AI-assisted automation
Using AI sensibly for recurring tasks
Not every process needs AI. Most recurring tasks in businesses follow clear rules – and for clear rules, conventional automation is more reliable, cheaper and easier to verify.
Language models are useful where rules no longer help: with texts, documents and free text whose content has to be understood before anything can be done with it. Categorising requests, pulling details out of letters, summarising long documents, preparing drafts.
I use AI where it offers a real advantage over conventional rules – embedded in a controlled workflow, with structured inputs and outputs and with human review where it is needed.
Typical starting points
- Incoming emails or requests have to be read, categorised and forwarded to the right place.
- Specific details are to be extracted from contracts, letters or reports and transferred into a table.
- Large volumes of documents have to be sifted and pre-sorted before someone processes them in detail.
- Free-text fields – feedback, descriptions or notes – are to be evaluated and turned into fixed categories.
- Recurring texts such as summaries, draft replies or report sections are written by hand every time.
Who is this for?
Businesses and departments that process texts and documents on a regular basis: case handling, customer service, administration, purchasing, legal and contract departments, consultancies. You don't need an AI strategy or a data team – just a specific workflow that costs time today.
What can be automated
Extracting information from texts
Read names, amounts, deadlines, addresses or other details from letters, contracts or reports and store them in structured form.
Classifying documents
Categorise incoming documents or requests by type, topic or responsibility and forward them.
Summarising content
Condense long documents, minutes or correspondence to the essentials – according to guidelines that fit your needs.
Structuring free text
Turn unstructured input into fixed categories and fields so it can be evaluated.
Generating drafts
Prepare draft replies, report sections or text modules that a person reviews and approves.
Pre-structuring large document volumes
Sift, group and tag collections so the actual processing goes faster.
Rule-based or AI-assisted – how I decide
The first question is always: Can the task be described with unambiguous rules? Calculations, fixed business rules, data transfers between systems, deterministic checks and simple file operations belong in conventional automation. It delivers the same result every time, is traceable and costs almost nothing to run.
A language model comes into play when the content of a text has to be understood and the cases are too varied to be captured in rules. Even then the model remains just one component: conventional software fetches the data, passes it on in structured form, checks the answer and decides what happens with it.
The decision also covers which data the model may see, where it is run, what that costs in operation and how often an error is acceptable. Where a single error would be expensive, a person checks the result before it is used further.
When does it make sense?
Makes sense when …
- the task comes up often and today ties up people who read and categorise texts.
- the cases are too varied to be covered with fixed rules.
- the criteria can be described and the results can be verified.
- an occasional deviation can be caught by review steps.
Less suitable when …
- the task follows unambiguous rules – then conventional automation is more reliable.
- every single error would have serious consequences and no review is planned.
- the data situation does not allow processing by an external model and local operation does not justify the effort.
How I work
- 01
Collect the task and examples
We go through the workflow and collect real examples: typical cases, difficult cases, edge cases. This shows what a model has to deliver and where rules suffice.
- 02
Check feasibility
With the examples I test how reliably a model solves the task. You see results before you decide on an implementation.
- 03
Build the workflow
The solution is built as a combination of conventional software and a language model: data access, structured inputs and outputs, review steps, logging and the points at which a person decides.
- 04
Roll out and observe
After rollout the results are monitored for a while. You receive documentation, source code and a walkthrough – and know where the limits of the solution lie.
What you get
- An automation that uses language models only where they add value
- Structured inputs and outputs, review steps and logging of every run
- Clearly defined points at which a person checks or decides
- Source code, documentation, a walkthrough and an honest description of the limits
Limits
- Language models make mistakes, even with good instructions. A solution without review steps is unsuitable for most business processes.
- Which data a model may process has to be clarified in advance – depending on the type of data, the provider and the operating model. I don't make blanket promises about that.
- Operation causes running costs that depend on the model and the data volume. For small volumes that is negligible, for large ones it has to be planned.
Frequently asked
- Which processes are suitable for AI automation?
- Processes in which people regularly read, categorise or summarise texts or documents, or transfer details from them – and in which the cases are too varied for fixed rules. Examples are categorising incoming requests, extracting details from letters and contracts, or pre-sorting large document collections.
- When is conventional automation better?
- Whenever the task can be described with unambiguous rules: calculations, data transfers, checks against fixed criteria, file operations. Conventional automation delivers the same result every time, is traceable and much cheaper to run.
- Can documents be evaluated automatically?
- Yes, if it is clear which details are needed. A language model can extract content from contracts, reports or letters and turn it into a fixed structure. Scanned documents need text recognition first. Reliability depends on the quality and consistency of the documents and is tested in advance with real examples.
- How reliable are AI outputs?
- Good enough for many tasks, but never error-free. How high the hit rate is can only be measured on the specific case – which is why I test with real examples before every implementation. The solution is built so uncertainties become visible and errors can be caught instead of propagating unnoticed.
- Does a person have to check the results?
- In most business processes yes, at least at the points where an error would have consequences. How much control is needed depends on the process: for pre-sorting a sample is often enough, for details that feed into contracts or payments an approval is needed.
- Can AI and conventional automation be combined?
- That is the normal case. Conventional software fetches data, prepares it, passes it to the model in structured form, checks the answer and processes it further. The language model only takes over the part that rules cannot handle.
Next step
Which texts or documents cost your team time on a regular basis?
Describe the workflow and the type of documents. I tell you whether AI actually helps here – or whether a simpler solution is enough.