Layer 04 — Automate
Automation
for the work
between
your tools
We automate the repeatable parts of intake, follow-up, sorting, drafting and reporting. AI goes in only where it makes the work faster without putting judgement, privacy or a customer relationship at risk. A person still decides the things that are expensive to get wrong.
01 — Position
Most of what gets sold as AI is a chatbot on a website
A chatbot is occasionally the right answer. Far more often the valuable automation is invisible: the work that happens between an enquiry arriving and a human deciding what to do about it.
Our rule is simple. Automate the part that is repetitive and cheap to be wrong about. Keep a person on the part that is judgement and expensive to be wrong about. Where a model does something that matters, a human sees it before it goes out.
02 — Where it pays
Six jobs
worth handing over
These have the same shape: high frequency, low variance, and a person currently doing them by hand between two other tasks.
Reading and sorting what comes in
Pulling name, intent, urgency and service from a free-text enquiry, an email or a voicemail transcript, and putting the result in the right queue with the right flags.
Drafting the reply, not sending it
A first draft written against your actual answers, waiting for a person to glance at it and hit send. Minutes instead of an afternoon, without anything unreviewed going out under your name.
Summarising a long history
Six months of notes and calls reduced to the three lines someone needs before picking up the phone.
Qualifying against your criteria
Sorting enquiries by your rules for who is worth calling first — and showing why it decided that, so a person can disagree with it.
Chasing on a schedule
The second and third contact attempts that depend on somebody remembering, and therefore usually don’t happen. Not AI at all, mostly, and often the highest-value thing on this list.
Answering the same question forever
Where a chatbot does earn its place: parking, insurance, hours, what a first visit involves. Answering from your documented answers, and handing over to a person the moment it is out of its depth.
03 — Rules we build to
Where we will not put a model
Stated up front, because these constraints change what a project costs and what it can do.
Non-negotiable
- Nothing that goes out under your name is sent without a person having seen it, unless you explicitly ask for that and understand the trade.
- Sensitive information does not get passed to a vendor who has not agreed in writing to handle it.
- An automated decision that affects a customer has to be visible and reversible by a human.
- If a platform’s own policy forbids a use, we don’t build it. Some messaging platforms prohibit automated handling in regulated industries outright.
How it gets built
- The manual version of the task is written down first. If the rule can’t be stated, it can’t be automated.
- It runs alongside the human for a period, and we compare the two.
- Every automated action is logged, so you can audit what it did and why.
- There is an off switch that a non-technical person can reach.
04 — A worked example
The second
phone call
On most enquiry pipelines the first call happens reliably and the second one doesn’t, because the first is prompted by the enquiry arriving and the second depends on somebody remembering two days later.
Nothing about fixing that requires a language model. It requires the system to know a call was attempted, to know what should happen if it went unanswered, and to put that person back in front of a human at the right hour. That is the shape of automation that earns money, and it is unglamorous.