Should you aim to automate 100% of a process?
No. Aim to automate the routine steps fully, and design a review step for anything a customer will see, anything that moves money or decides something about a person, the exceptions your rules do not cover, and anything AI wrote or decided. The last few percent of automation is the hardest and most expensive to build, and it is where the mistakes that matter tend to happen.
On this page
Key takeaways
- Aiming for 100% automation usually costs more and ships later, and the last few percent is where the mistakes that matter tend to happen.
- Keep a person on anything a customer will see, anything that moves money or decides something about a person, the exceptions, and anything AI wrote or decided.
- Ask four questions about each step. If any answer is yes, the system prepares the work and a person approves it, and if all four are no, automate it fully.
Looking at project briefs out there, aiming for 100% automation is definitely a new craze. I recently lost out on a project when I suggested that a system taking in and paying vendor invoices and generating marked-up client invoices, relying on AI for cross-channel parsing, multi-currency conversions and different VAT rules, should not be 100% automated. The system had real money moving through it, and multiple parties who could be impacted by one small mistake or hallucination.
A lot of clients come to me assuming that's the goal, and I understand why, as taking people out of the process completely sounds like the biggest saving. In my experience though, the last few percent is usually the hardest and most expensive part to automate, and it's where the mistakes that matter tend to happen, like an email going to the wrong customer, a payment approved on bad data, or an AI summary nobody checked. I like to design a review step in from the start, at least when someone is transitioning from fully manual. This guide covers what I usually keep human, what makes a good review step, and a four-question review test for deciding how each step in a process should run.
3 ways A step can run: automated, reviewed or manual. Most processes end up as a mix of all three.
4 kinds The steps I keep human: what customers see, money and people, exceptions, and anything AI wrote or decided.
1 test Four questions, asked of each step, decide whether a person checks it before it goes out.
Why not automate all of it?
You can get close, but the last few percent is usually the hardest part to build, and it's where the mistakes that matter tend to happen. A review step lets you go live sooner and keeps a person between a mistake and the people it would reach.
Aiming for 100%
- Every edge case needs its own rule
- The build takes longer and costs more
- Mistakes go straight out
- People stop trusting it
With a review step My default
- Rules cover the common cases
- It goes live sooner
- A person catches what looks wrong
- What they catch improves the rules
What are the three ways a step can run?
Most processes end up as a mix of all three, so the question to ask is which way each step should run, instead of whether the whole process is automated.
- Automated Runs on its own, and nobody needs to touch it.
- Reviewed The system does the work, and a person checks it before it goes out.
- Manual A person does it, with the system bringing them what they need.
What should stay human in an automated process?
Four kinds of step are where I keep a person involved, at least until the team trusts the output.
Anything a customer will see
Emails, reports and documents can be drafted automatically, and a person approves them before they go out, at least until the team trusts the output.
For examplePersonalised report emails wait in a queue, and someone approves each batch before it sends.
Money, and decisions about people
The system can screen out clear-cut cases, and a person makes the call on anything that pays out or is close to the line.
For exampleApplications that meet every criterion are flagged, and a reviewer confirms each approval.
Exceptions the rules don't cover
Anything that doesn't fit goes to a person, with what they need to decide in one place, and their decision can become a new rule later.
For exampleAn invoice without a purchase order number lands on a review list with the original attached.
Anything AI wrote or decided
AI is very good at drafting, sorting and summarising, and it can still be confidently wrong, so a person checks its output before anything acts on it.
For exampleAI drafts the reply to a customer email, and a person edits it and sends it.
The last few percent is usually the hardest and most expensive part to automate, and it's where the mistakes that matter tend to happen.
Klara Condac
What makes a good review step?
If reviewing is slow or awkward, people skip it, and then the step exists on paper only. Four things keep it quick enough to be used.
Your ticks are saved in this browser only. Nothing is sent anywhere.
Should a person check this step?
Ask these four questions about each step.
- Will a customer, or anyone outside the team, see it?
- Does it move money or decide something about a person?
- Would a mistake be hard to undo?
- Does it rely on AI, or on judgement that's hard to write as a rule?
If any answer is yes, the system prepares it and a person approves it. If all four are no, automate it fully.
What does this look like for supplier invoices?
A finance team that types every invoice in by hand
A finance team receives invoices by email and types each one into the accounting system by hand.
Automated
- Invoice details read from the email
- Matched to the purchase order
- Entered into the accounting system
A person checks Reviewed
- Anything that doesn't match
- Invoices from new suppliers
- Amounts above a set limit
How to measure it. The share of invoices that go through untouched, and errors caught before payment.
Questions people ask
Should I aim to automate 100% of a process?
Usually not. The last few percent is the hardest and most expensive part to automate, and it is where the mistakes that matter tend to happen, like an email going to the wrong customer or a payment approved on bad data. Automate the routine steps fully and design a review step for the rest.
What should stay human in an automated process?
Anything a customer will see, money and decisions about people, exceptions the rules do not cover, and anything AI wrote or decided. In each case the system can prepare the work, and a person approves it before it goes out.
What is a review step in an automation?
A point where the system has done the work and a person checks it before it goes out. A good one keeps everything waiting in one place, shows the original and why it was flagged, lets people approve, edit, override or resend without leaving the queue, and shows which rules to add next.
How do I know whether a step needs a person?
Ask four questions: will anyone outside the team see it, does it move money or decide something about a person, would a mistake be hard to undo, and does it rely on AI or on judgement that is hard to write as a rule. If any answer is yes, a person approves it. If all four are no, automate it fully.
How do I measure whether the automation is working?
Look at the share of items that go through untouched, and at the errors caught before they cause harm. For supplier invoices, that means the share of invoices that go through untouched, and errors caught before payment.
More resources
A note from me
Not sure where people should stay involved in yours?
I design systems to run on their own, and to hand the right decisions to a person. If you're planning an automation and want a second opinion on where people should stay involved, feel free to send me a message.