"It feels faster" is not a result
Relief is real, and it is a poor measuring instrument. Automations tend to remove the most irritating part of a task, the copying between two systems or the retyping of a PDF. The irritating part is often not the biggest part. People feel the difference in their mood and report it as a difference in hours.
Memory rounds. Ask three people how long it takes to handle a customer complaint and you get three answers, anywhere from five minutes to half an hour. All three are honest. One remembers the typical case, one remembers the worst one, one remembers yesterday's. A number from memory is useless as a baseline, because after go-live anyone can revise it upwards and nobody can prove them wrong.
The new work is invisible. After go-live somebody checks the output before it goes out. The cases the automation can't handle fall back to a person, and that person now spends longer on each one, because first they have to work out why the automation gave up. Then there are licences. None of this shows up in "it feels faster."
Unmeasured wins lose the next budget fight. There will be a next automation. If nobody can say what the last one did, the budget conversation becomes a contest of anecdotes, and the best storyteller wins.
Four numbers, two weeks
Before anything is built, measure the process as it runs today. Two weeks, because any single week might be the trade fair week or the week half the team is on holiday. For a process that only runs monthly, such as month-end closing, measure two cycles instead.
How often it runs. Count the triggers per week: incoming orders, supplier invoices, web form enquiries, whatever starts the process. Your ERP or a shared mailbox may already know the number. Use it, and still tally by hand for the two weeks, because some triggers arrive by phone and never touch a system.
How long one run takes. This is hands-on time, the minutes a person actually spends on one case. Time at least twenty runs spread over different days and different people. Include the colleague who is new, and don't only time the one who could do it asleep.
How often it goes wrong. Count the runs that need a second touch. A wrong article number, a missing cost centre, a call back to the customer to ask what they meant. A tally sheet next to the keyboard is enough.
Lead time from trigger to done. The order arrives Tuesday at 9:14, the order confirmation leaves Wednesday at 15:30. That gap is what the customer experiences. If twelve minutes of work sit inside thirty hours of waiting, making the twelve minutes faster changes very little for the customer, and the queue is the real problem.
The tool is one shared spreadsheet: date, trigger time, start, end, done time, rework yes or no, and a notes field. No new software, no time-tracking app.
Measure the process, not the people. Leave names out of the sheet and tell the team up front what the numbers are for. In Germany there is a legal reason as well. Under § 87 Abs. 1 Nr. 6 BetrVG, the Works Constitution Act, a company's elected works council, the Betriebsrat, has a right of co-determination over technical systems designed to monitor employees' behaviour or performance. The Federal Labour Court reads "designed to" as "objectively suitable for", so the employer's intentions don't matter. A tool that logs handling time per person automatically can fall under that. If you have a works council, talk to them before the measurement starts, not after.
One number, in writing, before the build
After two weeks you have four numbers. Pick one of them, the one the automation is supposed to move. Only one, because with four on the table you will, after go-live, report whichever one looks best. That's human. It just isn't measurement any more.
Choose it by the reason you're automating. If the team is drowning, pick hours per week. If customers complain about waiting, pick lead time. If wrong shipments are the pain, pick the rework rate. Keep tracking the other three, so you notice if the automation quietly makes one of them worse.
Write the target as one sentence. For example: "Lead time from order received to order confirmation sent drops from the measured X hours to under Y hours, measured with the same spreadsheet over two weeks." The managing director and the person who owns the process in the business confirm it by email. That is all the ceremony it needs.
Freeze the method. Same start point, same end point, same sheet, same two-week window. If the start point moves after go-live, you are measuring the new definition, not the automation.
Agree the result that ends it. Write down which result would mean switching the automation off again. It feels odd to do that before anything exists. That is the reason to do it then, because after three months of work and a signed licence there will always be a reason to keep going.
After go-live: same method, full cost
Skip the first weeks after launch. People are still learning the new routine, and the numbers describe the changeover, not the result. Then measure for two weeks again, with the same sheet.
Count the new costs. Take some made-up round numbers. Before, one run takes 12 minutes and happens 40 times a week, so 480 minutes, eight hours. After, 34 of the 40 runs go through on their own and someone spends two minutes checking each, which is 68 minutes. Six runs don't fit the pattern and get done by hand at 15 minutes each, another 90 minutes. The week now costs 158 minutes instead of 480. Convert the monthly licence into minutes at your hourly cost and add it, and add the time spent fixing and adjusting the automation. Only that total is the after.
Write the one-line statement. It always has the same shape: "[Metric] was [value] in [before period] and is [value] in [after period], measured with [method]. Running costs are [amount] per month." Attach the sheet. A managing director can check that in five minutes, and that is the whole point of writing it.
Say so when nothing moved. It happens. The bottleneck was somewhere else, or the exceptions were more frequent than anyone expected. There are two honest options: improve it against a new written target, or switch it off. Either is cheaper than a year of running something nobody can defend.
The baseline is the only part of an automation project you cannot do later. The code can be rewritten and the tool can be swapped, but once the old process is gone, so is the chance to time it. I wouldn't start a build until that one number is written down and agreed, because without it all anyone can report afterwards is how it felt.