Automated data entry: what to automate and what to check by hand

Many companies come to automated data entry with one question: "can we automate all the office work?". Put that way, the answer is almost always no. Office work is made of dozens of different steps. Some automate well; others, however, do not, or only in part. The real work, therefore, is telling them apart.

In this article I take a typical case: technicians' job reports. I use it to show the criteria I apply when deciding which data I can enter automatically and which must be checked by hand. The example is hypothetical. However, the steps are the ones you find in many small and mid-sized firms.

Why data entry lends itself to automation

Data entry has three strengths. It repeats, it follows rules that can be written down and it leaves a trail: the source document and the entered data can always be compared. So you can observe and measure it before you change it.

Also, much of the data already arrives in a structured form. A form filled in on a tablet has defined fields; a CSV export has named columns. In other words, they are not pictures to interpret. For data entry this is a big head start: the key data does not have to be guessed. A photographed paper sheet, by contrast, must be read first, and that is where the margin of error grows.

A hypothetical example: job reports

For example, picture a maintenance company with about ten technicians. Every job report follows the same path:

1. the technician fills in the report at the end of the job, on paper or on a tablet;
2. the report reaches the office as a photo, as a PDF or as a sheet at the end of the week;
3. someone reads it and looks up the customer and the installation in the business system;
4. they copy the date, hours, materials used and notes;
5. if something is missing or illegible, they call the technician;
6. finally they close the job, and the data is ready for whoever uses it next.

At first, each step looks small on its own. Together, though, they take hours every week. Above all, they depend on the memory of whoever does them.

Five criteria to choose what to automate

For this reason I ask five questions about each step. The answers tell me whether to enter the data automatically, to prepare the work for a person, or to leave it as it is.

Next, apply them to the example. Reading the reports filled in on the tablet passes all five. Likewise, matching each report to the customer and the installation is a good candidate, because checking against the customer records follows clear rules.

The hours worked, by contrast, have a high cost of error: a wrong hour turns into a dispute with the customer. Here data entry prepares the work, but a person confirms it. Similarly, the technician's notes need judgement. I can extract them and sort them by type; however, the person who knows the installation and the customer decides what to do with them.

Exceptions are half the job

In data entry, a common mistake is to plan only the happy path. In reality, some reports never match. The customer is not in the records, the material code does not exist, the handwriting cannot be read. How many there are in a real company, you only learn by measuring.

That is why, before setting up the processing, I collect a few weeks of exceptions. For each one I settle three things: how to spot it, who resolves it and with what information. Then data entry does something simple and useful. It puts the exception in front of the right person, with all the context ready.

In this way manual work does not vanish. Instead, it changes nature. As a result, people stop hunting for data and start making decisions. That is where the time saved becomes visible.

There is also a duty not to forget. The reports hold customers' names, addresses and contact details, that is, personal data. Article 5 of the GDPR requires them to be accurate and, where necessary, kept up to date. Data entry that copies an error into ten systems multiplies it: that is one more reason to handle exceptions rather than ignore them.

Data entry: rules or artificial intelligence?

Not everything needs AI. In fact, most of the work in the example runs on explicit rules. Checking a customer code or a material code is a check, not an interpretation. Moreover, rules always do the same thing and can be verified.

AI helps where the text is free-form: a handwritten report photographed on site, a technician's note, a customer's email attached to the report. At those points a model can extract data or classify the note. Even there, though, the result is checked before it reaches the business system.

As a result, my rule is simple. In data entry, rules come first, and AI only where rules fall short. That keeps the work cheaper, its errors predictable and its logic clear to whoever checks it.

How to measure data entry

Before changing the way of working, I measure the current one. A few numbers are enough, taken over a representative period:

• the average time from the end of a job to the report being entered;
• how many reports need a manual step;
• which exceptions come up most often;
• how many hours a week the person spends on this work.

Then, after the start, I measure the same numbers in the same way. Only that comparison shows whether automated data entry worked. It also shows where to act in the second round, especially when exceptions pile up.

Where to begin

The first step is not picking a tool. It is describing how the reports arrive today, exceptions included. In fact, the map alone often reveals double steps, or checks nobody remembers the reason for.

One last point matters as much as the rest: who owns the rules. After all, customers change, new materials arrive and internal procedures evolve. So every rule should live in a readable place, with the name of whoever may change it. Otherwise, after a few months, nobody knows why a report was rejected. At that point the automatic work stops helping and becomes one more source of doubt.

On the page about data processing and reports I describe how I set up this work: sources, calculation rules, checks and output format.

If a data entry task eats hours of your week, write to me and walk me through the steps. I will tell you which data I can enter and process automatically, which must be checked by hand and which are best left as they are.