AI Workflow Design · Process Architecture · Facilitation

Designing the AI layer,
not just using it.

Most teams add AI tools to an existing process. This project — an internal exploratory initiative within IKEA's digital experience design team — asked a different question: what does it look like to architect AI into a design workflow from first principles, with explicit checkpoints, structured agent roles, and a human decision layer that genuinely controls the output?

IKEA · Current Garaje de Ideas AI Workflow Design
Client IKEA / Ingka Group
Agency Garaje de Ideas / Groupe EDG
My role Design Operations Lead
Period 2025 — Present
Scope AI Workflow Design · Process Architecture · Multi-agent Orchestration · Facilitation
The problem

The wrong question was everywhere

The dominant question in design teams exploring AI was "which tool should we use?" The answer was almost always a product recommendation — a specific assistant, a specific plugin, a specific integration. Teams would adopt it, use it for a few weeks, and find that the outputs were inconsistent, the process was murkier than before, and nobody could explain why a particular output was good or bad.

The problem wasn't the tool. It was that AI had been added to a process that wasn't designed to accommodate it. Without explicit phases, defined agent roles, human checkpoints, and context management, AI-assisted work produces something more dangerous than failure: it produces plausible-looking output that drifts from intent without anyone noticing.

The insight

AI is a process layer, not a feature

The reframe that made this project possible: treat AI the same way you'd treat any other design systems problem. Define the inputs. Define the outputs. Define the handoffs. Define who decides what and when.

That reframe shifts the question from "what can AI do?" to "where does AI fit in this process, and how do we verify that it's working correctly?" The second question is harder — and far more useful. It's also the kind of question that systems thinkers are better positioned to answer than most, because it requires understanding the whole before optimising any part.

The framework

Three tiers. Five phases. Human gates throughout.

The framework organises AI-assisted design work into a three-tier agent hierarchy operating across five defined phases. Each tier has explicit scope, defined inputs and outputs, and clear escalation paths to human decision-makers.

Tier 1 Tier 2 Tier 3 Orchestrator Manages context, coordinates phases, escalates to human decision layer Phase Agent Owns a single phase — Discovery, Shaping, Building… Phase Agent Owns a single phase — Betting, Retrospective… Task Agent Problem decomposition Task Agent Constraint mapping Task Agent Solution evaluation Task Agent Decision logging Scope boundaries are explicit and enforced. No tier can substitute for human decision authority.
Three-tier agent hierarchy. Each tier has defined scope and escalation paths — the Orchestrator coordinates but does not decide; humans hold final authority at every phase gate.
Applied

Two threads from the same structure

The framework wasn't built in the abstract. It emerged from two parallel workstreams inside an enterprise design team, both of which exposed the same underlying problem: AI without process produces work nobody can own.

Shape Up facilitation

Shape Up shaping sessions are cognitively intensive — multiple stakeholders, ambiguous scope, competing constraints. I designed facilitation workflows that use AI to support the shaping process: structuring the problem before the session begins, surfacing edge cases before pitches are written, and synthesising multi-participant outputs into pitch-ready documentation. The AI handles structure; the facilitator handles judgment.

Multi-agent orchestration for complex features

For features involving product, engineering, content, and business stakeholders simultaneously, a single-agent approach produces generic outputs that satisfy nobody. The multi-agent approach distributes the shaping task: one agent decomposes the problem, another maps constraints, another evaluates solution directions. Outputs are structured to flow into existing documentation formats — so the AI layer and the process layer are the same system.

01 Discovery Problem statement · Appetite · Success criteria Human confirms: problem framing + appetite ⬡ GATE Problem confirmed 02 Shaping User needs · Constraints · Solution directions Human selects: direction + approves pitch ⬡ GATE Direction selected 03 Betting Pitch evaluation · Risk assessment · Commitment Human decides: what gets built this cycle ⬡ GATE Scope committed 04 Building Scope management · Decision logging · Progress Human manages: scope, deviation decisions ⬡ GATE Work completed 05 Retrospective Learnings synthesis · Framework evolution Human leads: what changes for the next cycle ⬡ GATE Framework updated Gates are hard stops, not formalities. The workflow does not advance without human sign-off.
Five-phase workflow with human gate checkpoints. Each gate requires a named human decision — not an acknowledgement. AI advances only after human sign-off.
What this requires

AI raises the bar for human judgment — it doesn't lower it

The most common misconception about AI-augmented workflows is that they simplify the designer's role. They don't. The role shifts from producing to directing, evaluating, and integrating. That shift requires skills that aren't traditionally part of design education:

  • Prompt craft — Setting direction with enough specificity to guide useful output, enough openness to allow genuine AI contribution. Closer to brief-writing than specification.
  • Output evaluation — Assessing AI output critically and quickly: what's right, what's wrong, what needs investigation. Requires domain knowledge and a clear model of what the work is trying to achieve.
  • Context architecture — Deciding what information must carry forward across phases, how it should be compressed, and what can be safely left behind. A new kind of information design skill.
  • Drift detection — Noticing when the AI's framing has shifted subtly from original intent. Requires keeping the original intent visible and referring back to it constantly.
  • Calibrated trust — Knowing when AI output is likely reliable and when it requires deeper scrutiny. High-confidence AI output is not a signal of correctness — it is a trigger for more careful review.

These are skills that senior designers, systems thinkers, and design leads are best positioned to develop — because they already understand process at a level that makes AI augmentation legible rather than magical.

Review efficiently High-confidence output types Structural organisation of information explicitly provided Summarisation of long documents (original available for verification) Formatting and presentation of human-confirmed decisions Identification of gaps in provided inputs Scrutinise carefully Outputs requiring deeper human review User needs synthesised from indirect inputs or training data Risk identification — surfaces known patterns, not necessarily actual present risks Recommendations aligning with apparent project direction Quantitative claims of any kind High-confidence output — fluency ≠ correctness Validate externally Do not use without external validation Any output referencing external facts, data, or precedents Predictions about user behaviour Competitive or market analysis
Trust calibration tiers. Calibrated trust — knowing what to review versus scrutinise versus validate externally — is a skill that develops with practice and should be made explicit across the team.
Outcome

A framework that separates process from tooling

The most durable result of this work is a process architecture that is independent of any specific AI tool or platform. Because the framework defines phases, agent roles, checkpoints, and context management protocols at the process level — not the implementation level — it survives tool migrations, platform changes, and team turnover.

That separation also makes it teachable. The framework is now being shared through an active AI literacy programme for senior design professionals, with a focus on workflow design rather than tool use. The target audience is designers and design leads who already think in systems — and who are ready to apply that thinking to how AI fits into the process, not just what AI can do.

"The AI is a layer. The process is the design."

The framework is documented in full as a standalone piece — written to be adopted, adapted, and evolved by any design team working seriously with AI. It draws from the real-world constraints and failure modes encountered during this engagement and represents a position: that responsible AI augmentation in design is a systems design problem, and should be treated as one.

Read the full framework — 18 min