Measure workflows before introducing AI: log one week

To find out where the time goes in an office, note for one week how often each recurring workflow occurs and how long one case takes. Frequency times minutes gives the weekly effort. Only then is it worth asking whether AI can help with one of these workflows. A suitable workflow occurs often, consists mostly of reading and compiling text, and ends in something a person can check.

1. Make a list Name 5 to 10 recurring workflows
2. Log for a week Minutes and role per case, no names
3. Calculate Example: 5 × 45 min = 225 min a week
4. Check suitability Frequent, text-heavy, checkable
5. Choose one One workflow for the first trial

Text version of the graphic

• 1. Make a list: Name 5 to 10 recurring workflows
• 2. Log for a week: Minutes and role per case, no names
• 3. Calculate: Example: 5 × 45 min = 225 min a week
• 4. Check suitability: Frequent, text-heavy, checkable
• 5. Choose one: One workflow for the first trial

Which workflows come into question

Almost every office in construction has the same workflows, just in a different mix. These make a good starting list for your week:

Business	Recurring workflows
Design practice	Meeting minutes, site meetings and site diary, checking interim invoices against measurements, tender and fee texts, searching for the valid drawing revision
Client-side project management	Checking interim and final invoices, checking variations, progress meeting and minutes, monthly report to the client, searching the project files
Building contractor	Incoming invoices up to approval, daily reports and defects, tender texts, submitting variations, sorting enquiries and documents

Cross out whatever you do only once a year. Note what your people call tedious anyway. Five to ten workflows are enough.

Logging for one week

You need no software for this. A spreadsheet or a sheet of paper at the desk is enough. For each case you enter three things: the workflow, the minutes from start to approval and the role of the person who handled it, such as project lead or office. You do not need names. Count working time only. Waiting time, for example for a reply, goes into a separate column, because it is not part of the effort that an AI could reduce.

At the end of the week you calculate, per workflow, the number of cases times the average minutes per case. The result is the office’s weekly effort for that workflow. If you multiply it by your firm’s hourly rate, you also see what the workflow costs in euros.

For orientation, here is an example for a project lead in client-side project management. The quantities and minutes are the author’s working assumptions, not measured, and serve only as a template for your own table:

Workflow	Cases per week	Minutes per case	Minutes per week
Check a variation	3	45	135
Check an interim or final invoice	5	45	225
Search for a document	10	15	150
Sort an enquiry	20	4	80
Total			590

590 minutes are almost ten hours a week. That is an assumption that may or may not hold for your office. A project lead who checks fewer invoices and runs more meetings ends up with a different table. That is exactly why you log. The example table leaves out meeting minutes and reports. According to the author’s working assumption, a set of minutes that you write yourself takes about four hours (30 minutes of preparation, 90 minutes of taking notes, 120 minutes of writing up), so four of them a week would come to 16 hours on their own.

per caseper week

• Check an invoice5 cases a week

  per case 45 min
  per week 225 min

• Search for a document10 cases a week

  per case 15 min
  per week 150 min

• Check a variation3 cases a week

  per case 45 min
  per week 135 min

• Sort an enquiry20 cases a week

  per case 4 min
  per week 80 min

Source: The author's working assumption, not measured (client-side project management, per project lead)

Text version of the graphic

• Check an invoice: per case 45 min, per week 225 min (5 cases a week)
• Search for a document: per case 15 min, per week 150 min (10 cases a week)
• Check a variation: per case 45 min, per week 135 min (3 cases a week)
• Sort an enquiry: per case 4 min, per week 80 min (20 cases a week)
• Source: The author's working assumption, not measured (client-side project management, per project lead)

The comparison shows what the table makes easy to overlook. Sorting an enquiry takes the least time per case at four minutes, and yet 20 enquiries a week add up to 80 minutes. Checking an invoice costs as much per case as checking a variation, 45 minutes, but with five cases instead of three it adds up to 225 minutes instead of 135. For the weekly effort, frequency counts as much as the duration of the single case.

How to recognise a suitable workflow

Three properties speak for a workflow as a first trial.

It occurs often. A workflow with 20 cases a week brings, at the same saving per case, ten times what a workflow with two cases brings. You also see after a few weeks whether anything changes.

It is text-heavy. Language models read, compare, sort and formulate. So they prepare: a deviation list for a variation, a draft of minutes, a checklist for an invoice. Calculating should be done by a fixed program or your spreadsheet, not by the language model. Why that is so is explained in the article Language models: what they can and cannot do.

It is checkable. At the end there is a result that a person can compare with the original and approve. Where a mistake goes unnoticed and only comes to light weeks later, the workflow is unsuitable for a start. The checking time never disappears. Anyone who leaves it out of the calculation overestimates the benefit. Which workflows you should not convert in the first place is covered in the article When you should not automate a process.

It is also worth a look at the data. For the first trial, choose a workflow without particularly sensitive information. Which data may be processed on which platform is something you clarify with your data protection officer before the start.

What to bear in mind when logging

Note per case and not per person. Then the table describes a workflow and not the performance of individual staff, and the team can see that the subject is the workflow. Legally, that settles nothing yet, so here are three general points (position as of November 2026).

Data protection: Employee data protection law governs when personal data of employees may be processed for the purposes of the employment relationship (section 26 of the German Federal Data Protection Act, BDSG). The provision is still in the statute in November 2026. On 30 March 2023 (C-34/21) the Court of Justice of the EU ruled on a Hessian provision that, according to the Hessian data protection commissioner, largely corresponds to section 26(1) first sentence BDSG. National rules on employee data protection must, according to the ruling, be left unapplied if they do not meet the conditions and limits of Article 88(1) and (2) GDPR, unless they are a legal basis within the meaning of Article 6(3) GDPR that meets the requirements of the Regulation. On 8 May 2025 (8 AZR 209/21) the German Federal Labour Court held, in a case about testing HR software, that section 26(1) BDSG is left unapplied. The general lawful bases of the GDPR then come into consideration. I have not checked whether a draft bill for a separate employee data protection act exists. How far all this carries for your logging is for your data protection officer to assess.

Co-determination: An existing works council has a right of co-determination, insofar as there is no statutory or collective-agreement rule, on the introduction and use of technical devices intended to monitor the behaviour or performance of employees (section 87(1) no. 6 of the Works Constitution Act, BetrVG). Whether your logging falls under this depends on how you carry it out, for example whether software records the times. If you conclude a works agreement with the works council, a rule of this kind must, according to the Court of Justice of the EU (19 December 2024, C-65/23), meet the requirements of Articles 5, 6(1) and 9 GDPR, and the courts may fully review whether the processing is necessary. The works council, your data protection officer and your legal adviser clarify that in the individual case.

AI literacy: Under Article 4 of the EU AI Act, providers and deployers of AI systems take measures to support the development of AI literacy among their staff and other persons who work with AI systems on their behalf. They do not have to guarantee a particular level for any individual. The duty has applied since 2 February 2025, and supervision by the national market surveillance authorities since 2 August 2026. You do not need this yet for the measuring week, but you will for a later trial. On its page for companies, the German Federal Office for Information Security (BSI) states that generative AI brings new IT security risks alongside its opportunities; its publication on the subject can serve as a basis for a systematic risk analysis.

Talk to the team beforehand, explain what the measuring is for and share the results.

What the measurement does not do

The table shows where time remains. It does not show how much of it an AI can take over. You get that figure only from a trial with the same log before and after, as described in the article Measuring time savings: cases times minutes, before and after. Nor does the table measure quality: a workflow that takes only 20 minutes today because nobody looks closely does not automatically get better with AI. And it replaces no decision. Which workflow you tackle first remains up to you.

What you can do now

1. List five to ten recurring workflows in your office, separated by role: project lead, site supervision, site manager, office.
2. Set up a table with the columns workflow, date, minutes, waiting time and remarks, and present it to the team.
3. Log for one week and calculate frequency times minutes per workflow at the end.
4. Check the three biggest items against the properties frequent, text-heavy and checkable, and choose one for the first trial.