# Is Taste the New Moat? What Still Defends Software When AI Builds Everything (2026)

> Taste is the new moat, says everyone in AI. The data says taste is a filter, not a wall. See what still defends software in 2026, with sources and a playbook.

- Source: https://www.taskade.com/blog/taste-moat-explained
- Published: 2026-11-15
- Author: John Xie

Bodiam Castle has stood in East Sussex since 1385. Its walls are thin. A serious siege would have broken them. What made the castle hard to take was the water around it, and that picture gave business strategy its favorite word. In 1986, Warren Buffett described GEICO's cost advantage as ["a kind of moat that protects a valuable and much-sought-after business castle."](https://www.berkshirehathaway.com/letters/1986.html)

Forty years later, AI drained a lot of moats. Code that took a team six months now takes one person a weekend. So the industry went looking for a new one, and in early 2026 it settled on a single word: **taste**.

This guide tests that claim against the evidence. The short version: taste matters more than it ever did, and it is still not a moat. It is the thing you use to build one.

> **TL;DR:** Taste is a filter, not a wall. App releases rose 60% year over year in Q1 2026 while studies show AI-assisted output converging, so judgment is scarce. But judgment alone can be copied. The moats that hold in 2026 are accumulated context, workflow depth, trust, owned distribution and taste written into process. [Build yours in Taskade Genesis →](https://www.taskade.com/create)

*Photo: Bodiam Castle and moat. Paul Gillett / Wikimedia Commons / CC BY-SA 2.0.*

## ⚡ Quick Answer: Is Taste a Moat?

Taste is not a moat on its own, because a moat must survive a competitor who can see your product and copy it. Taste decides what you build and what you refuse to build. It becomes defensible only after you encode it in process and compound it with private context. Until then it is an edge with a half-life.

| Question | Short answer | Why |
|---|---|---|
| Is taste scarce in 2026? | Yes | Supply of software exploded, attention did not |
| Is taste a moat by itself? | No | Visible choices can be copied, and models learn stable criteria |
| Can taste become a moat? | Yes | Written standards become process power, and recorded decisions become switching costs |
| What is taste for? | Choosing | It filters what gets built, shipped and killed |
| What defends software now? | Five things | Context, workflow depth, trust, owned distribution, encoded taste |

```mermaid
graph TD
    A[AI makes execution cheap] --> B[Supply of software explodes]
    B --> C[Output starts to look the same]
    C --> D[Judgment becomes the scarce input]
    D --> E{Is judgment written down and compounded?}
    E -->|No| F[Temporary edge that decays]
    E -->|Yes| G[Process power and switching costs]
    G --> H[A moat that survives copying]
    style A fill:#1e1e2e,color:#94a3b8,stroke:#475569
    style D fill:#0d1b2a,color:#7dd3fc,stroke:#38bdf8,stroke-width:1.5px
    style F fill:#1a140a,color:#ffb74d,stroke:#e65100,stroke-width:1.5px
    style G fill:#0d1a0e,color:#86efac,stroke:#22c55e,stroke-width:1.5px
    style H fill:#2a1225,color:#ff8fa3,stroke:#ff2d60,stroke-width:2px
```

## 🗣️ What Does "Taste Is the New Moat" Mean?

"Taste is the new moat" is the claim that when AI can produce any artifact on request, the lasting advantage belongs to people who know which artifact is worth producing. The phrase spread through tech in the first quarter of 2026, and it drew a counter-argument within days. Both sides are worth reading in their own words.

| Date | Who | What they said | Side |
|---|---|---|---|
| Feb 2002 | Paul Graham, ["Taste for Makers"](https://paulgraham.com/taste.html) | "We need good taste to make good things." | For, 24 years early |
| Jan 13, 2026 | Cong Wang | [Personal taste is the moat](https://wangcong.org/2026-01-13-personal-taste-is-the-moat.html) | For |
| Feb 14, 2026 | Paul Graham | ["In the AI age, taste will become even more important. When anyone can make anything, the big differentiator is what you choose to make."](https://x.com/paulg/status/2022604692178522562) | For |
| Feb 16, 2026 | Shrivu Shankar | ["Taste feels more like alpha: a decaying edge, only valuable relative to a rising baseline."](https://blog.sshh.io/p/taste-is-not-a-moat) | Against |
| Feb 2026 | Nan Yu, then at Linear | ["I hate to break this to everyone, but you probably don't have better taste than the AI."](https://x.com/thenanyu/status/2024142295235264735) | Against |
| Feb 2026 | Jens Axboe | ["Nonsense, taste has always been a core skill. It's just easier to see now."](https://x.com/axboe/status/2023559692375781687) | Neither |
| Mar 5, 2026 | Ellison Carter, Colorado State | ["Engineering under scarcity produces competence. Engineering under abundance demands taste."](https://hsph.harvard.edu/news/essay-3-taste-and-the-end-of-scarcity/) | For |
| Jun 24, 2026 | Forbes | [Every company is now an AI wrapper, so go-to-market is the new moat](https://www.forbes.com/sites/josipamajic/2026/06/24/every-company-is-now-an-ai-wrapper-so-gtm-is-the-new-moat/) | Distribution |
| Aug 26, 2026 | Tim Sullivan, a16z | ["If taste consists of stable criteria that experts consistently apply, then a modern AI can almost certainly parrot taste."](https://a16zfrontier.substack.com/p/beyond-slop-taste-and-ai-moral-panic) | Against |

In February 2026, Fortune reported Sam Altman's view that [good taste may get you a job](https://finance.yahoo.com/news/good-taste-may-just-job-224031255.html) in the AI era. By May, a video essay by Hak titled ["Is taste really the only moat left?"](https://www.youtube.com/watch?v=FlTp9Ij7Mbo) gave the argument its sharpest image: software just got fast fashion. This guide builds on that essay's three layers of taste and its three reps, and adds the strategy test and the data that the debate was missing.

```mermaid
graph TD
    A[2002: Taste for Makers] --> B[Jan 2026: first taste-is-the-moat essays]
    B --> C[Feb 14: Paul Graham prediction]
    C --> D[Feb 16: Taste is not a moat rebuttal]
    D --> E[Mar 5: Carter essay on abundance]
    E --> F[Jun 24: Forbes says distribution]
    F --> G[Aug 26: a16z says models can learn it]
    style A fill:#1e1e2e,color:#94a3b8,stroke:#475569
    style C fill:#2a1225,color:#ff8fa3,stroke:#ff2d60,stroke-width:2px
    style D fill:#1a140a,color:#ffb74d,stroke:#e65100,stroke-width:1.5px
    style E fill:#2a1225,color:#ff8fa3,stroke:#ff2d60,stroke-width:2px
    style F fill:#0d1b2a,color:#7dd3fc,stroke:#38bdf8,stroke-width:1.5px
    style G fill:#1a140a,color:#ffb74d,stroke:#e65100,stroke-width:1.5px
```

Notice who is on the "against" side. One of the strongest sceptics worked at Linear when he posted it, the product that everyone on the "for" side holds up as proof. That is a hint that the two camps are describing different things. One camp describes a skill. The other describes a business defense. A skill can be valuable and still fail the test for a moat.

## 📈 Why Is Everyone Talking About Taste in 2026?

Because supply exploded. In the first quarter of 2026, new app releases rose 60% year over year across both app stores and 80% on iOS, [according to Appfigures data reported by TechCrunch](https://techcrunch.com/2026/04/18/the-app-store-is-booming-again-and-ai-may-be-why/). When the cost of making software falls toward zero, the scarce resource stops being the ability to make it.

| Signal | Number | Source and date |
|---|---|---|
| App releases, both stores, Q1 2026 | +60% year over year | [Appfigures via TechCrunch, Apr 2026](https://techcrunch.com/2026/04/18/the-app-store-is-booming-again-and-ai-may-be-why/) |
| iOS app releases, Q1 2026 | +80% year over year | Same |
| New GitHub repositories | More than 230 a minute | [GitHub Octoverse, Oct 2025](https://github.blog/news-insights/octoverse/octoverse-a-new-developer-joins-github-every-second-as-ai-leads-typescript-to-1/) |
| GitHub commits in 2025 | About 986 million, +25% | Same |
| Projects built or updated on Lovable | About 200,000 a day | [TechCrunch, Mar 2026](https://techcrunch.com/2026/03/11/lovable-says-it-added-100m-in-revenue-last-month-alone-with-just-146-employees/) |
| Registered domain names, Q3 2026 | 411.6 million, +8.7% | [Verisign DNIB](https://blog.verisign.com/domain-names/q3-2026-domain-name-industry-brief-quarterly-report/) |
| Share of Microsoft code written by AI | "Maybe 20%, 30%" | [Satya Nadella via CNBC, Apr 2025](https://www.cnbc.com/2025/04/29/satya-nadella-says-as-much-as-30percent-of-microsoft-code-is-written-by-ai.html) |

```mermaid
xychart-beta
    title "Software supply is growing faster than attention (year over year, %)"
    x-axis [App releases Q1 2026, iOS releases Q1 2026, GitHub commits 2025, Domains Q3 2026]
    y-axis "Growth (%)" 0 --> 100
    bar [60, 80, 25, 8.7]
```

Carter's line explains the shift in one sentence. When building software took scarce engineers and long timelines, knowing *how* to build was the work. Time did the filtering for you: anything that took six months had to earn those six months. Now the first 80% of a product arrives in an afternoon, so the filter has to come from the builder. We covered the supply side of this in [the state of vibe coding](https://www.taskade.com/blog/state-of-vibe-coding) and the casualties in [the vibe coding graveyard](https://www.taskade.com/blog/vibe-coding-graveyard).

The result is familiar to anyone who has opened a launch feed this year. Onboarding flows that ask everything and explain nothing. Settings pages that look like spreadsheets. Newsletters you can identify as machine-written three sentences in. We have a name for the output, and a full guide to it: [AI slop](https://www.taskade.com/blog/ai-slop-explained).

## 🏰 What Is a Moat, Exactly?

A moat is a structural advantage that lets a business keep its customers and its margins even when competitors can see what it does. The test is simple and harsh: if a well-funded rival copied your product tomorrow, what would still stop your customers from leaving? Whatever survives that question is the moat.

The best working list is Hamilton Helmer's [*7 Powers*](https://www.goodreads.com/book/show/32816087-7-powers) (2016). Here is each power, and what AI abundance did to it.

| Power (Helmer) | What it is | What AI abundance did | Status in 2026 |
|---|---|---|---|
| Scale economies | Unit cost falls as volume grows | Compute rewards scale, so this power moved to model labs and clouds | Holds for the few |
| Network economies | Each user adds value for others | Agents can be users too, but networks still need people or data inside | Holds |
| Counter-positioning | A model the incumbent cannot copy without hurting itself | Credits versus seats is the live example | Holds, for a while |
| Switching costs | Leaving costs the customer something | Feature lock-in fell, and context lock-in rose | Changed shape |
| Branding | Customers pay more for the name | Trust is worth more when everything looks the same | Stronger |
| Cornered resource | Preferential access to a scarce asset | Private data and accumulated context are the new scarce assets | Stronger |
| Process power | An organization that does hard things routinely | This is where encoded taste lives | Stronger |

Read the table again and look for taste. It is not one of the seven. A skill is not a power until it is built into the company. That is the gap in most of the writing on this topic: it argues about whether taste is valuable, when the useful question is which power it feeds.

```mermaid
graph LR
    T[Taste: the judgment of people] --> P[Process power]
    T --> B[Branding]
    C[Context: what the product remembers] --> S[Switching costs]
    C --> R[Cornered resource]
    D[Distribution you own] --> N[Network economies]
    style T fill:#2a1225,color:#ff8fa3,stroke:#ff2d60,stroke-width:2px
    style C fill:#0d1b2a,color:#7dd3fc,stroke:#38bdf8,stroke-width:1.5px
    style D fill:#0d1a0e,color:#86efac,stroke:#22c55e,stroke-width:1.5px
    style P fill:#11111b,color:#cdd6f4,stroke:#00d4aa
    style B fill:#11111b,color:#cdd6f4,stroke:#00d4aa
    style S fill:#11111b,color:#cdd6f4,stroke:#00d4aa
    style R fill:#11111b,color:#cdd6f4,stroke:#00d4aa
    style N fill:#11111b,color:#cdd6f4,stroke:#00d4aa
```

## 🎨 What Is Taste? Three Layers Most Definitions Blur

Taste is the ability to recognize that something is right, or wrong, before you can explain why. It is a real form of knowledge, and it is not the same as preference. Most arguments about taste go in circles because the word covers three different layers, and people defend one while attacking another.

| Layer | What it is | Where you see it | Can a model do it? |
|---|---|---|---|
| 1. Recognition | Knowing an output is off before you can say why | Rejecting all 40 drafts and rewriting the prompt | Partly, for stable criteria |
| 2. Plumbing | Quality in the parts users never see | Empty states, defaults, shortcuts, pacing | Only with your context |
| 3. Restraint | Deciding what should not exist | The feature you refused to ship | No, it needs accountability |

```mermaid
graph TD
    L1[Layer 1: Recognition] --> Q1[Is this output right?]
    L2[Layer 2: Plumbing] --> Q2[Is the unseen part right?]
    L3[Layer 3: Restraint] --> Q3[Should this exist at all?]
    Q1 --> V[What users see in the first minute]
    Q2 --> F[What users feel in the first month]
    Q3 --> K[What users never have to deal with]
    style L1 fill:#0d1b2a,color:#7dd3fc,stroke:#38bdf8,stroke-width:1.5px
    style L2 fill:#1a0d24,color:#ce93d8,stroke:#6a1b9a,stroke-width:1.5px
    style L3 fill:#2a1225,color:#ff8fa3,stroke:#ff2d60,stroke-width:2px
    style V fill:#11111b,color:#cdd6f4,stroke:#00d4aa
    style F fill:#11111b,color:#cdd6f4,stroke:#00d4aa
    style K fill:#11111b,color:#cdd6f4,stroke:#00d4aa
```

**Layer 1 is recognition.** A piano tuner hears a flat note before measuring it. A chef knows a sauce needs something before knowing what. Designer Dan Alldis calls this ["the ability to feel the delta between almost and actually."](https://danalldis.substack.com/p/tasty) In AI work, it is the moment you look at a wall of generated options and realize the prompt was wrong.

**Layer 2 is plumbing.** Many products copied the surface of the tools they admire: dark mode, keyboard-first navigation, the same typography. Few copied the discipline underneath. Linear co-founder Karri Saarinen puts the difference this way: ["For quality, you need a team that views the spec as the baseline, not the finish line."](https://www.figma.com/blog/karri-saarinens-10-rules-for-crafting-products-that-stand-out/) Dieter Rams compressed the same idea into three words in 1995: *Weniger, aber besser*. [Less, but better](https://www.cooperhewitt.org/2018/04/16/less-but-better/).

**Layer 3 is restraint.** This is the uncomfortable one. Builders used to ask "can I ship this?" The answer is now always yes, so the only useful question left is "should this exist?" Paul Graham made the constraint concrete back in 2002: ["When you're forced to be simple, you're forced to face the real problem."](https://paulgraham.com/taste.html)

A note on a famous line. You will often read that Michelangelo said the statue was already in the marble and he only carved away what was not David. [Quote Investigator traces that story](https://quoteinvestigator.com/2014/06/22/chip-away/) to an 1858 book review and to later jokes credited to Ruskin and others, with no evidence that Michelangelo said it. It is a good image and a false attribution, which is a small lesson in taste by itself: check the plumbing.

## ⚖️ Is Taste Actually a Moat? The Case For and Against

No. Taste passes the scarcity test and fails the durability test. It is scarce because few people exercise it. It is not durable by itself because its output is public, and anything public can be studied, copied and learned by people and by models. The honest verdict takes both sides seriously.

| The case for | The case against |
|---|---|
| Supply is infinite, so selection is the bottleneck | A moat must survive copying, and shipped choices are visible |
| Homogenized output makes distinct work stand out | Models learn stable criteria, so the baseline keeps rising |
| Restraint needs accountability, which a model lacks | "Taste" can be a status word with no measurable content |
| Linear, Apple and 37signals built durable brands on it | Those companies also have process, brand and distribution |
| It decides what gets built, which decides everything after | An edge "only valuable relative to a rising baseline" is alpha, not a wall |

Shankar's word for it is the right one: **alpha**. In investing, alpha is an edge over the market that shrinks as others discover it. Taste works the same way. A fresh interaction pattern is alpha on the day you ship it and table stakes two quarters later.

The companies that look like they have a "taste moat" did something more specific. They converted taste into things that do not decay at the same rate.

```mermaid
stateDiagram-v2
    state "Personal taste" as PersonalTaste
    state "Shipped choices" as ShippedChoices
    state "Copied by rivals" as Copied
    state "Written standard" as WrittenStandard
    state "Process power" as ProcessPower
    state "Recorded decisions" as RecordedDecisions
    state "Compounding context" as CompoundingContext
    [*] --> PersonalTaste
    PersonalTaste --> ShippedChoices: you build with it
    ShippedChoices --> Copied: competitors study the output
    ShippedChoices --> WrittenStandard: you write the rules down
    WrittenStandard --> ProcessPower: every person and agent applies them
    ProcessPower --> RecordedDecisions: approvals and rejections are stored
    RecordedDecisions --> CompoundingContext: the system learns your preferences
    CompoundingContext --> [*]
    Copied --> [*]
```

There are two exits in that diagram. Taste that stays in one person's head exits through "Copied." Taste that is written down, enforced and recorded exits through compounding context. Apple is the classic example: the taste of a few people became a review culture, then a brand, then a habit for a billion customers. We traced that arc in [the history of Apple](https://www.taskade.com/blog/apple-history) and in [Bill Atkinson's HyperCard](https://www.taskade.com/blog/hypercard-history), where a single designer's standards became a tool that millions of non-programmers used to build.

## 🧪 Does AI Make Everything Look the Same?

Yes, at the group level, and the evidence is consistent. AI assistance raises the quality of an average individual output and lowers the variety across outputs. That combination is exactly the condition under which judgment becomes valuable: everyone gets the same strong first draft, so the difference is made after the draft.

| Study | Finding | Link |
|---|---|---|
| Doshi and Hauser, *Science Advances*, 2024 | Stories written with AI help were rated more creative and were more similar to each other | [science.org](https://www.science.org/doi/10.1126/sciadv.adn5290) |
| Padmakumar and He, ICLR 2024 | Writing with an instruction-tuned model reduced the diversity of essays | [arXiv](https://arxiv.org/abs/2309.05196) |
| Anderson, Shah and Kreminski, 2024 | Ideas from ChatGPT users were more alike at the group level | [arXiv](https://arxiv.org/abs/2402.01536) |
| Agarwal, Naaman and Vashistha, CHI 2025 | AI suggestions pushed writing toward Western styles | [ACM](https://dl.acm.org/doi/10.1145/3706598.3713564) |
| Holzner and others, 2025 meta-analysis | Creativity up (g = 0.27), diversity of ideas down (g = -0.86) | [arXiv](https://arxiv.org/abs/2505.17241) |

The meta-analysis is the number to remember. A positive effect of 0.27 on creative performance, and a negative effect of 0.86 on diversity. The floor went up. The spread collapsed.

```mermaid
xychart-beta
    title "AI assistance: a small gain in quality, a large loss in variety"
    x-axis [Gain in creativity, Loss in diversity of ideas]
    y-axis "Effect size (absolute g)" 0 --> 1
    bar [0.27, 0.86]
```

*Source: Holzner and others, 2025 meta-analysis. The diversity effect is negative. The chart shows its size.*

This is why so many 2026 products feel alike when you first open them. They come from the same factories, trained on the same patterns. It is also why the editorial habit matters more than the generating habit. The working rule from practitioners is short: **what the model gives you is the starting line, not the finish line.**

## 👗 How Did Software Get Fast Fashion?

Software got fast fashion when production cost fell below the cost of judgment. Clothing went through the same change: production got cheap, patterns got copied, fabric got thinner. Shein lists ["more than 6,000 new fashion, beauty and lifestyle products daily"](https://builtin.com/company/shein-technology-llc) by its own account, and [up to 10,000 by NPR's reporting](https://www.npr.org/2023/10/13/1204983212/shein-america-fast-fashion-legal-issues). Much of it does not outlast the season.

| | Fast fashion | Fast software |
|---|---|---|
| What got cheap | Cutting and sewing | Writing code, copy and layouts |
| What got copied | Patterns from runways | Patterns from training data |
| What got thinner | Fabric | Architecture, the parts nobody inspects |
| Daily output | Thousands of new items per retailer | About 200,000 project builds a day on one platform |
| What the buyer feels | Looks right, falls apart | Demos well, nobody returns |
| Who still wins | Brands with a point of view and repeat buyers | Products with a point of view and retained users |

```mermaid
graph TD
    subgraph Fashion
        F1[Production gets cheap] --> F2[Patterns get copied]
        F2 --> F3[Thousands of new items a day]
        F3 --> F4[Market splits: volume or point of view]
    end
    subgraph Software
        S1[Code gets cheap] --> S2[Patterns come from training data]
        S2 --> S3[About 200,000 builds a day on one platform]
        S3 --> S4[Market splits: slop or craft]
    end
    style F4 fill:#1a140a,color:#ffb74d,stroke:#e65100,stroke-width:1.5px
    style S4 fill:#2a1225,color:#ff8fa3,stroke:#ff2d60,stroke-width:2px
    style Fashion fill:#1a140a,stroke:#3a2a1a
    style Software fill:#0a1420,stroke:#1a2a45
```

The comparison is useful for what it does *not* say. Fast fashion did not kill clothing brands. It split the market. One part competes on price and volume. The other part competes on fit, material and trust, and it charges more than before. Software is heading for the same split, and the middle is where it hurts.

A fair word for the tools: cheap production is a real gift. Ideas that used to die in the gap between thinking and shipping now get built. The problem is not speed. The problem is shipping the prototype as if it were the product. A first draft is supposed to be rough. The failure is calling it finished and walking away.

## 🏚️ What Do monday.com and Perplexity Teach About Rented Advantage?

They teach that an advantage you rent can be repriced by the landlord. monday.com rented its distribution from ad and search platforms. Perplexity rents its intelligence from model vendors. Both are real businesses with real revenue, and both spent 2026 showing where the exposure sits.

### monday.com: rented distribution

monday.com is one of the great growth stories in software. It also built that growth on paid acquisition. Its IPO filing shows sales and marketing at [$191.4 million in 2020, which was 119% of revenue](https://www.sec.gov/Archives/edgar/data/1845338/000110465921067423/tm211978-4_f1.htm). The plan worked: the company passed [$1 billion in annual recurring revenue in August 2024](https://ir.monday.com/news-and-events/news-releases/news-details/2024/monday.com-Achieves-Significant-Financial-Milestone-with-1-Billion-in-Annual-Recurring-Revenue/default.aspx).

Then the channel changed. On the Q2 2025 earnings call, management told analysts it saw ["some softness within the down market due to the changes in the Google algorithm."](https://www.fool.com/earnings/call-transcripts/2025/12/23/mondaycom-mndy-q2-2025-earnings-call-transcript/) The stock fell 29.8% in a day. In February 2026 the company guided to 2026 revenue of $1.452 billion to $1.462 billion and stopped discussing its 2027 targets, and the stock fell another 21%. In July 2026 it announced a cut of [about 20% of its staff](https://www.calcalistech.com/ctechnews/article/udx1nmbdq). The full sourced timeline is in our [history of monday.com](https://www.taskade.com/blog/monday-com-history).

### Perplexity: rented intelligence

Perplexity built a product people love on top of models it does not own. The Information reported that in 2024 the company took in about $34 million in revenue and spent at least $57 million on AI models and web services ([summary](https://www.aicerts.ai/news/perplexity-margins-debate-accounting-choices-drive-profit-story/)). From 2025, the response was predictable: caps on heavy features, a [$200 a month Max tier](https://techcrunch.com/2025/07/02/perplexity-launches-a-200-monthly-subscription-plan/), and a credit system for its agent product.

The story is not a collapse. By August 2026, annualized revenue had [passed $750 million](https://www.fastcompany.com/91600914/maybe-we-were-wrong-about-perplexity), even as its share of AI web traffic fell. The lesson is narrower and more useful: when your cost of goods is someone else's price list, your margin is someone else's decision. See the [history of Perplexity](https://www.taskade.com/blog/perplexity-ai-history) for the full ledger.

```mermaid
graph TD
    R1[Paid search and ads] --> R2[The platform changes the rules]
    R3[Third-party models] --> R4[The vendor changes the price]
    R2 --> X[Rented: growth slows, margin shrinks]
    R4 --> X
    style R1 fill:#1e1e2e,color:#94a3b8,stroke:#475569
    style R3 fill:#1e1e2e,color:#94a3b8,stroke:#475569
    style R2 fill:#1a140a,color:#ffb74d,stroke:#e65100,stroke-width:1.5px
    style R4 fill:#1a140a,color:#ffb74d,stroke:#e65100,stroke-width:1.5px
    style X fill:#1a0d0d,color:#ef9a9a,stroke:#b71c1c,stroke-width:1.5px
```

```mermaid
graph TD
    O1[Context earned through use] --> O2[Gets deeper with every session]
    O3[Standards written into process] --> O4[Applied by every person and agent]
    O2 --> Y[Owned: leaving costs the customer something]
    O4 --> Y
    style O1 fill:#0d1b2a,color:#7dd3fc,stroke:#38bdf8,stroke-width:1.5px
    style O3 fill:#1a0d24,color:#ce93d8,stroke:#6a1b9a,stroke-width:1.5px
    style O2 fill:#11111b,color:#cdd6f4,stroke:#00d4aa
    style O4 fill:#11111b,color:#cdd6f4,stroke:#00d4aa
    style Y fill:#0d1a0e,color:#86efac,stroke:#22c55e,stroke-width:1.5px
```

| | Rented | Owned |
|---|---|---|
| Distribution | Ads, search rank, platform feeds | Community, marketplace, habit, word of mouth |
| Intelligence | A frontier model through an API | The context and instructions around the model |
| Data | Public web, the same for everyone | What your product learned from your customers |
| Standards | A style copied from a leader | A written spec your team enforces |
| What a rival needs to match it | A budget | Time, and your customers' permission |

This is the same force behind the February 2026 software selloff and the move away from pure seat pricing. We covered both in [the SaaSpocalypse explained](https://www.taskade.com/blog/saaspocalypse-explained) and [the great SaaS unbundling](https://www.taskade.com/blog/great-saas-unbundling).

## 🧱 What Moats Still Work in the Age of AI?

Five moats hold up in the 2026 evidence: accumulated context, workflow depth, earned trust, owned distribution and taste encoded as process. Each one shares a property. A competitor cannot buy it with a bigger model budget, because it is produced by time spent with real customers.

```text
            THE ABUNDANCE MOAT STACK (2026)

  5  ENCODED TASTE        written standards, review gates
     ----------------------------------------------------
  4  OWNED DISTRIBUTION   community, marketplace, habit
     ----------------------------------------------------
  3  EARNED TRUST         brand, reliability, security
     ----------------------------------------------------
  2  WORKFLOW DEPTH       the product runs the work
     ----------------------------------------------------
  1  ACCUMULATED CONTEXT  memory the product earned
     ====================================================
     RENTED LAYER         models, compute, ad channels
     (everyone can buy this, so it defends nothing)
```

| Moat | What it looks like | Helmer power | Evidence |
|---|---|---|---|
| Accumulated context | The product remembers decisions, people and preferences | Cornered resource, switching costs | Every approval or rejection teaches the system |
| Workflow depth | The product executes the work, not just displays it | Switching costs | At AI Ascent 2026, Sequoia placed the AI moat in customer intimacy, workflow integration and earned trust ([recap](https://www.theaiopportunities.com/p/sequoia-ai-ascent-2026-the-future)) |
| Earned trust | Customers believe the output without checking | Branding | Convergence makes a trusted name worth more |
| Owned distribution | A channel no platform can turn off | Network economies | The monday.com search story above |
| Encoded taste | Standards that people and agents both follow | Process power | The Linear and Apple pattern |

### Is context really a moat if memory is portable?

Context at the model layer is not a moat, and the early evidence points one way. In March 2026 Anthropic shipped a tool that [imports your memories from other assistants](https://www.macrumors.com/2026/03/02/anthropic-memory-import-tool/). A 2026 paper surveyed 388 AI chat users. The 170 active users in it used [2.83 AI platforms on average, and only 17.6% used a single one](https://arxiv.org/pdf/2603.25220). Chat memory moves.

Context at the **workflow** layer is different. A list of facts about you is portable. A working system is not: the client records, the agents trained on your documents, the automations wired to your tools, and six months of approve and reject decisions that taught the system how you work. Agency builder Luuk Alleman [put it in a single sentence](https://www.youtube.com/watch?v=fLgKYFvuL9k): "The switching cost is the memory." For the long version, see [the history of agent memory](https://www.taskade.com/blog/agent-memory-history) and [context engineering](https://www.taskade.com/blog/context-engineering).

## 🤖 Who Are You Building For: Humans or Agents?

Both, and they want opposite things. An AI agent reading your product does not care how it looks. It wants structure, clean interfaces and predictable patterns. A human in a world full of generated sameness cares more about feel than before. The temptation is to serve the agent and ship slop to the human. That is backwards.

| | Agents want | Humans want |
|---|---|---|
| Interface | Structured data, stable APIs, Markdown | Clarity, pacing, a point of view |
| Tolerance for boring | Unlimited | Low and falling |
| What earns a return visit | Reliability | How the product made them feel |
| Cost of slop | None | Churn |
| Your job | Be legible | Be worth returning to |

```mermaid
graph TD
    P[One product] --> A[Structured surface: APIs, Markdown, stable patterns]
    P --> H[Crafted surface: pacing, defaults, a point of view]
    A --> AG[AI agents read it and act]
    H --> HU[People feel it and return]
    AG --> O[Work gets done]
    HU --> O
    style P fill:#2a1225,color:#ff8fa3,stroke:#ff2d60,stroke-width:2px
    style A fill:#0d1b2a,color:#7dd3fc,stroke:#38bdf8,stroke-width:1.5px
    style H fill:#1a0d24,color:#ce93d8,stroke:#6a1b9a,stroke-width:1.5px
    style O fill:#0d1a0e,color:#86efac,stroke:#22c55e,stroke-width:1.5px
```

This is why "agent-readable" and "delightful" are not a trade-off. The best products in 2026 expose a plain, structured surface for agents and keep a crafted one for people. Our own site does this with Markdown copies of documentation pages for agents and designed pages for readers. The bar for human-facing software went up, not down. For the agent side of that design problem, see [what AI agents are](https://www.taskade.com/blog/what-are-ai-agents) and [the agent harness explained](https://www.taskade.com/blog/agent-harness-explained).

## 💗 Is the MVP Dead? MVP vs Minimum Lovable Product

The MVP is not dead, but it stopped being a goal and became a starting condition. A minimum viable product proves that something works. When anyone can produce a working version by Sunday night, "it works" no longer separates you from 100 other versions of the same idea. The bar moved to the minimum lovable product.

| | Minimum viable product | Minimum lovable product |
|---|---|---|
| Question it answers | Does this work? | Will anyone come back? |
| Cost in 2026 | A weekend and a subscription | The weekend, plus judgment |
| What it tests | Demand for the function | Demand for *your* version |
| Typical failure | Nobody needed it | Nobody remembered it |
| Scarce input | Engineering time (in the past) | Taste |

```mermaid
stateDiagram-v2
    state "Shipped as is" as ShippedAsIs
    state "Refined with people" as Refined
    state "Forgotten" as Forgotten
    state "Lovable" as Lovable
    [*] --> Idea
    Idea --> Prototype: a weekend with AI
    Prototype --> ShippedAsIs: call the draft the product
    Prototype --> Refined: edit, cut, test with people
    ShippedAsIs --> Forgotten: nobody returns
    Refined --> Lovable: people come back
    Forgotten --> [*]
    Lovable --> [*]
```

Brian de Haaff of Aha! has written that he [first wrote about the minimum lovable product in 2013](https://www.aha.io/blog/minimum-viable-product-vs-minimum-lovable-product). The idea waited a decade for its moment. Speed is no longer a reason to skip the lovable part, because speed is no longer scarce.

## 🛠️ How Do You Turn Taste Into a Moat?

Turn taste into a moat in five moves: write it down, gate on it, record every decision, feed the record back, and cut on a schedule. The goal is to move judgment out of one person's head and into a system that every teammate and every AI agent applies the same way, and that gets better with use.

| Step | What you do | What it produces |
|---|---|---|
| 1. Write the spec | One page: what we always do, what we never do, and five examples of each | A standard people and agents can read |
| 2. Add a gate | Nothing ships without a named human approving it | Accountability |
| 3. Record decisions | Store every approve, edit and reject with the reason | A private dataset of your judgment |
| 4. Feed it back | Give the record to your agents as knowledge | Output that starts closer to your bar |
| 5. Kill weekly | Remove one feature, section or project every week | Focus, and room for the right thing |

Step 1 is where most teams stop, and it is the cheapest step. A taste spec does not need to be clever. It needs to exist.

```text
TASTE SPEC (one page, read by people and agents)

WE ALWAYS
- Lead with the answer, then the detail
- Show an empty state that teaches the next action
- Name the human who approved the output

WE NEVER
- Ship a first draft as final
- Add a setting when a good default will do
- Use a number we cannot source

EXAMPLES OF RIGHT   (5 links)
EXAMPLES OF ALMOST  (5 links, with one line on what is off)

KILL LIST THIS WEEK
- 1 item, removed, not deprioritized
```

The "examples of almost" line carries the most weight. Models and new teammates learn a standard fastest from near misses, because near misses show where the line is.

```mermaid
sequenceDiagram
    participant H as Human
    participant A as AI agent
    participant S as Taste spec
    participant M as Decision record
    H->>A: Ask for a draft
    A->>S: Read the standards
    A->>M: Read past approvals and rejections
    A-->>H: Return a draft
    H->>M: Approve, edit or reject, with a reason
    M-->>A: Next draft starts closer to the bar
```

That loop is the whole mechanism. The human stays in control of every decision. The system keeps the reasons. After six months, the record is something no competitor has and no model vendor can ship, because it is made of your decisions. This is the approve, edit or reject pattern that agent frameworks document as [human-in-the-loop](https://docs.langchain.com/oss/python/langchain/human-in-the-loop), and it is the practical bridge from taste to switching costs.

## 🏋️ How Do You Develop Taste? Three Reps

You develop taste the way you develop any judgment: repetition, feedback, and exposure to people who see differently. There is no shortcut, and it cannot be outsourced to a model. Three habits do most of the work, and each takes less than an hour a week.

| Rep | Practice | Why it works |
|---|---|---|
| Study | Pick one product you admire each week. List what its makers refused to build, its defaults, its empty states | You absorb other people's decisions as patterns |
| Ship in public | Show unfinished work to people who will be honest | Judgment is calibrated by other people |
| Kill | Remove one thing every week | Restraint is a muscle, and resistance is the signal |

```mermaid
graph TD
    S[Study one product a week] --> P[Ship your work in public]
    P --> K[Kill one thing a week]
    K --> J[Sharper judgment]
    J --> W[Write it into the taste spec]
    style S fill:#0d1b2a,color:#7dd3fc,stroke:#38bdf8,stroke-width:1.5px
    style P fill:#1a0d24,color:#ce93d8,stroke:#6a1b9a,stroke-width:1.5px
    style K fill:#1a0d0d,color:#ef9a9a,stroke:#b71c1c,stroke-width:1.5px
    style J fill:#0d1a0e,color:#86efac,stroke:#22c55e,stroke-width:1.5px
    style W fill:#2a1225,color:#ff8fa3,stroke:#ff2d60,stroke-width:2px
```

On the second rep, Carter leans on Hannah Arendt: ["For Arendt, taste is the exercise of judgment in the presence of others."](https://hsph.harvard.edu/news/essay-3-taste-and-the-end-of-scarcity/) Comedians do not write jokes only in their heads. They go up at an open mic and learn in real time which lines die. Your version is the friend who says the empty state feels sterile, or the user who says the AI feature feels bolted on.

On the third rep: expect to hate it. The thing you are cutting has time, thought and tokens invested in it. That resistance is the feeling of taste forming.

One more model helps here, from [a talk by Joel Torana on mapping ideas](https://www.youtube.com/watch?v=ASOYg71a_eA). Your experience is the macro view, the era you live in is the meso view, and the specific thing you make is the micro view where the two meet. Authority comes from that intersection, which is why two people with the same tools produce different work. If you like thinking in levels, our guide to [systems thinking](https://www.taskade.com/blog/systems-thinking-explained) goes deeper.

## 🧬 Where Does Taskade Genesis Fit?

Taskade Genesis fits at the bottom two layers of the stack: accumulated context and workflow depth. It cannot give you taste, and no product can. What it gives you is a place where your standards, your data and your decisions live together, so that what you build keeps what it learns. We call that Workspace DNA: Memory, Intelligence and Execution.

| Moat layer | What you need | How it works in Taskade |
|---|---|---|
| Accumulated context | A memory the product earns | Projects and databases hold the record, and Taskade EVE keeps project memory in the workspace |
| Workflow depth | Software that runs the work | [Automations](https://www.taskade.com/automate) with 100+ integrations execute the steps |
| Encoded taste | Standards agents can read | [AI agents](https://www.taskade.com/agents) take custom instructions and knowledge, including your taste spec |
| Accountability | A human gate | An automation can pause for a human approval before it acts |
| Owned distribution | A channel you control | Publish apps on your own domain (Pro and up) or share them in the [Taskade App Factory](https://www.taskade.com/apps) |

An honest note on our own position. Taskade also routes frontier models from top AI labs, and it also meters AI usage with credits. We rent intelligence like most of the industry does. Our bet is that the durable value sits in the workspace around the model: the memory, the agents and the automations that compound as a team works. The idea is covered in [how Workspace DNA works](https://www.taskade.com/blog/how-workspace-dna-works) and [code versus runtime](https://www.taskade.com/blog/code-vs-runtime).

A practical way to start: build the taste review board from this article. One table of decisions with a kill list, one agent that checks each draft against your written rules, and one automation that sends a weekly review. [Taskade Genesis](https://www.taskade.com/create) builds that from a single prompt, and Pro is $10 a month billed annually. If you would rather begin from a working example, clone one from [AI apps](https://www.taskade.com/ai/apps).

## ✅ The Defensibility Test: 10 Questions

A product is defensible when it passes most of these ten questions. Score one point for each yes. The test works for a startup, a side project or an internal tool, and it takes five minutes.

| # | Question | Moat it tests |
|---|---|---|
| 1 | If the model vendor shipped our core feature tomorrow, would customers stay? | Rented intelligence |
| 2 | If our main acquisition channel changed its rules, would we still grow? | Rented distribution |
| 3 | Does the product get better for a customer the longer they use it? | Accumulated context |
| 4 | Would leaving cost a customer more than exporting a file? | Switching costs |
| 5 | Does the product do the work, or only display it? | Workflow depth |
| 6 | Are our standards written where a new hire or an agent can read them? | Encoded taste |
| 7 | Does a named human approve what ships? | Accountability |
| 8 | Did we remove something this month? | Restraint |
| 9 | Do customers recommend us without being asked? | Trust |
| 10 | Can we say what we refuse to build? | Point of view |

| Score | Reading |
|---|---|
| 8 to 10 | You have a moat, and taste is feeding it |
| 5 to 7 | You have an edge. Write it down before it decays |
| 0 to 4 | You have a prototype. That is a fine place to start and a bad place to stop |

## 🔑 Key Takeaways

1. **Taste is scarce.** App releases rose 60% in a year while studies show AI-assisted output converging.
2. **Taste is not a moat by itself.** It is alpha: an edge that shrinks as the baseline rises.
3. **A moat must survive copying.** Buffett's 1986 picture and Helmer's seven powers still set the test.
4. **Rented advantage gets repriced.** monday.com rented distribution, and Perplexity rents intelligence.
5. **Five moats hold:** accumulated context, workflow depth, earned trust, owned distribution, encoded taste.
6. **Chat memory is portable. Working systems are not.** The switching cost is the record of your decisions.
7. **Build for two audiences.** Structure for agents, feel for humans.
8. **Convert taste on purpose.** Write the spec, gate on it, record decisions, feed them back, kill weekly.

A moat is something you defend. Taste is something you become, and then something you build into the walls. When everything can be made, what you choose not to make becomes the work.

Memory holds what you learned. Intelligence applies it. Execution ships it. ▲ ■ ●

## 💬 Frequently Asked Questions

### Is taste the new moat?

Not by itself. Taste is a filter that decides what gets built and shipped, and a filter can be copied once its output is visible. Taste becomes a moat only when a team writes it down, builds it into process, and compounds it with context that competitors cannot see.

### What does "taste is the new moat" mean?

The phrase means that when AI makes software cheap to produce, the scarce skill moves from building to choosing. Paul Graham wrote in February 2026 that when anyone can make anything, the big differentiator is what you choose to make. Critics answer that judgment is an edge that decays, not a wall.

### What is a moat in AI?

A moat in AI is a durable advantage that stops competitors from taking your customers even when they can copy your product. The candidates are accumulated context, workflow integration, trust, owned distribution and taste encoded as process. The model alone is rarely a moat because most companies rent it.

### Can AI learn taste?

Partly. If taste is a stable set of criteria, a model can learn to apply those criteria. What a model does not have is your context: your customers, your constraints and the decisions you already made. The durable part of taste is tied to private context and to accountability for the result.

### Does AI make everything look the same?

At the group level, yes. A 2024 Science Advances study found that stories written with AI help were more creative on average and more similar to each other. A 2025 meta-analysis found a positive effect on creativity (g = 0.27) and a large negative effect on diversity of ideas (g = -0.86).

### Is an AI wrapper a real business?

It can be, if the product owns something the model vendor does not: the customer relationship, the workflow, the data that accumulates with use, or the distribution. A product that only resells model output at a markup is exposed to vendor price changes and to the vendor shipping the same feature.

### What moats still work for AI startups in 2026?

Five hold up: accumulated context and memory, deep workflow integration, earned trust and brand, distribution you own, and taste encoded as process. Speed, feature lists and access to a frontier model do not hold, because every competitor can buy the same things.

### What is the difference between an MVP and a minimum lovable product?

A minimum viable product is the smallest version that proves a product works. A minimum lovable product is the smallest version people want to come back to. When anyone can ship a working MVP in a weekend, the question moves from "does it work" to "will anyone return."

### How do you develop taste?

Through repetition with feedback. Study one product you admire each week and list what its makers refused to build. Ship your work in public so other people calibrate your judgment. Remove one thing every week from a roadmap or a draft.

### How do you turn taste into a moat?

Write it down, enforce it, and compound it. Put your standards in a short spec that people and AI agents both read. Add a review gate with a named approver. Record every approval and rejection so the system learns your preferences. The written standard gives you process power, and the record gives you switching costs.

### Is distribution the only moat left?

Distribution matters more as products converge, but rented distribution is not a moat. monday.com spent 119% of its 2020 revenue on sales and marketing, then told investors in August 2025 that changes in Google search softened demand from small customers. Distribution defends a business when the company owns the channel.

## 🔗 Further Reading

### The abundance series
- [Creative Destruction Comes for SaaS](https://www.taskade.com/blog/creative-destruction-saas): Schumpeter's argument applied to software
- [The History of Lovable](https://www.taskade.com/blog/lovable-history): 200,000 projects a day
- [AI-Native vs AI-Bolted](https://www.taskade.com/blog/ai-native-vs-ai-bolted): why architecture shows through the surface

### Building with judgment
- [One-Person Companies](https://www.taskade.com/blog/one-person-companies): what a single operator can run with agents
- [AI Agents for Solopreneurs](https://www.taskade.com/blog/ai-agents-solopreneurs): the managed AI employee model
- [Best AI Second Brain Tools](https://www.taskade.com/blog/ai-second-brain-tools): where your context should live
- [How to Build an Approval Workflow](https://www.taskade.com/blog/build-approval-workflow): the human gate, step by step