Student research project · not a product
Two Sides
of the Same Call
One tax advisor who works alone. Her clients. One call at 12:10.
Walk through it from both ends — eleven stations, about five minutes.
The little boxes show where a claim comes from: I interview R research A my assumption V vendor claim. Click one anywhere.
Station 1 · the call
You are the advisor.
You’re deep in a difficult annual statement. Keep the file number in your head. That’s your focus.
A game, not data.
Station 2 · two sides
Same call. Two sides.
Meet her clients
Proto-personas from one advisor interview — assumptions, not data. I haven’t talked to clients yet.
Station 3 · the price tag
Every call has a price tag.
Up to 10 calls on a normal day, up to 20 around Christmas. 5 to 20 minutes each, then 5 to 10 minutes to refocus. Her rate: €150 an hour.
per year, routine calls only
€0
These are costs, not savings — she still has to answer in writing. And my guesses differ ×5. So the first job is a one-week call log.
Bonus drain: some clients call to get a short tax question answered without a bill. She then often writes a follow-up email to document and bill it — double work.
Station 4 · the second drain
It runs both ways.
Every year she sends her clients a checklist of documents. Some points stay open — as an example, she mentioned 7 of 10 points answered. She chases the rest by email or phone, again and again. At the end of the interview she brought it up again, unprompted.
reminders sent: 0
One example, not an average. The list items are my illustration A. New hypothesis N1: chasing interrupts her as often as incoming calls — open, the call log will tell.
Station 5 · the waiting game
Now you’re the client.
You wait for your tax return — twice, one after the other. Notice how each wait feels.
Wait A
- …
Wait B
Status lines are an illustration A, neutral on purpose. Nothing is stored.
Station 6 · wrong guesses
An AI should answer prevent the call.
The interview changed my thesis. Bet on what it showed:
Station 7 · the market
Find one that prevents the status call.
Tap every option. Vendor claims, Nov 2026, not verified V
Station 8 · the rules
Who’s calling?
You’re the phone assistant now. Three people want information.
She may not even confirm that someone is her client. Third parties have tried. (§ 203 StGB, § 57 StBerG — not legal advice.)
Say it’s an AI
An assistant must disclose it is an AI (AI Act, Art. 50). But disclosure lowered acceptance in sales calls (Luo et al., 2019).
A call never made avoids this tension.
Who holds the phone?
Later, an assistant could answer status questions itself — after an identity check. Voices won’t do: humans cannot reliably detect speech deepfakes (Mai et al., 2023).
Concept, not built.
Station 9 · the response
Three gates. Drop a caller in.
Pick a client — and watch where the call ends up.
And beyond her office: the chamber could offer model secrecy clauses for AI tools (§ 62a StBerG); practice software could make status updates without a login standard.
Station 10 · the test
The bars are set. The data isn’t in yet.
I fixed each bar before collecting data, so I can’t move it later. Each bar decides what happens next. Try your luck:
Station 11 · the end
One interview is a signal,
not a pattern.
- One tax advisor. I know her personally, so politeness may have biased her answers.
- No clients interviewed yet. The personas are my assumptions.
- Estimated numbers. The call log replaces them.
- Unchecked vendors. Everything marked V is the providers’ own word.
Sources, AI use, how this was made
I used Claude (Anthropic) to search for and summarise literature, to draft the competitor analysis, to translate and summarise the interview transcript and to correct language in parts. I conducted the interview myself; all findings, interpretations and decisions are my own. This site was coded with Claude Code (Anthropic) from my brief. No cookies, no trackers, no third-party requests.
Individual Assignment, Part 2 · BUU44712 Digital Entrepreneurship & AI · Trinity Business School · November 2026