Most AI sales pitches describe a future. This is about the present, which is less exciting and considerably more useful.
AI is good at a narrow band of work: reading unstructured text, drafting a first pass, classifying things, and pulling structure out of mess. Almost every genuinely valuable automation we build sits inside that band.
What is actually worth automating
Reading things so a person does not have to. Inbound emails sorted by what they are. Invoices with the numbers pulled out. Applications summarised into a row. This is the highest-value category and the least glamorous.
Drafting a first version. A reply, a quote, a description. A human still approves it, but they edit instead of starting from nothing.
Answering the same question for the fortieth time. If your team answers the same five questions daily, that is a well-defined job with a clear success measure.
Moving data between systems that do not talk. Half of what gets called AI automation is really plumbing, and it delivers most of the value.
What is not worth automating
Anything where being wrong is expensive and nobody checks. Pricing, medical, legal, anything touching money without a human in the loop.
A process you cannot describe. If you cannot write down the rules, automation cannot follow them. "Use your judgement" is not a specification.
Work that happens twice a month. Automation has a build cost and a maintenance cost. Ten minutes a fortnight will never repay either.
Relationships. Automated outreach that pretends to be personal is worse than no outreach. People can tell, and it costs you more trust than it saves you time.
The test we use
Before we quote anything in our AI and automation tier, a task has to pass three:
- It repeats — weekly at minimum
- It is describable — you can write the rules down
- It costs real hours — enough that removing it is worth the build
Fail any one and the honest answer is that you should not automate it yet.
What it looks like in practice
Not a robot. Usually a small, dull thing that runs on a schedule:
- Every inbound enquiry classified and routed, with a drafted reply waiting
- Documents read into your system instead of typed in
- A daily summary of what needs attention, assembled from three systems
- Records kept in sync without anyone copying anything
Boring, and that is the point. The impressive demos are rarely the profitable ones.
The part vendors skip
Automation needs looking after. Models change, formats change, an upstream system adds a field. Something that has run for six months will eventually stop being right, and if nobody notices, wrong data spreads quietly.
Anyone selling you automation with no ongoing cost is selling you something nobody will maintain. Ask what happens in month seven.