Most companies can only afford to automate one thing first. The two candidates are almost always the same pair: the queue of repetitive requests, and the pile of paperwork somebody types into a system.
They do not pay back at the same speed, and the difference is not about which is bigger. It is about how clearly the rules are written down.
The paperwork usually goes first
Paperwork has a shape. An invoice has a supplier and a set of line items, and every invoice has them somewhere. A model reading one is doing extraction, which is a solved problem with a measurable error rate.
That measurability is what makes it pay back fast. You can hold back the low confidence ones for a person, watch the error rate for a fortnight, and know exactly what you have bought.
What a person still does
Somebody reviews the extractions the model flagged as uncertain, which is a fraction of the pile rather than all of it. That review is the thing that makes the rest trustworthy.
The queue is harder than it looks
A support queue looks repetitive because the questions repeat. The answers do not. Two customers asking the same thing often need different replies. One is inside their contract and the other is not, and that rule usually lives in somebody's head.
Automation does not fail because the model is weak. It fails because nobody wrote the rule down.
| Paperwork | The queue | |
|---|---|---|
| Input shape | Consistent enough to extract | Whatever the customer typed |
| Rules | Usually already written down | Usually in somebody's head |
| Error visible | Same day, against the source | Weeks later, from a complaint |
| Time to payback | Short, and measurable | Longer, and harder to prove |
How to tell which one you have
- Can somebody write the rule down in an afternoon, or does it need three people in a room
- Is the input the same shape every time, or does it arrive however the sender chose to send it
- When the process goes wrong, does anybody find out that week
The one with the clearest rules goes first. A pilot dies when the messy one goes first, because the model gets blamed for a policy nobody had agreed.
The rules you cannot write down are the rules you are not ready to automate.
Both eventually get done. The only question is which one earns you the confidence to fund the second one.
Wondering what this would take against your own systems?
The review is a fixed price, and you keep the costed plan and the risks whether you go ahead or not.
Book an automation reviewCommon questions
Can we automate both at once?
You can. On a first project it is usually a mistake. Two builds at once means two sets of rules being argued about at the same time. Neither result comes out clean enough to make the case for funding a second one.
How long before we know whether it worked?
For document work, about a fortnight of real traffic. That is long enough to see the error rate settle and short enough that nobody has committed to a second phase on a guess.
Hitasoft engineering
LinkedInHitasoft has built software since 2008, for companies that mostly do not have an IT department. These pieces are written by the people who do the integrations rather than by anybody in marketing.
