We build AI automation for businesses, so the next sentence is against interest, which is usually where the useful sentences live: a meaningful share of the automation being sold right now should not be built, and some of what has been built is quietly costing its owners more than the manual work ever did. The problem is rarely the technology. It is that “can this be automated?” and “should this be automated?” are different questions, and only one of them appears on most proposals.
So here is the filter we actually use in scoping, in full, with the reasoning attached, so you can run it on your own operation before anyone runs a proposal past you.
The filter, in plain terms
Automate the work that is rules, routing and retyping. If a task is done the same way every time, moves information from one place to another, or exists because two systems do not talk to each other, it is a candidate, and usually a wonderful one. Data entry between platforms. Enquiry routing and acknowledgement. Report assembly. Invoice chasing. Appointment logistics. Document generation from information the business already holds. The test that identifies this category: could you write the rule down completely? If a competent temp could do the task perfectly on day one from a one-page instruction, a machine can do it perfectly forever, and the word people use for this work, “mindless”, is the giveaway. Machines were built for the mindless; removing it gives you back your team’s actual attention, which is exactly what our automation work targets.
Keep the work that is judgement, relationships and risk. Anything where the expensive outcome is a wrong decision confidently executed: pricing unusual jobs, handling an unhappy customer, anything a regulator or a lawyer would later read. The category’s signature is that the rule cannot be written down completely, because the task is precisely the handling of exceptions, context and consequences. Machines execute; they do not carry accountability, and a process should never be more automated than your willingness to own its worst output. The test we use in scoping: if this goes wrong on a Friday night, who finds out, and when? If the honest answer is “a customer, publicly”, a human stays in the loop, and the automation’s proper job becomes preparing that human faster: gathering the context, drafting the option, flagging the case, and stopping.
And keep the moments that are the relationship. Some touches are inefficient on purpose. The personal reply from the owner, the call that could have been an email, the thank-you that was typed by a person: customers can increasingly tell, and the businesses automating their warmth away are spending trust to save minutes. The distinction that matters here is between automating around a moment and automating through it. Around: the system books the meeting, assembles the customer’s history, drafts the agenda, so the human arrives prepared and present. Through: the customer receives machinery where a person used to be. The first makes the relationship cheaper to honour. The second quietly cancels it.
The false economy nobody audits
The costliest automation failures are not the ones that break; they are the ones that run. A workflow automated at the wrong boundary generates exceptions, and exceptions handled manually inside an automated process cost more than the process ever did manually, because now every correction requires understanding both the task and the system. The pattern to watch for in your own operation: a tool everyone quietly works around, a spreadsheet that has grown up beside the system to fix what it gets wrong, a team member who has become the full-time translator between the machine and reality. Each of those is an automation bill still being paid, off the books. The fix is rarely more automation; it is moving the boundary back to where the rules actually end.
The tell in any automation proposal
Ask what the proposal declines to automate. Good automation work has a boundary drawn on purpose and reasons for it; it sequences the safe, high-volume wins first and treats the judgement calls with respect. A proposal that automates everything it can reach is not a strategy; it is an inventory of what the tools can do, and you will pay to discover the difference. This is why our own scoping starts with the workflow, not the software: mapping where the hours actually go, then applying the filter above, and telling you plainly when the answer to a workflow is “leave that one human.”
How to run the filter yourself
An hour with your team gets most of the way there:
- List where the hours go. Not job titles, tasks: the ten activities that consume the most repeated time each week.
- Sort each into the three categories. Could the rule be written down completely? What happens when it goes wrong, and who finds out? Is a relationship living in this touch?
- Rank the first category by volume and pain. The best first automation is high-frequency, rule-complete and low-consequence: the win that proves the approach without betting anything.
- Write down what you are deliberately not automating, and why. This list is worth as much as the build list; it is the boundary, on purpose, and it is what you test proposals against.
- Audit what is already automated against the false-economy signs above. Sometimes the highest-return project is un-automating one wrong boundary.
What it looks like done right
The best automation is invisible in the customer experience and enormous in the back office: the team stops retyping, the enquiries never sit unacknowledged, the reports assemble themselves, and the humans spend their recovered hours on the judgement and the relationships, the work that was being squeezed. One of our distribution clients put the outcome in one sentence we have never improved on: “Fundamentally changed how our business operates.”
Get the honest read on your operation
If you want the honest read on your own operation, which hours are automatable, which should stay human, and what the recovered time is worth, our free AI-readiness review does exactly that, reviewed by a person, with you within a working day.
Questions answered
Is this just an argument for buying less automation?
It is an argument for buying automation that stays bought. The rules-routing-retyping category in most businesses is larger than anyone expects once it is actually listed, and automating it properly is transformative; we build exactly that. The filter exists because the returns live in that category, and the costs live in pretending the other two categories belong to it.
Will automation replace our team?
Built to the filter, it replaces the part of everyone's job they would pay to lose: the retyping, the chasing, the assembling. The judgement, the relationships and the accountability stay human because that is where their value is, and the recovered hours go there. Teams generally discover the machine took the work they resented, not the work they were hired for.
How do we stop an AI system making things up?
By never putting one where invention can escape. In our builds, systems that generate content work from your actual data, humans review anything customer-facing that carries risk, and the judgement-and-risk category above simply does not get delegated to a model's confidence. The question is less "how do we stop it" than "where do we never place it", which is the filter again.
Basis: our automation practice and the counted market analysis behind our AI automation pages, 2026. Client quote from our locked records, Anglers Resource.
