Presented at SDLCAI 2026, a seminar day on AI in software development, 13 Oct 2026, Marsio, Espoo, in the session “Views from academia” · Tapio Pitkäranta, PhD candidate, Aalto University
An AI agent builds it from the stakeholder’s own words, in minutes.
The stakeholder uses it as if it were real, instead of reading about it.
It answers a question about the problem, then it is thrown away.
of project effort goes to requirements and design
Boehm 1981; COCOMO IIthe cost of a defect found after delivery, vs. one found in requirements
Boehm & Basili 2001of effort spent on rework
Boehm & Basili 2001Discovers: does a solution fit?
Working software the stakeholder can use. In the field: meeting-room booking, a cash-flow forecast.
Field · 69 probesDiscovers: the process, its constraints
Today’s process re-enacted with real timings, which the stakeholder can stop, correct or fork.
Illustrated · 3 casesDiscovers: which assumptions hold
Today and a possible future side by side, so assumptions can be judged before anything is built.
Illustrated · what-ifsDiscovers: what the question needs
UI, a chat overlay, a value-chain diagram, voice narration: chosen for the stakeholder.
OpenA domain expert who can describe the need in plain language. Writes no code.
Turns ideas into tractable prompts, checks safety, protects the agent queue.
A coding agent (Claude Code) reads the queue, builds the Probe, deploys and reports.
New processes are tried here, and probes are created at industrial speed.
APIs, connectors and modernisation bridges. Validated probes move inward only through deliberate, reviewed steps.
The stable foundation. The outer layers extend it; they do not replace it.
Every requested probe was built and shown back live, in the room.
Queue management, prompt hygiene and session tracking all fell back onto one person.
A small tool batched briefs and dispatched agents in parallel. Production stopped limiting.
Any question a stakeholder can step through, and judge, can become a probe.
The assembly line is now an agent fleet; the factory manager writes prompts, reviews pull requests and validates output.
Eight engagements, two voluntary organisations. May not transfer where people cannot describe pains in conversation.
Probes validate fit, not production quality. The multiplier must not be projected onto delivery.
Every probe came from Claude Code. Patterns may or may not be vendor-independent.
50 lines per day varies by domain and maturity. It moves the number, not the qualitative finding.
The authors often orchestrated. An independent or less skilled one might get different results.
Field scale covers the solution-proposing mode. The healthcare cases are illustrative, n = 3.
The role that turns sentences into probes. It keeps prompts tractable and the agent queue protected, and it is the first thing to automate as you scale.
Give probes a safe place to run, away from the legacy core, and a reviewed path for the ones that hold up to move inward.
Count questions asked and probes judged, not lines of code. Idea supply and validation speed are where the next constraint lives.