AI as a controlled input aid for option lists

Use AI for option lists to generate proposals, not as an automatic publish button. Have the system structure text from a source file, validate the result technically and have a content owner check every question, price and condition before customers see the list.

Manually retyping a Word document containing dozens of options takes time and invites copying errors. AI can speed up the initial conversion, but does not automatically understand the contractual meaning, dependency or exception your organisation intends.

Limit the task to recognisable structure

Yindle Portal can extract text from a DOCX and generate a proposed questionnaire. The generator converts options into questions with positive and negative choices, amounts in cents, omitted work as a negative amount and unknown prices as blank. A schema limits quantities and text lengths.

That is useful input assistance, but no guarantee that the source is unambiguous. A line such as “price on request, only with package B” needs human interpretation and possibly a different form structure.

Set a fixed review order

First check completeness: does the proposal contain every relevant source line? Then check meaning: is the yes/no question correct and has omitted work been interpreted properly? Next assess price, VAT context, quantities, dependencies and language. Adjust sorting and presentation only afterwards.

Never publish directly from the model response. Retain a draft status and have a second reader compare a sample against the source file. A large financial impact may require a complete second check.

Worked example: fictional home renovation

In this fictional example, contractor Dijkzicht has a document with 48 renovation options. AI converts “Additional skylight, €1,250 including VAT” into a yes/no question. For “Omit window sill, −€90”, the proposal creates a negative price. A line about a technical condition is included in the description but not yet implemented as blocking logic.

The work planner compares every amount with the source, flags four unclear lines and splits one combined option. A colleague then checks the final customer summary. This fictional example shows where automation stops and editing begins.

Use a source-to-question log

Source line Proposed question Check Decision
Additional skylight Add skylight? Price and description are correct Accept
Package-dependent Add option? Condition cannot be enforced Model manually
Price on request Yes, add — to be determined Follow-up process missing Add a contact route

This log lets you trace errors and improve the prompt or input process. Do not publish the internal log to customers; they need a clear, approved list.

Test changes and confidentiality

Use only source files your agreements permit the chosen AI service to process. Remove personal data and irrelevant confidential information. Record who may start generation and how long source and result are retained.

Also test repeated imports of a changed list. The system must not silently overwrite existing customer answers. Compare versions, publish deliberately and determine whether current drafts need reviewing again.

Create a separate error set for evaluation. Include amounts with dots and commas, several options in one paragraph, negative amounts, tables, footnotes and a line without a price. Repeat the same set after a model or prompt change. This assesses fixed difficult cases rather than a handful of successful examples.

Distinguish content rejection from technical rejection. Technical rejection includes invalid JSON, too many options or an amount out of range. Content rejection includes a misinterpreted condition or unclear customer wording. Only the first category can be fully caught by schema validation.

Track human corrections without making confidential source content widely available. If “per property” is often read as a quantity when it means a fixed price, you can improve the instruction specifically. Then measure whether that type of error decreases. Speed becomes a gain only when the checking workload also falls.

Plan an alternative for a document that cannot be processed reliably. Administrators must be able to start manually or use only the useful lines. Do not force anyone to repair a poor proposal because the automatic route has already started.

Also note why the automatic route was abandoned, so a similar source file can be assessed faster and more consistently later.

Read how choices fit within a customer portal and explore Yindle Portal. To convert one source list responsibly, discuss the review route with Yindle.

Previous insight
Next insight

Start with your question

Where does your team get stuck?

Your first enquiry does not need to be a technical brief. Tell us which systems you use and where work gets stuck. Together we discuss what is possible and which step you can take next.

Discuss my project