Is it going to replace my staff?
A horse didn't remove the need for someone who knows the way.
In a business your size, it takes over tasks, not people — the reading, the drafting, the sorting, the retyping. What it cannot do is know when the answer is wrong, and that judgment sits with the person who understands the work. Most owners I meet don't end up with fewer people. They end up with people spending their day on the part that needed a person.
How do I know it actually did what it says it did?
You can see every gate it went through.
Because the model asks for a tool and software you control performs it, every single action leaves a record — which file, when, what it wrote. That is the practical difference between an assistant and a black box, and it is why the "keep a log" rule is worth having from day one. If someone offers you an automation you cannot inspect afterwards, that is the thing to be suspicious of, not the AI.
Everything feels automatable. Where do I actually start?
The dullest job you do every week.
Not the most valuable one, and not the most annoying one — the most repetitive one, where you would notice immediately if the answer came out wrong. Repetition means you get many goes at improving it; obvious failure means a mistake teaches you something instead of costing you something. The next chapter is the eight steps, and step one is exactly this.
What if it sends something to a customer by mistake?
It can't, unless you gave it the postbag and told it not to ask.
Sending is a tool like any other, and it is the one to keep on a short rein longest. The sensible arrangement for a first year is that it drafts and you send — you keep the two seconds of judgement and lose none of the time saving, because writing the thing was the slow part.
Plain words
Everything the series has named so far. Today's are marked.
- Large language model (LLM) 001
- The engine that predicts the next word. The horse.
- Prompt 001
- What you type. A pull on the reins.
- Chatbot 001
- A model you can talk to, turn by turn. Horse plus bridle.
- System prompt / instructions 001
- Standing orders sent with every message. The saddle.
- Context window 001
- How much of the conversation it can hold at once. How far it can see on this ride.
- Token 001
- A chunk of a word — how length is counted and how you're billed. Sugar cubes.
- Hallucination 001
- Making something up and sounding certain. Shying at a shed snakeskin.
- Knowledge cutoff 001
- The date its training stopped. The last day it was out in the world.
- Tool 002
- A specific action it's allowed to take in the real world. A saddlebag.
- MCP 002
- An agreed standard for plugging tools into a model. Standard-size buckles.
- Agent 002
- A model given a goal, tools, and permission to keep going. A working horse, not a show pony.
- Agentic loop 002
- Look, decide, act, check, repeat. The ride itself.
- Human in the loop 002
- A person approves before something real happens. A hand on the reins.
- Automation 002
- The same work happening without you starting it each time. The horse knows the route.
Everything on one page
Photograph this before you go.
1 · The model
Predicts the next word. Knows a lot. Knows nothing about you.
2 · The chatbot
Add reins and you can steer it. Where everyone starts.
3 · The harness
Standing instructions, fitted once. Every ride starts right.
4 · The agent
Tools in the bags, a goal, and a loop. Now it does a job.
And the rider stays on. You set the goal, you set the boundary, you check the end of it. That doesn't get automated — it's the part that was always yours.
#003 — Why Leave the Field
Everything so far has happened in a browser tab, which means you have been the one carrying every fact out to the horse and every answer back. Next time we find the building where your actual files live, and why professionals do not work in a chat window. Nothing to install before you come — lids stay down.
Bring the task you picked today. We will use a real one from the room.
I do this for a living, and love to help people.
Thirty-plus years in enterprise software and a computer science degree — which mostly means I have watched a great many horses bolt, and I can usually tell you which ones are worth saddling before you spend money on the saddle.
In person
Your team, your room, your actual work on the whiteboard instead of somebody else's examples.
Over video
The same session, at your desks, with fewer chairs to stack afterwards.
Built for real
When the idea survives the workshop and somebody has to go and build the thing — that part I do too.
Fun fact — I rode to school on horseback as a kid in South Africa. So the metaphor isn't borrowed. I've done the falling off in person.
Emile du Toit brainitconsulting.com
Workshop #002 · Saddlebags and the Ride
© 2026 Emile du Toit, BrainIT Consulting · Print it, share it, leave my name on it.