Design in the Age of AI: The Human Layer

At OSS Ventures, the most awesome SaaS venture builder specialized in manufacturing in the world (we’re the only one), we prided ourselves during the first four years in the central role of design. We were going to factories, interviewing users, inferring cool SaaS ideas and creating pristine figma UX designs.

Then AI came.

1 · The Paradox

For a brief moment, we thought design might matter less.

AI could already generate full mock-ups, UX flows, and pitch decks. Prompts replaced wireframes. Figma plug-ins spat out whole apps. Some guy invented “AI interviewing”, where you basically sit down and vibe interview chatGPT roleplaying as your future user.

The myth was seductive: that maybe, finally, design would become an afterthought — a thin varnish on top of a self-building world.

It lasted three months.

Because once we started building real AI products, the opposite became true.

The surface became easy — the understanding became hard.
 Design turned from a craft of pixels into a discipline of cognition.

2 · Speed Changed the Game

The first shock was tempo.
 What used to take weeks now takes hours.
 A designer can jump from a conversation to a functional prototype before lunch. That part feels like magic. Recently, we went from idea to mockup in 12 hours with real client data.

But the compression of time creates a void somewhere else.
 When mock-ups are instantaneous, the scarce resource becomes insight.
 It’s no longer about how fast you can make something look good — it’s about how precisely you can capture how people think.

Turns out for now, we haven’t been able to uncover unconventional insights from vibe interviewing chatGPT.

Designers today spend less time on Figma, more time decoding decision logic, shadowing planners, or asking what data they actually trust.
 The job moved from pixels to patterns.

3 · Rethinking “As-Is” and “To-Be”

At OSS Ventures we always mapped customer journeys.
 In the age of AI, those maps broke.

Because “as-is” and “to-be” are no longer linear flows of clicks; they’re systems of data, decisions, and algorithms — some digital, some human.

So we changed our process.
 Every journey now includes two extra layers:

1. The system layer — what tools, data streams, and databases people use while doing the job.
2. The mental-algorithm layer — how they decide, what shortcuts they take, what heuristics they trust.

We discovered that understanding a workflow now means understanding both the SQL and the synapse.

4 · The Hidden Algorithms in People’s Heads

One project made the lesson brutally clear.

We built an AI planning tool for a manufacturer.
 The AI produced schedules that were mathematically optimal — and instantly rejected by every planner.

“It’s wrong,” they said.
 Except it wasn’t.
 It just wasn’t what they would have done.

So we unpacked it. Over weeks of interviews and shadowing, we uncovered roughly 550 micro-rules living in their heads:
 “If supplier A is late and machine B is hot, prioritize line 3 before line 2.”
 “If it rains on Tuesday, shift maintenance to Thursday because Jean is off.”

None of this existed in any system. Yet it defined the truth of operations.

The insight was humbling: before you can align machines and data, you have to align the human algorithms that run the company.
 
Design became the process of reverse-engineering intuition.

5 · Designing for Human Algorithms

Once we saw it, we couldn’t unsee it.
 Every role — buyer, planner, maintenance manager — has its own invisible model of reality.
 They’re not wrong or right; they’re localized truths.

In deterministic software, you standardize and enforce one truth.
 In probabilistic software, you expose, compare, and orchestrate many. And give the keys to a final human boss who’s in charge.

That’s now design’s job: to make those mental models legible.
 To build UIs that don’t dictate decisions but show why people differ.
 To help humans debug themselves and take the probabilistic view when it matters, and the deterministic one elsewhere.

6 · From User to Manager of AI Agents

A new archetype appeared this year: the AI manager.

These users don’t click buttons; they supervise fleets of autonomous agents and deterministic pieces of software, making decisions on their behalf.
 They need control, not features.

The interfaces we designed started to look eerily like IDEs:
 logs, error handling, variable inspection, version control of prompts.
 We weren’t designing dashboards anymore — we were designing cockpit instruments.

It’s a subtle shift with massive implications: UX moves from action to supervision.
 The product becomes a negotiation between automation and human judgment.
 Design becomes about giving confidence, not control.

7 · The Invisibility Challenge

Now, some of our AI systems have no interface at all. The recommendations appear directly inside legacy tools.

There, “UX” happens in the background.
 The only design that remains is the exception layer: how people understand, override, and trust the invisible.

We learned to design invisibility.
 To craft the boundary between automation and awareness.
 In these systems, the only visible UX is for the power users — the ones who configure, monitor, and debug.
 Everyone else just feels that “things work better.”

That’s still design. Maybe the purest form of it.

8 · Designing for Organizations, Not Screens

AI rarely changes a single job; it reshapes the entire workflow.
 Which means you can’t stop at personas — you need power maps.

In every deployment we now chart:

• Who gains or loses decision authority,
• Who provides data vs. consumes it,
• Where trust gaps appear once algorithms enter the room.

The design deliverable is no longer a wireframe; it’s an organizational diagram.
 A visualization of how intelligence — human and machine — flows through the system.

We started calling this systemic design: the craft of shaping collaboration between humans, AIs, and institutions.

9 · The New Material of Design

For most of design history, the material was visual: pixels, typography, motion.
 Then it became structural: data models, information architecture.
 Now the material is probability.

You’re designing around uncertainty.
 Every output of an LLM or an ML model is a distribution, not a fact.
 So the designer’s palette includes confidence intervals, failure modes, and human fallback.

In deterministic systems, you could aim for consistency.
 In probabilistic ones, you design for variance.
 Manufacturing solved that a century ago with tolerance bands and process control.
 Design is rediscovering the same thing — except our “parts” are text and recommendations.

10 · Tooling and Workflow

Ironically, AI made Figma obsolete for parts of our work — and indispensable for others. Our designers are now spouting a wireframe in half a day. That’s no longer where the edge is.

Fast mock-ups mean designers can test structure instantly, freeing time for field work. We spend more hours in factories, control rooms, and Teams calls. The best insights come from watching someone curse at their screen.

11 · Examples from the Trenches

The Planner Rulebook
 550 rules uncovered. The design output?
 A “bias editor” where planners could visualize, adjust, and share their personal heuristics. The tool learned from them before advising to them.
Design’s role : do the work to get the 550 rules.

The Invisible Assistant
 A maintenance tool where the AI ingested incredible amount of data in the backed. Users never saw the algorithm — only the 2AM maintenance guy in Mexico is now behaving as their best expert based in Dallas.
 Design’s role: create transparency reports for managers, not UIs for operators.

The AI IDE
 A AI for quote product where power users tune agent behaviors.
 We designed trust, band of probabilities, — an IDE for quoting, really.
Design’s role : create the new interface for supervision, not delivery

12 · The Questions We Still Don’t Have Answers For

We’re still early. A few open fronts keep us up at night:

• What’s the optimal UX for an invisible system?
• How do you represent uncertainty without destroying trust?
• Should AI systems have a visible “personality,” or be neutral infrastructure?
• How do you document and version “prompt design” the way developers version code?
• At what point does designing an organization become designing ethics?

These aren’t philosophical questions — they’re daily ones.
 Each new deployment gives a slightly different answer, and that’s fine.
 Probabilistic worlds don’t reward certainty. We’re all busy learning.

13 · Why This Matters for Builders and Investors

For builders: AI shortens the distance between idea and product, but lengthens the distance between product and adoption.
 Design bridges that gap.
 Invest in design early, not late — because every wrong assumption compounds exponentially in self-learning systems.

For investors: the moat in AI products isn’t the model; it’s the human understanding embedded in the workflow.
 The best products we’ve seen weren’t the smartest — they were the most aligned with how people actually reason under pressure.

For operators: design is your mirror.
 If the AI feels off, it probably misunderstood you, not your data. Contact your local, friendly designer.

14 · A Historical Echo

In the 1990s, software teams discovered usability.
 In the 2020s, design became strategy.
 In the 2030s, design will be infrastructure — the connective tissue between humans and autonomous systems.

Factories once learned to manage mechanical variability with process control and Six Sigma.
 We’re now learning to manage cognitive variability with design.

It’s the same story, one layer higher in abstraction.

15 · The Future Designer

The new designer profile is strange.
 Part psychologist, part data-analyst, part storyteller.
 Able to read SQL schemas and body language in the same meeting.
 Fluent in prompts, ethics, and production KPIs.

They’ll work closer to product managers and engineers than ever — because in probabilistic systems, everyone is designing.

And they’ll spend more time outside the office: on the field, watching the gap between what people say and what they do.
 Because that’s where the real product lives.