Thoughts

·

5 min read

Set the taste. Let the variants come.

A set of 162 variants is no threat to a team that sets its own taste. A clean system, from words to parts, makes every choice traceable and every step reversible.

A sea stack in the surf, wrapped in mist

In short

We set the taste: the flavours, the gradations, the fine distinctions. A clean system from words to parts makes each choice traceable and reversible, so even 162 variants are material, not a burden.

01 The stance

Taste is set, not found. The team decides it and writes it down.

02 The system

Words, field, parts: every decision traces back and can be walked back.

03 The split

Agents widen and sort. People define and decide, and keep a share for bets.

An image model makes five hundred variants of a product before lunch. For a team that has set its taste, that is material, not a problem. Taste here is not a mood. It is a set of decisions: which flavours, which gradations, which fine distinctions, held together across form, technology, requirements and the views of seven departments. With a clean system behind them, every choice can be traced and every step walked back. AI makes this easier, because it lets a team work with more layers than it could ever hold by hand.

Making variants is solved. Judging them is not.

A designer used to sketch hundreds of ideas and render a handful. Tools such as Vizcom now turn a sketch into a photorealistic render in seconds. The cost of making a variant has fallen close to zero. The cost of judging one has not.

The evidence points the same way outside design. When applicants to MIT Solve began to use AI, submissions doubled within a year. Most ideas still fail: at Microsoft, only about a third of well-run experiments improved the metric they were built for. More options do not raise that rate. They raise the number of decisions.

There is a useful twist. In a caption study reported in MIT Sloan Management Review, AI used to generate ideas narrowed their diversity, while AI used to select among them kept diversity close to human levels. Where AI enters the process matters as much as whether it is used at all.

Why swiping fails

Swiping through 162 renders feels like judgement. It is mostly fatigue. Under load, evaluators lean on rules of thumb that punish the unfamiliar, so the novel options go first. Fluent, finished output reads as quality even when it is not. In a field experiment at MIT Solve, reviewers with AI help rejected 9 per cent more proposals, and the AI’s written reasons swayed experts as much as novices.

Models also share blind spots. Language models rate their own text higher than human readers do, and their creative answers are far more alike than people’s, even across model families.

A finished render is judged as a whole, on polish. Fit is hard to see in a picture and easy to check against a sentence.

Taste sits between the departments

Taste is often treated as the private preference of the design team. In a company it is something else, and it is set, not found. Strategy sets the direction and finance decides what gets paid for. Brand and marketing want a promise kept, sales wants something it can sell, product wants the user served, engineering wants functions that work and production wants parts it can make. Neighbours in this ring work most closely together, and each overlap gives a point of orientation. The design language sits in the middle and touches all seven (Exhibit 1).

Exhibit 1

Design language sits where the departments overlap. Taste is choosing inside that overlap.

Seven departments in a ring, each in its own colour, ordered from intent to market to making to money. Neighbours work most closely together; where they overlap sits an orientation point, numbered and coloured as a mix of both. The light blue centre is the design language. The small dots inside it are shared ground, where a department already agrees with it.

Seven departments, one design language Concept

Circle one department, in its colourNumber where two agree, a mix of bothLight blue design languageSmall dot shared ground
Brand and marketingSalesProductEngineeringProductionFinanceStrategy1234567Design languageWhere all seven meet1Brand and salesHow it is shown and sold2Sales and productWhat customers ask for3Product and engineeringFunctions and interfaces4Engineering and productionMaterials, joints, tolerances5Production and financeUnit cost and price point6Finance and strategyBudget and priorities7Strategy and brandPositioning and promiseShared groundWhere a department already agrees with the design languageBrand and marketingSalesProductEngineeringProductionFinanceStrategy1234567Design languageWhere all seven meet1Brand and salesHow it is shown and sold2Sales and productWhat customers ask for3Product and engineeringFunctions and interfaces4Engineering and productionMaterials, joints, tolerances5Production and financeUnit cost and price point6Finance and strategyBudget and priorities7Strategy and brandPositioning and promiseShared groundWhere a department alreadyagrees with the designlanguageBrand and marketingSalesProductEngineeringProductionFinanceStrategy1234567Design languageWhere all seven meet1Brand and salesHow it is shown and sold2Sales and productWhat customers ask for3Product and engineeringFunctions and interfaces4Engineering and productionMaterials, joints, tolerances5Production and financeUnit cost and price point6Finance and strategyBudget and priorities7Strategy and brandPositioning and promiseShared groundWhere a department already agrees with thedesign language

What it means. Taste is not a private preference of the design team. It is the judgement that holds seven sets of requirements together, including who pays. Agents can read all seven at once and keep track of the overlaps. The choice inside them stays with people.

Sources and method

My own model. Departments follow a classic functional organisation; orientation points are examples, not a client organisation.

This is where AI helps first. Agents can read all seven sets of documents at once and keep track of where they agree and where they collide. The choice inside the overlap stays with people.

One process, two kinds of work

My process runs through four phases: discover, define, design, deliver. The amount of material in play opens and closes twice (Exhibit 2). Six directions, each tried in three rounds of nine variants, give 162 variants: enough to compare, few enough to judge. Agents earn their place where the work is wide: reading every source, drafting a word map, making variants and sorting them. People hold the narrow points: setting the core terms, drawing the field, judging and deciding.

Exhibit 2

Agents widen and sort, people define and decide: one brief becomes 162 variants and three decisions.

Four phases from left to right. Three indigo layers are three rounds of iteration, each opening and closing in its own way, the core darkest. Above the dashed line, variants move away from the last winner; below it, they test what is familiar. Green explore lines run on both sides, through every phase.

Discover, define, design, deliver Illustrative

Indigo layers three rounds of iteration, the core darkestDashed line the last winner
Green line an explore line, a bet followed on its ownDot keptRing dropped
Black dot a decisionRing agentsDot people
DISCOVERDEFINEDESIGNDELIVER1brief120inputs20boards162variants6families3decisions+ NEW− FAMILIARAGENTSRead strategy, brand, salesand researchTurn photos into wordsDraft the word mapSort boards by focusMake variantsTag topics, cluster familiesCheck against the no-goCheck that every topic arrivedPEOPLEInterviews and observationSet core terms and draw thefieldSketch, judge each family inone sentenceDecide with the departmentsDISCOVERDEFINEDESIGNDELIVER1brief120inputs20boards162variants3decisions+ NEW− FAMILIARAGENTSRead strategy,brand, sales andresearchTurn photos intowordsDraft the wordmapSort boards byfocusMake variantsTag topics,cluster familiesCheck against theno-goCheck that everytopic arrivedPEOPLEInterviews andobservationSet core termsand draw thefieldSketch, judgeeach family inone sentenceDecide with thedepartmentsDISCOVERDEFINEDESIGNDELIVERFrom 1 brief to 120 inputs, 20 boards, 162variants and 3 decisions.1120201623Above the line: new. Below it: familiar.DISCOVERRead strategy, brand, sales and researchTurn photos into wordsInterviews and observationDEFINEDraft the word mapSort boards by focusSet core terms and draw the fieldDESIGNMake variantsTag topics, cluster familiesCheck against the no-goSketch, judge each family in one sentenceDELIVERCheck that every topic arrivedDecide with the departments

What it means. Exploring is not a strand at the end. It runs through every phase: some bets rejoin the main work, some are kept, most are dropped. AI earns its place where the work is wide. The narrow points stay human, and that is where taste is applied.

Sources and method

My own process. Invented counts and shapes to show the way of working, not measured data. The counts follow one example: 120 inputs from documents, interviews and photos; 20 boards, four topics with five focuses each; six directions, each tried in three rounds of nine variants (162); one family per direction; three decisions.

From words to meaningful objects

Inside that process sits the framework I use to keep judgement traceable. It has four pillars and one rule: every decision traces back to a word (Exhibit 3).

Language comes first. Strategy, brand, sales material, personas, competitor research and positioning were written for other purposes. Agents lift them out of that context and condense them into a word map: core terms, their neighbours and their opposites. Photos of things that suit the brand, described in words, add terms the documents miss. A core term such as quality needs its own definition in each product category. Does it include robustness? Where does it stop? Each core term becomes a topic, one colour that runs through everything that follows.

Atmosphere translates the words into colour, material, finish, light and emotion. Each moodboard has a focus: a detail, a surface, a contrast, a mood, a technical theme. First sketches and design directions grow out of the boards, so the qualities fit the brand’s own world rather than someone else’s. Form is design development in 3D, where the directions become concrete variants. Presence is the whole collection, and here every topic is dosed. Products are built differently, in how they are operated and serviced, in a button, a switch or the way two surfaces meet. So for each product the team decides how strongly a topic should show, and whether directly or only indirectly. The last check is simple: has every topic arrived, at the right strength? If something does not work, the thread shows how far to go back: to the form, to the board or to the word.

Exhibit 3

Every decision traces back to a word: four topics run from the first words to the finished object.

Four pillars from left to right. Each colour is one topic from the word map. A dot shows where the topic appears in a pillar, and the line follows it to the object. Grey dots are experiments that drop out. In presence every product carries all four topics, each dosed: the size of a dot shows its strength, and a ring marks a topic that is present only indirectly.

From the first words to the finished object Illustrative

QualityCalmRobustInviting
Line one topic, followed throughGrey dot variant that drops outRing topic present indirectly
LANGUAGEATMOSPHEREFORMPRESENCEQualitydurable, preciseCalmquiet, softRobustsolid, safeInvitingwarm, openCore terms and their neighbours,worked out with strategy andmarketing until each has its placeMoodboards, first sketches anddesign directions that carry eachtermSketches and 3D. Grey variants aretoo weak, do not combine or cannotbe builtThe whole collection. Each termdosed per product, directly orindirectlyLANGUAGEATMOSPHEREFORMPRESENCEQualityCalmRobustInvitingCore terms and theirneighbours, worked outwith strategy andmarketing until eachhas its placeMoodboards, firstsketches and designdirections that carryeach termSketches and 3D. Greyvariants are too weak,do not combine orcannot be builtThe whole collection.Each term dosed perproduct, directly orindirectlyLANGUAGEATMOSPHEREFORMPRESENCE

What it means. Do not ask which variant looks best. Ask which thread it carries. A variant without a thread drops out, however well it is rendered, and a thread that never reaches the object is a gap in the brief.

Sources and method

Framework: From Words to Meaningful Objects, my own method. Invented values to show the method, not measured data. The four topic names are examples.

Draw the field before you fill it

Moodboards do a second job. Together they mark out a field: a thesis that sets the direction, an antithesis that tests it from the other side, and a no-go zone that shows on purpose what the team rules out. Too cold. Too abstract. Not inviting. Showing the no-go matters, because those looks keep coming back in meetings until someone has put them on the wall and named them.

The same field can be drawn as a map. Take two words from the word map as axes, place the competitors, and look for the space that is free and fits the thesis (Exhibit 4). On generic axes every brand lands in the middle. On your own words the gaps appear. Agents can redraw the map for every pair of words in minutes, which used to take a workshop.

Exhibit 4

Competitors crowd two corners. Demand sits between them, and so does the space that fits the thesis.

Horizontal: soft to precise. Vertical: quiet to expressive. Each bubble is a competitor, sized by market share and coloured by segment. Dashed rings mark where demand sits. Light blue is the free space inside that demand, green the antithesis to test, hatched the corners ruled out.

Positioning on two words from the word map Illustrative

PremiumMainstreamChallengerSize market share
Light blue our spaceGreen to testDashed rings where demand sitsHatched ruled out
Too coldToo playfulOur spaceprecise, calm, not coldTo testWhere demand sitsBrand ABrand BBrand CBrand DBrand EBrand FBrand GBrand HBrand IBrand JBrand KBrand LBrand MBrand NEXPRESSIVEQUIETSOFTPRECISEToo coldToo playfulOur spaceTo testWhere demand sitsBrand ABrand BBrand CBrand DBrand FBrand GBrand HBrand IBrand JBrand KBrand LEXPRESSIVEQUIETSOFTPRECISEToo coldToo playfulOur spaceTo testBrand ABrand FBrand GEXPRESSIVEQUIETSOFTPRECISE

What it means. White space only counts where there is demand. On generic axes every brand lands in the middle. On words from your own map the gap appears, and agents can redraw the map for every pair of words in minutes.

Sources and method

Invented brands, shares and positions to show the method. Not measured data and not real companies. On market share as bubble size and demand on positioning maps: Wikipedia, Perceptual mapping.

Judge the part, not the render

Now the 162 variants arrive, and they are not judged as renders. The product is taken apart: vents, joints, surface transitions, the bezel of a display, buttons, light. Each part gets a spectrum between two words, with competitors next to your own variants, each shown alone and in the same style (Exhibit 5). Researchers who study innovation recommend this separation of presentation from evaluation, because polish is otherwise mistaken for quality.

Exhibit 5

Judge the part, not the render: only variants with a topic inside the target zone are kept.

Six parts of an electrical device. Each row places the variants for one part between two words from the word map, soft and precise. The light blue pill is the target zone from the thesis. Rings are competitors.

Six parts of an electrical device Illustrative

QualityCalmRobustInviting
Grey dot carries no topicRing competitorLight blue target zone from the thesis
PARTKEPTSOFTPRECISEVents3 of 8Joints1 of 8Surface2 of 8Display bezel2 of 8Buttons1 of 8Light2 of 8Grey inside the blue zone: fits the form, carries no topic, so it is dropped.PARTKEPTSOFTPRECISEVents3 of 8Joints1 of 8Surface2 of 8Display bezel2 of 8Buttons1 of 8Light2 of 8SOFTPRECISEVents3 of 8 keptJoints1 of 8 keptSurface2 of 8 keptDisplay bezel2 of 8 keptButtons1 of 8 keptLight2 of 8 kept

What it means. A render hides its parts. Split the product and each decision becomes small enough to judge in one sentence. Grey dots inside the zone fit the form but carry no topic: they are the ones a swipe would keep.

Sources and method

Invented values to show the method, not measured data. Positions are judgements between two words, not measurements. Competitor positions are examples.

Agents tag each variant with the topics it carries, cluster similar ones into families and flag what crosses into the no-go zone. To avoid shared blind spots, use different models for making and for checking. People then decide on families, with one written sentence: which topic this family carries, and why it is the best expression of it. In the MIT Solve experiment, plain AI recommendations improved decisions, while the same recommendations with persuasive written reasons did not. A reason written by a person keeps the judgement with the person.

Keep a share for bets

Variants derived from the last winner climb the same hill, a long-standing criticism of A/B testing. The remedy is an explore share: a fixed part of the variants that deliberately does not derive from the last winner. Coca-Cola’s content plan set aside 10 per cent for high-risk ideas in the same spirit. The field gives the share a rule. Bets come from the antithesis, never from the no-go zone.

There is a catch. If every team runs the same share with the same models, the bets converge too. A first study found that newer models give less varied creative answers than older ones. Your own photos, your hand sketches and the words in your own map are the inputs a competitor cannot copy.

What would prove me wrong

If teams that judge whole renders by eye, with an explore share near zero, hold the same impact over three years as teams that work from a word map, this method is overhead.

Next

When a brand’s AI learns mostly from the brand’s own earlier work, its taste narrows. How to keep feeding it from outside is the next note.

Sources (11)
Judging ideas
Lane and others, field experiment with MIT Solve, HBS Working Knowledge, 2025De Freitas and others, A Human Bottleneck Framework, preprint 2026
Sameness
Boussioux, Doshi, Hauser and Hosanagar, MIT Sloan Management Review, 2026Panickssery, Bowman and Feng, NeurIPS 2024Wenger and Kenett, Creative Homogeneity Across LLMs, preprint 2025Patel and others, Are LLMs becoming similarly creative?, preprint 2026
Experiments
Kohavi, Crook and Longbotham, Online Experimentation at Microsoft, 2009Jason Cohen on the A/B testing trap, 2010Chris Said, Four pitfalls of hill climbing, 2016
Practice
Coca-Cola Content 2020, as reported by TopRankVizcom on AI rendering for product designers
The Milky Way in a night sky
Shingle of many colours at the water’s edge

Create a free website with Framer, the website builder loved by startups, designers and agencies.