01Rare tasks: the arithmetic seldom works
Our automation page says that if your team does the same thing more than once a week, it can be automated. That is the floor for asking the question, not a reason to build. An automation has to be built, tested, watched and repaired when a connected tool changes something, and that work does not shrink because the task is rare. Take an example, not a client: a 30-minute report that someone files four times a year costs two hours a year. It is hard to see a build paying for itself against two hours, and the automation will sit idle for months, which is exactly when a vendor changes an API and nobody notices. A one-page checklist costs nothing and gets the job done the same way every time. We hold no break-even figure and publish none, because it depends on what your hour costs and how long the build takes.
- Happens less than weekly: write a one-page checklist before you think about a build
- Count the full cost: building, testing, watching, and fixing when a connected tool changes
- A task that costs a couple of hours a year rarely repays a build
- We publish no break-even figure, because it depends on what your hour costs
02Judgment calls: automate the preparation, keep the decision
Some work looks repetitive and is not. A discount request, a hiring shortlist, whether to keep a client who pays late and raises his voice: the form is the same every time, the answer is not. Our automation page lists work that needs human judgment on every case, and any compliance decision with legal accountability, as not a good fit. Our AI assistant page adds medical diagnosis, legal advice and financial recommendations. A rule can only return what someone wrote into it, and when it is wrong it is wrong at volume, with your name on it. What you can automate is the work around the decision: pulling the order history, the payment record and the latest messages onto one screen, so the person decides in one look instead of five tabs. The handoff rules in our post on what happens when a chatbot cannot answer follow the same logic.
- Discounts, exceptions and complaints: a person decides, a system gathers the facts
- Compliance decisions with legal accountability stay with a named human
- Medical diagnosis, legal advice and financial recommendations are listed as not a good fit on our AI assistant page
- Automate the lookup and the reminder, not the verdict
03Moving rules and broken steps: automation copies what you have today
Two different problems share one cure, which is to wait. If the rule changes every few weeks, the automation keeps enforcing the old version until someone remembers it exists. That is worse than a slow manual job, because nobody sees it going wrong. If the process is broken, the damage is bigger. Our automation page says a broken process runs faster when automated, and that disagreement about the steps, inconsistent data or weekly rule changes get executed faithfully at scale. A hypothetical example: invoices go out whenever someone remembers, and half carry the wrong price. Automating the sending delivers the wrong price on time, every time. Fix the steps by hand first, run them for a while, and automate the version that survived.
- A rule that changes monthly: wait until one version has held through a few cycles
- The team disagrees on the steps: settle that in writing first
- Inconsistent data: clean it by hand before a system reads it
- Run the fixed process manually, then automate the version that survived
04What to do while it stays manual
Staying manual is not standing still. Write the process on one page, even if it comes out ugly. Time it for two weeks, so you have your own number instead of a guess. Put a review date in the calendar and a trigger beside it: it now happens weekly, two people get the same result from the page, the rule has stopped moving. When the trigger fires, the process is a candidate. Until then a checklist, a mailbox rule or a template reply does most of the work at no build cost. Our post on the signs your business is ready to automate covers readiness in general, and our post on which processes to automate first covers the order once you have candidates. This post goes further on three cases those posts only touch: the rare task, the broken process and the decision that needs judgment, with a trigger for coming back.
- Write the process on one page and time it for two weeks
- Set a review date and a trigger: weekly volume, two people getting the same result, a rule that has stopped moving
- Use a checklist, a template reply or a mailbox rule in the meantime
- When a process passes the trigger, the signs and the order are in the two linked posts
05Where PROOF48 is the wrong choice
AUTOMATE costs 2,997 EUR a month. It includes an AI website chatbot that qualifies and books, AI DM and email auto-responders, booking automation with a CRM, automated review management and two custom automations a month. It does not cover WhatsApp or phone. If your list of repeating work is short, rare or still moving, that plan is the wrong purchase, and a checklist is the right one. Our automation page says we map the process before touching the code and say so if it is wrong, and our AI assistant page says we refuse builds we can see failing from day one. We hold no count of the requests we have turned away, so we state none.
- AUTOMATE: 2,997 EUR a month, website chatbot, DM and email responders, booking with a CRM, two custom automations
- AUTOMATE does not cover WhatsApp or phone
- A short, rare or moving list of tasks: skip the plan and use a checklist
- We map first, and say so before building if the process is wrong
A rule a machine can own, and decisions it should not
Korean Cosmetics (kcosmetics.ro) shipped one rule a machine can own: Stripe checkout re-prices every line on the server, so prices sent from the browser are never trusted. The same entry lists decisions with legal weight: every product claim rewritten in all 12 languages for EU cosmetics regulation, and one product pulled from sale for exceeding an EU ingredient limit. We would not hand a call like that to a rule running alone, and the entry does not say how those calls were made. The store is in pre-launch, so we publish no sales figures.
See the build: Korean CosmeticsQuestions, answered
How often does a task have to happen before it is worth automating?
Our automation page says that if your team does the same thing more than once a week, it can be automated. Can is not should. Weigh the build, the upkeep and what your hour costs. We have no measured break-even point and publish none.
Can I automate part of a process that needs judgment?
Yes, the part around the decision: gathering the facts, sending the reminder, logging the outcome. The decision itself stays with a person, and in a chatbot the handoff to that person is set before launch.
What if my process is broken and I want to automate it anyway?
We map the process first. Our automation page says that if the process is wrong, we say so before we build. Fix the steps by hand, run them for a while, then come back with the version that held.
Does PROOF48 turn automation work down?
Our AI assistant page says we refuse builds we can see failing from day one, and our automation page says we tell you before we build when a process is wrong. We publish no count of how often that has happened.
Where PROOF48 fits
Keep a process manual while it is rare, depends on judgment, keeps changing or is broken. Automate the preparation around a decision, never the decision. Write the process on one page, time it, and set a trigger for coming back to it. PROOF48's AUTOMATE plan costs 2,997 EUR a month, and our automation page says we map first and tell you when a process is not ready.
Cosmin Tiu · Founder, PROOF48
Cosmin has worked in online marketing and software development since 2018, and has built with AI every day since January 2023. He still writes code from scratch daily, in Python, React, Rust and C#. PROOF48 ships business websites in 48 hours and invoices afterwards, so every opinion on this blog was paid for in real builds, not in theory.