AI Engineer World's Fair 2026

You Can't Prompt the Room: The Last Skill AI Won't Replace

Balázs Horváth· Founder, VisualLabs15:45

Loading slides…

Read the talk

You Can't Prompt the Room: The Last Skill AI Won't Replace

As implementation gets cheaper, requirements discovery becomes more consequential: map the workflow, define the decision, and test whether people return to what you ship.

From a talk by Balázs Horváth

Choosing what deserves to be built

At an internal hackathon, Balázs Horváth’s team started with about 21 agent ideas and abandoned 17. Some lacked access to the necessary data; others did not make enough business sense to build. He reports that the four remaining ideas substantially changed how the team worked. The useful distinction was not which agent could be implemented, but which deserved to be implemented.

That distinction draws on what Horváth describes as thirteen years bridging business and IT. He began by testing functional designs and specifications, then writing them, worked on large ERP and CRM programs in the US and UK, and founded VisualLabs. His team learned to elicit requirements that could become specifications for developers, configuration work for consultants, and now instructions for AI. The implementation tools have changed much faster than the work of understanding customers. Reading the room is part of building valuable software.

0:250:48
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:25 · section reference included

Stakeholder access becomes the bottleneck

When access to code stops being the main constraint, another constraint becomes visible: getting stakeholders and decision makers into a room, securing enough of their time, and discovering what they actually need. A prompt can produce code or an entire specification. It cannot supply the missing conversation that determines whether that specification describes worthwhile work.

Horváth uses the familiar horse-to-car analogy: asking customers what they want may produce a request for more or faster horses, while understanding their underlying need may suggest a car. The saying is commonly attributed to Henry Ford, although The Henry Ford museum has not satisfactorily traced it to him. Here its purpose is to distinguish improving a familiar solution from discovering a different one. Horváth characterizes AI as tending toward common answers; that is his explanation for why human discovery must guide generation beyond reproducing existing approaches, rather than a technical definition of how every model behaves.

2:272:40
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

2:27 · section reference included

Map the workflow and slice the first release

This shifts attention toward the analyst’s toolkit: story mapping, the Business Model Canvas, value canvases, and design thinking. Story mapping is the technique Horváth singles out. Start with a backbone of user activities, then place the stories that support each activity beneath it. The backbone preserves the overall process while the details explain what users need at each step.

For a support system, the process runs from contacting support through triaging and resolving or routing, then closing the case. A first release can cover that entire journey with a small set of capabilities:

Workflow stageRelease 1 capability
ContactCapture intent
TriageClassify urgency
Resolve or routeDraft a grounded answer
CloseLog to a system of record

These are the first four stories of the MVP. Sentiment reading, team routing, next-action suggestions, chatting, and satisfaction checks can sit in the later backlog. Keeping the map at this level makes it possible to choose a release without losing sight of the complete support workflow.

Story map with Contact, Triage, Resolve or route, and Close columns. Release 1 includes capturing intent, classifying urgency, drafting a grounded answer, and logging to a system of record. A Later row lists additional capabilities.Open full source frame
A support story map separates Release 1 from later capabilities.

The map does not finish requirements discovery. Each story becomes a starting point for discussion with stakeholders: what should this capability actually do, and what would make it useful in their process? That conversation turns a short label such as “classify urgency” into a requirement an agent can act on.

4:044:29
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:04 · section reference included

Turn stories into specifications and tests

Consider the urgency story: a support lead needs open cases ranked by urgency so that escalations do not slip. Its structure carries three distinct pieces of context:

  • Persona: the support lead who needs the capability.
  • Need: a ranking of open cases by urgency.
  • Purpose: preventing missed escalations.

The purpose matters because it gives the requested behavior a reason. Horváth argues that this familiar user-story structure also helps AI interpret requirements through pattern recognition; he presents that as practical guidance, without a measured model comparison.

Add acceptance criteria to make the story testable, then derive test cases from those criteria. For the support-lead story, the stakeholder conversation still has to establish what urgency means and what would count as a missed escalation; the sentence alone does not settle those rules. Once the stories are connected, they can describe a coherent system from which to create the specification and code. The development lifecycle remains recognizable: requirements, acceptance conditions, tests, and implementation still need to agree, even when AI changes the tools used to produce them.

6:266:40
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:26 · section reference included

Four questions that make requirements concrete

Before deciding what to build, work through four discovery questions:

  1. Whose problem is this? Name a specific person or persona and make the problem concrete enough to quantify.
  2. What does winning look like for them? Identify the successful outcome, including whether reaching it faster, more smoothly, or more safely matters.
  3. What would make them refuse to use it? Look for an unsupported platform, cumbersome interaction, or data-security concerns that would block adoption.
  4. Would it change a decision? Name the decision and explain how the system would help the person make it better.

The final question connects the proposed capability to an action in the user’s work, rather than treating the existence of a new feature as success.

Keep the answers in a Markdown file in the repository so AI has access to the same context. For the support-lead example, a working requirements.md can retain the established story and make the remaining discovery questions explicit:

markdown

# Rank open support cases

## User story
As a support lead, I need open cases ranked by urgency
so that no escalations slip.

## Discovery questions to resolve
- Whose problem: Which support lead owns this workflow,
  and how often are escalations currently missed?
- Winning: What observable outcome means escalations no longer slip?
- Refusal: Which platform, usability, or data-security constraints
  would prevent use?
- Decision: Which case-handling decision should the ranking change?

## Acceptance criteria to agree
- Define urgency with the support lead.
- Establish examples of the expected ordering of open cases.
- Derive test cases from the agreed examples.

The unanswered items remain questions, not silently invented business rules. This gives the next discussion a concrete starting point and is more useful context than a generic instruction to build an agent that handles support.

7:568:11
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:56 · section reference included

Value → Architecture → Design

Requirements discovery also needs an order of operations. First establish what constitutes value and how the current process creates it. Then understand the architecture that supports that process. Only then begin designing the system. Horváth calls this VisualLabs’ VAD: Value → Architecture → Design:

  1. Value: determine what the customer is trying to gain and how that value is created.
  2. Architecture: understand the process flow and the systems supporting it.
  3. Design: shape the system around the desired value and process, including any process changes needed along the way.

Starting with value prevents a convenient implementation from becoming the goal before the team understands the work it is supposed to improve.

Much of this is familiar product management. Its significance changes when teams can reach for the same implementation tools and coding models. Under those conditions, the differentiator is how well a team understands the business need before asking a model to build. The old skill becomes more consequential as the economics of implementation change.

9:359:59
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:35 · section reference included

When shipping speed hides a product problem

High development velocity can coexist with weak adoption. A team may ship features rapidly while few people use them. Even an initial login or trial of a new feature does not establish that it has lasting value. Horváth recommends looking at the frequency of a relevant activity, rather than relying on time spent on the site: do people come back to perform the work the system is intended to support?

Two further warning signs concern delivery itself. A polished demo is easy to mistake for a finished product, but a demo is not a live system that people actually use. Likewise, a product requirements document, or PRD, that has never been tested with real users lacks the feedback needed to establish whether the proposed product belongs in their work. Production use and real-user feedback are part of the path to delivery, not consequences to assume from a convincing presentation.

Four cards read “Velocity up, adoption flat,” “Never used a second time,” “The demo is the deliverable,” and “A PRD no real user tested.”Open full source frame
Four warning signs of building the wrong thing fast.
11:0511:25
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:05 · section reference included

Move expertise upstream and measure repeat use

If implementation becomes cheaper, choosing the work deserves more attention from experienced people. Horváth’s organizational proposal is to move expertise upstream: closer to customers, business problems, and the decisions about what gets built. Begin by auditing which measurements and time allocations currently reward building the wrong thing.

Horváth proposes replacing the count of features shipped last quarter with the count of shipped features actually used more than twice. This is a practical suggestion for shifting attention toward repeat use, not a validated universal threshold; he does not specify a user denominator or observation window. Its purpose is to change what the team pays attention to after shipping.

The corresponding staffing change is to give subject-matter experts more influence over customer-facing decisions. That does not require every engineer or expert to become a functional consultant, product manager, or product owner. It requires including their knowledge of what has worked and failed before while the team is still deciding what to build.

12:3112:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:31 · section reference included

Run a mapping session before building

Before the next implementation, run a mapping session. Use a user story map, a Business Model Canvas, or another map that makes value creation visible. The artifact matters less than whether the team can explain where the value lives and how the proposed system contributes to it.

Three cards under “Monday morning”: “Audit the wrong-thing rate,” “Move one senior person upstream,” and “Run one mapping session.”Open full source frame
Monday morning: audit, move upstream, and run a mapping session.

Then try the same AI-assisted use case with and without user stories and compare the results. Horváth reports that doing this himself persuaded him to keep user stories in his development process; he offers a personal observation, not a quantified controlled evaluation. The next useful experiment is available before any elaborate process change: take the thing you already want to build, articulate the users’ needs and purposes, and see how that changes the result.

14:3014:52
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:30 · section reference included

Resources

From the talk

Transcript download
Full transcript

Loading transcript…