The criteria we apply, what we look for, and the principles that govern which agencies appear on this site.
Every agency has been assessed against five criteria. No numeric scores are assigned — the criteria determine whether an agency meets the threshold for inclusion and how it is positioned in the directory.
The central challenge in evaluating AI UX agencies is distinguishing agencies that have genuinely solved AI-specific design problems from agencies that have applied standard UX practice to AI-adjacent products and relabeled it. The criteria below are specifically designed to make that distinction.
The primary criterion: whether the agency has shipped AI products where machine learning model outputs directly affect user decisions in production. Not prototypes. Not products that contain AI features in the background. Not AI tools used during the design process. Products where the UX challenge was specifically and primarily about how users interact with AI-generated outputs — generative content, probabilistic recommendations, AI agent actions, or ML-powered decisions.
We assess this through live deployed products currently in use by credited clients. We check whether the AI features described in case studies are still active in the live product, and whether the design decisions attributed to the agency are visible in the current experience. Case studies that describe AI product work without showing the specific AI interaction patterns — uncertainty states, confidence indicators, output refinement flows — are given less weight regardless of how polished the screens look.
Whether the agency has developed documented approaches to the design problems unique to AI products. Standard UX methodology was not built for probabilistic systems, and the agencies that have done the most AI product work have developed new frameworks for the problems that standard UX doesn't address.
We look for evidence of methodology in published thinking, case study detail, and process documentation: how the agency approaches uncertainty communication, how it designs fallback states for AI failure modes, how it handles the trust design challenge of systems with variable outputs, and how it tests AI UX with real model behavior rather than simulated data. Agencies that can describe these approaches specifically have developed genuine methodology. Agencies that describe their process in standard UX terms without AI-specific adaptation have not.
Whether the agency's designers understand AI system behavior well enough to design for it accurately. An interface designed around how an LLM ideally behaves will fail in production because LLMs don't behave ideally. Agencies with genuine technical fluency prototype with real model integrations, design for edge cases and failure modes from the start, and can describe how their design decisions account for specific model behaviors.
We assess this through case study specificity — whether the work described reflects understanding of how the underlying system works — and through evidence of prototyping with real model behavior rather than static mockups that simulate AI outputs.
Whether the agency's demonstrated AI product portfolio matches the types of clients and AI products it claims to serve. Consumer generative AI products and enterprise AI dashboards are different design problems requiring different expertise. An agency with deep generative AI UX experience may be genuinely wrong for an enterprise ML operations brief, regardless of overall portfolio quality.
Each agency is assessed against the AI product types it has documented experience with. Stage fit is also evaluated: startup-accessible agencies with appropriate pricing and process models are distinguished from enterprise-tier agencies whose model is not suited to early-stage AI companies.
Verified reviews from named clients on identity-checked platforms specifically referencing AI product work, industry recognition for specific AI product designs, and documented business outcomes tied to AI UX decisions. General UX validation is noted but given less weight than AI-specific validation.
We distinguish meaningful AI product validation from noise: a review that specifically describes the AI UX challenge solved and the outcome achieved carries more weight than a general satisfaction review for design quality. Awards for specific AI product work — not general design industry recognition — are the relevant signal here.
Agencies that use Midjourney, Figma AI, or other AI tools to accelerate design work are not qualified for this directory on that basis. The criterion is designing UX for AI products, not using AI tools during design.
An agency that designed a marketing website for an AI company has not demonstrated AI product UX experience. The relevant work is designing the product interface itself.
Every agency now claims AI expertise. We evaluate only what the portfolio and client evidence demonstrate — not what the services page or homepage states.
Both are noted in profiles as context. Neither is a selection criterion.
An agency can produce excellent design work and still lack the specific AI product UX experience this directory requires. General portfolio quality is assessed under criterion 1 only as it relates to AI product work.
Inclusion requires at least one verifiable shipped AI product where model outputs directly affect the user experience, documented evidence of AI-specific UX methodology, and sufficient public information to evaluate across all five criteria.
The directory is reviewed once per year. Individual profiles are updated on a rolling basis when significant new AI product work, team changes, or validation signals emerge. Factual corrections and profile updates can be submitted via the Contact page with supporting documentation.
Agencies are identified through our own research process. If an agency is not listed, it either did not meet the inclusion threshold, lacked sufficient verifiable AI product portfolio work, or has not yet been reviewed.
We check live deployed products for the AI features described in case studies. We look for the specific interaction patterns that AI products require — uncertainty states, confidence indicators, output refinement flows, fallback states — in the live experience. We assess whether the case study describes the AI-specific design decisions made, not just the visual output. Agencies that can only show static screens without describing the AI interaction architecture have not demonstrated AI product UX depth.
AI product UX means designing the user experience of the AI system itself — how users interact with model outputs, how uncertainty is communicated, how trust is built when outputs vary. Working with an AI company as a client is not equivalent unless the specific work involved designing that interaction. An agency that designed an AI company's brand identity, marketing site, or onboarding flow has client history in the AI sector but not AI product UX experience.
Several possibilities: their AI product portfolio didn't meet the inclusion threshold, their work with AI companies was primarily in brand or marketing rather than product UX, or they have not yet been reviewed. Reputation and client prestige are not substitutes for demonstrated AI product UX capability.
The agencies in this directory serve fundamentally different AI product types — generative consumer tools, enterprise ML dashboards, autonomous systems, LLM-powered SaaS. Scoring a consumer generative AI specialist against an enterprise intelligent dashboard specialist on the same numeric scale implies a comparability that doesn't exist. Profile-level positioning with explicit best-for and not-a-fit-for guidance communicates fit more accurately than a ranking.
Once per year in full. Individual profiles are updated on a rolling basis when significant new AI product work, methodology changes, or validation signals emerge. Given how rapidly the AI product market moves, agencies that have shipped notable new AI product work between annual reviews may be updated before the next cycle.