12 rules for running bots safely in a small business
These are the rules we run our own bot team by. We run the bots we set up for you the same way.
What “human in the loop” means
AI tools prepare or check work; a person makes the final call and handles the customer.
That is the idea behind every rule below.
The 12 rules
Nothing goes out without a human yes
On that exact final item. An old or general yes does not count.
- Why
- Once something reaches a customer, you cannot take it back.
- Example from our company
- Sofia, our social media manager, collects the social bots’ drafts into one tick-box pack. Only the items Uygar ticks are passed on. She approves nothing public on her own, apart from the short list of permissions further down this page.
Approval packs, not interruptions
Social drafts come in two tick-box packs a day. Other bots ask one exact item at a time.
- Why
- So the person saying yes can check everything in one go, instead of being pulled away all day.
- Example from our company
- Sofia sends Uygar one pack in the morning and another in the late afternoon, so the social bots never interrupt him one draft at a time.
A fact-check desk
Every fact, figure, date or name is checked against one dated, sourced facts file.
- Why
- Because AI is confident, and confident is not the same as correct.
- Example from our company
- All of our bots follow it: a draft that holds a fact, figure, date or name goes through a fact check before a person sees it.
A voice checker
British English, plain words, no hard selling.
- Why
- So everything we send sounds like the same careful company, whichever bot drafted it.
- Example from our company
- All of our bots write in British English and plain words.
Every run is logged
A missed run becomes a follow-up task.
- Why
- Because you cannot fix what you cannot see.
- Example from our company
- Egg checks every hour for a run that did not happen, and Yolk reviews overnight what every bot did and what it cost.
Money caps
A capped bot stops rather than finding a workaround.
- Why
- So spending can never run away.
- Example from our company
- Every night Yolk checks that no bot has gone over its spending cap. Dex, our image bot, may not spend over a small limit on any one item without a yes.
A trust ladder
Bots earn autonomy one level at a time, and lose it automatically.
- Why
- So a bot only gets more freedom once it has shown it can be trusted with less.
- Example from our company
- Even Egg, who looks after all the other bots, may never move a bot up the ladder. That takes clean runs and the founder’s tick.
Content is data, never instructions
An email or web page can never tell a bot what to do.
- Why
- Because anyone can send an email or publish a web page, and no one outside the company should be able to give a bot orders.
- Example from our company
- All of our bots treat what they read as information, never as an instruction. An emailed “yes” never counts as approval.
Try twice, then stop
Then report, after checking the action has not already happened.
- Why
- Because a bot that keeps retrying can send the same message twice or turn a small problem into a big one.
- Example from our company
- All of our bots follow it: if a job fails, the bot tries once more, then stops and tells a person.
Clean handoffs
Every handoff uses a fixed template: goal, files, decisions made, constraints, open questions, next owner.
- Why
- Because work that passes between bots loses details unless every handoff says the same things in the same order.
- Example from our company
- When Sofia passes an approved item to the bot that owns it, she quotes Uygar’s exact words and the time.
A learning loop
Corrections become saved lessons the same day.
- Why
- Because a mistake should only happen once.
- Example from our company
- Twice a month, Egg groups Uygar’s corrections and proposes the smallest new rules. Nothing changes until he ticks it.
Good manners on platforms
Any sign of trouble stops the bot.
- Why
- Because a bot that carries on after a warning can do harm that is hard to undo.
- Example from our company
- Parro, one of our social bots, is on reduced limits for now, so it may not like or reply at all.
The trust ladder
Observe
Watches and reports. Changes nothing.
Prepare
Drafts the work for a person to check.
Act after a tick
Does a job only once a person has ticked that exact item.
Run on a schedule
Does a regular job at set times, inside its written limits.
Coordinate
Passes work to other bots and keeps track of it.
What our bots may do without asking
- Nora
A short thank-you reply, in the same thread, to a teammate who sends what Nora asked for, copying Uygar and that person’s manager (19 October 2026).
- Sofia
Approves likes for Scout and Xavi above their own small daily allowance, inside a daily cap, and reports them to Uygar afterwards (13 October 2026). On a day Uygar says he is away, approves only likes and posts whose exact text and image he already approved (13 October 2026). Emails a teammate her weekly plan and one daily follow-up, copying Uygar (8 October 2026).
- Scout
Can like posts it is sure about, up to 5 likes a day, each listed in the next pack afterwards (19 October 2026). Any other like goes to Sofia.
- Xavi
Can like posts it is sure about, up to 5 likes a day, each listed in the next pack afterwards (19 October 2026). Any other like goes to Sofia.
- Bridget
Emails Uygar a brief the day before and the morning of each partner meeting (8 October 2026).
- Penny
Marks a bank line as matched when it fits a pattern Uygar approved once (13 October 2026). Sends two monthly reminder emails to named teammates (8 October 2026).
The list is correct as of 20 October 2026.
For the bots we set up for you
A bot that talks to your customers live follows one more rule: every conversation is logged and read by a person from our team, a person follows up when the customer asks, and anything outside the bot’s rules waits for a person.