Process documentation brings together knowledge from many heads, and the people who hold it are kept busy by their day-to-day work. AI process documentation starts right here: the AI asks the questions itself. Each person talks about their work in a short conversation, and the documentation is written from the answers.
This redistributes the work, because the largest effort in any process documentation lies in gathering the knowledge. Phrasing it takes far less time. This article shows what actually happens in an interview, what the finished step-by-step documentation contains and where the limits are. It also covers how the approach differs from process mining, wikis, chatbots, screen-recording tools and consultant interviews.
Three Generations of Process Documentation
Each of the three generations has its strengths, and each asks for the most effort at a different point.
Generation 1: Word templates. Someone fills out a template and files it on the shared drive. The template provides a clear structure, while most of the work lies before and after it. One person writes down knowledge that is spread across twenty heads, on top of their actual job, and then keeps it current. Often three solid process descriptions come out of this before day-to-day business takes up more room again.
Generation 2: Diagramming tools. BPMN editors and flowchart software make workflows clear and precise. Getting the knowledge out of people's heads, putting it into boxes and keeping it current is still a human task, because the tool draws what it is given. The notation also takes some practice. So the people who know a process best often cannot check or correct a BPMN diagram themselves.
Generation 3: AI that interviews. The AI holds a conversation of roughly 15 to 20 minutes with each employee, via chat and without preparation. The documentation emerges from the answers on its own. People who have never drawn a process diagram contribute simply by describing how they work.
The effort moves from gathering to review, which usually goes faster, because reading and correcting takes less time than writing. Compared with workshops or observation, the one-on-one interview is a good way to capture the real as-is process, because each person describes their own part in detail. With AI, it now reaches many people at once.
This is the approach behind wissa.ai, and you are welcome to try it with the free test package.
What Happens in Those 15 to 20 Minutes
"The AI interviews your employees" means something different to everyone, so here is how it works in detail. People receive a personal invitation by email and start the conversation with one click in their browser, then answer in a chat. The conversation can be paused and resumed later.
The first minutes establish context: what is your role, what lands on your desk in the morning? Then it moves into the work itself, in everyday language rather than process vocabulary:
- What reaches you, and from whom?
- What do you do with it, step by step?
- Which system do you work in, and what happens alongside it on paper, by email or in conversation?
- What is the most common exception, and how do you handle it?
- Who does this when you are on holiday, and who do you hand off to?
Much of the insight comes between the questions. The AI follows up as soon as an answer leaves a gap:
AI: You said the batch is released as soon as the lab gives the green light.
How do you find out?
A: They call. Or it's written on the routing slip when they drop it off.
AI: And if neither happens because the lab is under time pressure?
A: Then I call them myself at some point. Happens twice a week.
AI: Who would do that if you weren't there?
A: Good question. Ms Wendland, I suppose. Whether she knows she has to chase
it, I honestly couldn't say.
These few lines contain an unregulated control step, a failure mode with a frequency and an open question about cover. Such things rarely come to mind when someone writes a work instruction. The AI then suggests Ms Wendland for a conversation of her own.
Where the time goes
The time is split between the people who know the process and the person who coordinates the documentation. It helps to plan both honestly from the start:
| Who | What they do | Time |
|---|---|---|
| Each interviewed employee | One conversation by chat, whenever it suits them | Roughly 15 to 20 minutes |
| Coordinator | Set up the project, approve suggested invitations, decide contradictions | Usually a few minutes per decision |
| Process owner | Read, check and correct the documentation | Rule of thumb: about an hour per five or six processes |
Nobody has to write minutes or consolidate notes, because the AI turns the conversations into documentation. The time that remains is the time only people can give: telling how the work runs, and deciding what counts.
From Interview to Step-by-Step Documentation
At the end of the interviews there is documentation per process: the steps in order, the role responsible for each step, the systems involved, the handovers and the exceptions. Every statement carries its source, meaning the person, the conversation and the question round it came from. Several mechanisms make this possible.
The snowball principle. When someone says in an interview "Max and Hanne handle that," the AI suggests both for the next conversation. You approve the invitation with one click, or let such invitations go out automatically. It works its way from one knowledge holder to the next, following the actual handovers in daily work. Along the way, someone may surface who was on no project list at all.
The overview and the structure emerge with it. From the same conversations, the system builds an overview of all areas and processes with roles and handovers, the basis for a process map. It also builds an org chart based on how people actually work together, and a tool register. You do not have to define a structure up front.
Living documentation. New conversations flow into the process documentation, and every new version comes with a diff against the previous one. The record grows with you. You can start with five conversations, run ten more later and build on what is already there.
Contradictions become visible. Management says "it runs through SAP", and the team says "it runs through three Excel spreadsheets". Both statements then stand side by side with their sources, and you decide which one counts. Often that very contradiction brings the most telling insight, because it marks the spot where the intended process and the lived one part ways.
Insights with evidence. Bus factor and knowledge risks ("only one person knows customs clearance") are each traceable back to the interview round. The same applies to duplicate work, manual handoffs between systems and shared tools. That traceability helps a department trust the documentation. "That's not how it works" can be answered with "it comes from the conversation on 12 March."
Onboarding included. For each role, an onboarding plan is generated: processes in learning order, systems along with the colleagues who know them, points of contact. Screenshots that people share during their interviews are embedded in the documentation.
Example: A Food Producer With 45 Employees
Composite example: it combines workflows and findings that are common in food businesses of this size. The company and the figures are a worked example, not a customer project.
A regional deli food producer with 45 employees makes salads and spreads for retail and food service. It is IFS-certified and runs three shifts, plus quality assurance, purchasing, sales, shipping and administration. The production manager has been there for 19 years and retires in two.
Much of what he knows is written down nowhere. Which supplier is unreliable for which raw material? When is a batch held back despite green lab values? How is the line actually changed over between two recipes? The IFS documentation is tailored to the audit, and the work in between relies on the tacit knowledge of the people involved.
They started with four interviews: production manager, shift supervisor, QA and purchasing. Those produced eight further names. Ten people were interviewed, including two from shipping nobody had thought of. In total there were fourteen conversations across three weeks, 15 to 20 minutes each. That is roughly four hours of staff time, plus about five hours of management time for reading, deciding and checking.
The result was 21 documented processes, six of which had appeared in no overview before. Among them were complaint handling with the retail partner and the release of recipe changes. Nine processes had a single knowledge holder, seven of them the same person.
Two contradictions were particularly telling. Purchasing assumed QA documented supplier assessments, and QA assumed the opposite, so for a year and a half nobody had taken on the task. Shift handover ran through a logbook on early shift but verbally on late shift. That explained the recurring uncertainty about the status of half-finished batches. By rough estimate, an approach with two workshop days plus write-up would have taken 60 to 80 hours for the same coverage.
What AI Process Documentation Is Often Confused With
AI process documentation shares ground with several established methods. Each of them is built for a particular job, and knowing which one helps you combine them well.
Process mining
Process mining reconstructs workflows from system traces: timestamps from ERP, MES or ticketing, turned into a model with real cycle times. Where processes run entirely inside one system, it is the more precise method, because it measures.
It requires system traces, and it sees what happens inside the system. A log shows that an order sat for eight days. The reason is something you learn by talking to the people involved.
Generic chatbots and language models
"Write me a process documentation for order handling" produces a well-written text in thirty seconds. That text describes an average process from training data. The model learns about your customers, your systems and your exceptions once someone tells it about them. What makes the documentation useful for your company is therefore the gathering from the people who do the work.
Classic documentation tools and wikis
Confluence, SharePoint and similar tools give you structure, versioning, permissions and search, and they are a good choice for storing things. They are filled by people who take time to write. AI interviews complement them by gathering the content, which can then find its place in the wiki.
Consultant interviews
An experienced process consultant on site for two weeks works thoroughly and brings recommendations and experience from many companies. The result describes the company at the moment of capture. For strategic redesigns, consulting is a strong choice. For continuously recording how work happens in the company, AI interviews that keep writing with every conversation are a good fit.
Screen-recording tools for step-by-step guides
Screen-recording tools capture your screen while you work in an application. From the clicks they build a step-by-step guide with screenshots. For operating a single piece of software, that is a fast and precise way to produce click-through instructions.
Interviews are built for a different part of the job: workflows that run across people, departments and several systems. Why a step happens, what comes before it and who takes over when a colleague is out sick is something you learn in conversation. The two approaches fit together well. The interview documents the process, and a screen recording or a work instruction covers the click path of a single step.
What the Method Cannot Do
It does not replace professional sign-off. The AI documents what was said, with a source. Whether that is technically correct, legally sound and actually intended is still a human decision. For quality-management documents, the gathering is automated while release stays a step with a name and a date. As a rough guide, budget about an hour of reading per five or six processes.
It cannot find processes nobody talks about. The method is only as complete as the set of people interviewed. The snowball principle closes many gaps, because others name the people who were forgotten, though it cannot guarantee completeness. And what someone prefers to keep to themselves stays with them in an AI interview as well. That is part of voluntary participation, and it is worth it.
It does not measure cycle times. If three people say approval "usually takes two days," that goes into the documentation with a source, as a statement. Defensible numbers on duration, cost and utilisation come from measurement. The documentation prepares for it: once you know which processes exist and where they hand over, a process analysis can focus on the places that deserve it.
After Documentation: Optimize, Digitize, Automate
Solid process documentation is usually the starting point for further steps. What comes next is covered in four separate guides:
- Process optimization: five phases and the established methods from Lean and Kaizen to value stream mapping.
- AI process optimization: where AI noticeably improves workflows, how AI process analysis works, and why the documented current state comes first.
- Process digitization: removing media breaks, choosing the right processes, e-invoicing and GoBD in seven steps.
- Process automation: methods from workflow to RPA and AI, the business case, and why documented exceptions carry automation.
Data Privacy and the Works Council
wissa.ai is built for GDPR-compliant use: the platform is hosted in Germany, AI processing takes place in the EU, and a data processing agreement (DPA) under Article 28 GDPR is part of the offer; the contract documents are available in German. The interviews describe workflows, roles and systems. wissa does not rate people: there are no scores and no rankings. Because under German law it is enough that software is objectively capable of monitoring performance or behaviour, introducing it in a company with a works council usually requires codetermination under section 87(1) no. 6 BetrVG. We also recommend making participation voluntary.
Four questions typically come up:
- Could this turn into performance or behaviour monitoring? What gets evaluated are processes, systems and responsibilities. An evaluation of handling times or a comparison of who works faster is not part of it. Per person, the status of the invitation and conversation is visible, along with their statements.
- What happens to what people said? It is processed to produce the documentation and serves as the source record, which names the person. Interviewees can correct their own statements during the conversation, and anyone who wants to withdraw a statement can say so.
- Is participation voluntary? That is for your company to set, and voluntary participation also helps the quality, because people who enjoy telling their story tell it openly and precisely.
- Where does the data sit? Platform and database run on servers in Germany, with AI processing in the EU; all subprocessors, including delivery and email sending, are listed in the data processing agreement.
It makes sense to involve the works council right at the start, before the first interview. Tell the workforce in two sentences what this is about: securing knowledge that would otherwise leave with a resignation or a retirement.
What It Costs
Billing is in interview credits based on actual AI usage; roughly 10 credits correspond to one completed interview on average, including the analysis, with short conversations using fewer and long ones more. To get to know it, there is a free test package with 10 credits, usually enough for one interview (once per company, valid for one month, requires company details and a concluded DPA). After that:
| Plan | Price | Credits | Enough for |
|---|---|---|---|
| Test | €0, one-off | 10 | usually one interview |
| Starter | €79 per month | 100 per month | roughly 10 interviews |
| Business | €149 per month | 250 per month | roughly 25 interviews |
| Enterprise | on request | as needed | more capacity, payment on invoice |
| Credit top-up | €79 | 100, valid for 12 months | roughly 10 interviews |
Starter and Business can be cancelled monthly. Interviewees need no licence and no account of their own; the interviews use the credits of your account. All prices net of VAT.
For the food producer above that means 14 conversations, roughly 140 credits, more if conversations run long. One month on the Business plan with 250 credits would likely cover the capture, for €149 net.
Start for Free Today
wissa.ai is ready for you right after sign-up. You invite your team, the AI runs the conversations and writes the documentation from them. The page on process documentation software shows the features in detail, and the free test package is usually enough for a first interview.
Conclusion
In the first two generations of process documentation, the largest effort lay in gathering the knowledge. AI interviews take over that part: roughly 15 to 20 minutes per conversation, snowballing from one knowledge holder to the next. The result is step-by-step documentation with sources and a diff for every new version. Professional sign-off stays with you, and that is a good place for human time.
FAQ
What is AI process documentation?
An AI conducts structured interviews with employees and automatically turns them into process documentation. The AI asks, writes from the answers and extends the documentation with every further conversation. The gathering inside your own company creates the connection to your business: without interviews, a language model describes a general average process.
Can AI create step-by-step guides from interviews?
Yes, for processes. From the interviews the AI writes the steps in order, with the responsible role, the systems, the handovers and the exceptions for each step, and every statement has its source. Click-by-click guides for a single application come faster from a screen-recording tool or a work instruction, and both fit well alongside the process documentation.
How much time does AI process documentation take for the team?
Each interviewed employee spends roughly 15 to 20 minutes in a first conversation, by chat, and can pause and resume at any time; open points can go into a short follow-up session. The coordinator approves invitations and decides contradictions, usually in a few minutes each. As a rule of thumb, reading and checking take about an hour per five or six processes. Nobody writes minutes, because the AI produces the documentation.
How is this different from process mining?
Process mining works with system data from ERP or ticketing systems, shows what runs digitally there, and measures real cycle times. AI interviews capture the knowledge in people's heads, including the informal agreements, exceptions and that Word template from 2019 that never show up in any log. Process mining answers "how long," interviews answer "why" and "who else is involved."
How many employees do we have to interview?
Often fewer than expected: as a rough guide, in companies with 40 to 80 employees, 12 to 18 interviews cover the core processes, because the snowball principle suggests the people who are named in the conversations. Start with four or five people from different areas.
Is this compatible with data privacy and the works council?
wissa.ai brings the groundwork: a setup built for GDPR-compliant use, with hosting in Germany, AI processing in the EU, a data processing agreement under Article 28 GDPR, and no scores or rankings of people. What gets documented are workflows, roles and systems. In companies with a works council, introducing it usually requires codetermination (section 87(1) no. 6 BetrVG), so involve the works council and data protection from the start; details are best clarified with your data protection or legal advisers.
How do I know the documentation is accurate?
Every statement is traceable back to its interview round. If statements contradict each other, the system shows both versions side by side with their sources, and you decide which one counts. Professional sign-off stays with a human and mainly takes reading time.
What happens when a process changes?
You run a new conversation with the people involved, or a freshness check follows up after a while (both use credits). The documentation updates from it and shows a diff against the previous version, so you only read and check the change. That way the documentation keeps growing with your company.
Does this replace our quality management system?
It complements your quality management system. The gathering is automated, and document control remains your job: release with a name and date, version status, retention. What you get is a description of how work actually happens, including owners and interfaces, the lived as-is that a quality management system can build on well.
Sources
- Michael Polanyi: The Tacit Dimension, Garden City (NY) 1966 - tacit knowledge that cannot simply be written down
- Ikujiro Nonaka, Hirotaka Takeuchi: The Knowledge-Creating Company, New York 1995 - externalization: turning experience into explicit knowledge through dialogue
- ISO 9001:2015, clause 4.4 (Quality management system and its processes) and clause 7.5 (Documented information)
- General Data Protection Regulation (GDPR), Article 28 - processor
- German Works Constitution Act (BetrVG), section 87(1) no. 6 - codetermination on technical systems capable of monitoring behavior or performance