What can be automated in your business? and how to tell that this particular process qualifies

01

How do you recognise a process that is ready to automate?

The question is not whether the technology can handle it. It almost always can. The question is whether anything countable is left after go-live, and whether the flow survives its first exception. That is why we score a candidate on five traits at once, not on one.

The second trait has changed over the last two years. A process used to need a hard rule: if A, then B. Today it is enough that it relies on reading content to a fixed pattern, because an AI agent pulls the data out of an email or a PDF and writes it where it belongs. That opens up a whole category of tasks that used to bounce back to a person.

• Repeatable: it happens at least a few times a week, always in the same order
• Rule based or reading based: you can write it as „if this, then that”, or point at what you are looking for in a document
• The data is in a system: in the CRM, in accounting, in a spreadsheet or in a database, not in one person's head
• The result is measurable: hours, number of operations, number of corrections, or time from event to result
• The cost of an error is acceptable: a mistake is visible and can be reversed before it does damage

02

How do you check in five minutes whether a process is a candidate?

You run this test on paper, one process at a time. „I do not know” counts as a no. Five yes answers means a candidate for a pilot. Four means you first close the missing piece, usually the data or the measurement.

1. 01

   How many times a week?

   Count it from the calendar or from a system, not from memory. Below three repetitions a week the gain rarely covers the cost of keeping the flow alive.

2. 02

   Can you explain it to a stranger?

   Write the instruction in ten points. If „it depends who is doing it” shows up by point three, the process is ready to be described, not to be automated.

3. 03

   Where does the data sit?

   Name the system and the field. Data living in inboxes and on desktops counts too, but then step one is tidying it up rather than building anything.

4. 04

   What will you measure after 30 days?

   Pick one number you can read out of a system without asking anyone: hours, number of tickets, response time, or number of corrections.

5. 05

   What happens when the automation gets it wrong?

   „The customer receives the wrong document and nobody notices for a week” does not disqualify the process. It means a human approval step belongs at the end.

03

Which areas of a company get automated most often?

The table below is what we see in our workshops, not market research. The last column says where the effect shows up, because every area measures it differently: finance in hours, sales in response time, inventory in the number of stockouts.

This is not a shopping list. A company with a team of twenty takes one row from the table, not six. The next row gets added after the 90-day measurement, not before.

04

Where does the first step sit in each area?

In sales the effect shows up fastest in response time. An analysis of web form enquiries published in Harvard Business Review in 2011 found that an attempt to make contact within an hour gives close to seven times the odds of reaching a decision maker compared with an attempt an hour later. The automation does not sell for your rep. It only makes sure the lead does not sit there until Monday.

In customer service the line runs elsewhere. Classifying and assigning a ticket happens without a person from day one, while the reply to an unusual complaint is still written by a human. We set up AI customer service so that this line lives in a rule instead of in somebody's head.

In finance, HR, inventory and reporting the third trait is met before you start, because the data already sits in systems. Assembling one summary from several sources takes 3 to 5 hours a week in our own quoting, and the questions between reports are picked up by BOND, our company AI agent that answers from your company data in 5 to 10 seconds.

• Sales: lead routing by a rule, instead of by whoever spots the form first
• Customer service: ticket categorisation, where an AI agent reads the email and sets the priority and the owner
• Finance: invoicing from the CRM, so data from a closed deal reaches your invoicing system without retyping
• HR: employee onboarding, where accounts, documents and the task list appear on the day the contract is signed
• Inventory: stock updates, so the shop, the warehouse system and the marketplace listing show the same number
• Reporting: a weekly report assembled from several sources in one fixed format, with nobody in the middle

05

What should you NOT automate?

This list matters more than the catalogue above. Failed implementations almost always started with one of the four items below.

On decisions, two pieces of law are worth knowing. Article 22 of the GDPR gives a person the right not to be subject to a decision based solely on automated processing where it produces legal effects concerning them or similarly significantly affects them. The AI Act puts recruitment and employee evaluation among high risk uses that require human oversight. In practice this means one thing: the system prepares the decision, a person signs it.

• A process that changes every month: the flow gets rewritten more often than it runs. Stabilise the process first.
• A decision somebody is accountable for: hiring, firing, trade credit, a concession on a complaint. The system prepares the material, a person decides.
• A task done once a quarter: four times a year at two hours each is eight hours a year. Maintaining the integration eats more than it gives back.
• A process nobody can describe end to end: automation will freeze the mess in place and give it speed.

06

Which process do you start with when several of them fit?

When several processes pass the test, take the one you can launch without waiting for anybody outside the company. The order is always the same: one process in one department, a pilot on real data, production, then a measurement after 30 and after 90 days. We take at most two implementations at a time, so a pilot gets attention rather than a place in a queue.

The table lines up four typical first processes against what each one needs on the input side. Anything missing from the last column is a job for the coming week, before the conversation about tools even starts.

All four share one trait: the first measurement can be taken after 30 days without asking anyone to fill in a survey.