The Surface Is The Stack

AI design is not a category

A founder generates a landing page in a prompt-to-app tool and ships it to a campaign URL. The form silently routes nowhere. The page looks production-ready; the integration is missing. A different founder turns a long memo into a 14-card investor deck in a slide generator, exports cleanly, and presents it. By the seventh deck of the quarter, every pitch looks identical because the underlying templates are the same. A third founder vibe-codes an internal records tool, connects it to live customer data, and discovers two months later that nobody can update the schema without breaking a downstream invoice flow.

Three different surfaces, three different failure modes, one category mistake. AI design hasn't become one tool; it has become a stack of single-surface tools that share a marketing label.

The same AI that drafts a clean campaign page is reckless for an accessibility-bound signup flow. The same generator that nails a first logo concept fails across dark mode, app icons, and brand consistency at scale. The prompt-to-app tool that produces a useful product demo creates a hidden maintenance burden the moment real customer data touches it. The mistake is not the tool. The mistake is the absence of a surface map.

The right question is not which AI design tool is best. The market answers that question with a different tool every quarter, and the answer keeps changing. The right question is sharper: what do I need to ship this week, what tool gets me to a credible version fastest, and where does the output stop being safe to ship without a designer or an engineer?

Seven surfaces matter for an operator without a designer on staff: landing pages, slide decks, brand assets, prototypes and mockups, working app prototypes, internal tool UIs, and documents with diagrams. Each has a different tool category that earns the slot today, a different ceiling on what ships clean, and a different return-point for the specialist. Treating them as one category is how teams waste budget on overlapping subscriptions and end up with assets that look right and fail in motion. The map below names the primary tool per surface and the adjacent fits that look like overlap and usually are not the right call.

The seven surfaces

The surface decides the tool. Below, each surface gets the same treatment: the tool category that earns the slot today, what ships clean, where the ceiling sits, and the action plan for week one. The action plans assume you're starting from zero on that surface; if a tool is already in use, the answer is to keep using it and skip the section.

Landing pages and marketing sites

Campaign pages, waitlists, product pages, simple CMS sites. The slot belongs to AI-assisted site builders, split by intent: Framer for speed and single pages, Webflow for production marketing sites with CMS or a real design system. Relume sits upstream of Webflow when sitemap structure is the bottleneck. Free tiers are sandboxes; custom-domain publishing requires paid plans, with Webflow's Site plans listed at $14, $23, and $39 per month billed yearly across Basic, CMS, and Business.

What ships clean: a campaign page, waitlist, product explainer, or webinar page, with a structurally useful first draft and a responsive baseline. The ceiling appears at site systems: navigation logic, CMS taxonomy, SEO migration, accessibility, localization, form routing, CRM integration, and analytics events. Neither tool removes your review on mobile, performance, accessibility, or claims.

If you start this week, open Framer for a single campaign page. Generate from a short brief, replace the copy, connect the form, publish to a staging URL, test on mobile. Ship if the form routes correctly and a non-internal reviewer can explain the offer in 10 seconds. Don't reach for Lovable, Bolt, or v0 unless you're shipping a product prototype. A marketing page needs editing, publishing, and analytics more than a generated React app.

Slide decks and pitch materials

Investor decks, sales decks, internal strategy decks, webinar decks. The slot belongs to AI presentation tools: Gamma for speed and the long-memo-to-deck workflow, Beautiful.ai for teams with brand-controlled slide governance. Claude Design now generates decks from prompts and exports to PPTX, PDF, or Canva, but as a research-preview product on paid Claude plans it is not yet the team default; use it for ideation.

What ships clean: sales decks, webinar outlines, internal strategy decks, founder updates, product education decks. Gamma is strongest when you have a long memo, transcript, or outline. The ceiling shows up on investor and executive sales decks: AI decks over-explain, include too many cards, and fall back on generic iconography. Brand fidelity also breaks under repetition. The first deck looks clean; the seventh looks like a template farm.

If you start this week, use Gamma to turn an existing memo or pitch outline into 10 to 12 cards rather than starting from a blank prompt. Replace generic visuals, cut 30% of the cards, export to PPTX or Google Slides, and present internally once. Move to Beautiful.ai only when slide governance, locked layouts, or team analytics become the bottleneck. Don't pay for two deck tools in month one.

Brand assets and recurring marketing creative

Logos, social posts, banners, ad variants, campaign visuals, thumbnails, podcast covers, lead-magnet covers. The slot belongs to AI-assisted editable design suites: Canva by default, Adobe Express plus Firefly when the team already lives in Adobe. Canva's advantage is not that it generates the best image; it's that the output is editable by non-designers and saves as a reusable template. Looka generates a placeholder identity for a company with no logo at all, but that is a starting system, not a brand.

What ships clean: social posts, event banners, ad variants, podcast covers, simple PDFs, screenshots inside frames, basic brand kits. The ceiling is brand fidelity at scale. AI design tools produce clean-looking output that has no underlying brand logic. Logos are the highest-risk asset: the first mark looks plausible, the system around it fails across sizes, dark mode, icon-only use, and legal distinctiveness. Ad creative also hits a claims and compliance ceiling no design tool can resolve.

If you start this week, open Canva. Define logo, two fonts, three colors, and three template types: social post, one-page PDF, banner. Use AI to generate variants, then save the final choices as templates. Ship one post, one PDF, and one banner by end of week. Move to paid Canva only when export, brand kit, team reuse, or AI allowance limits become the bottleneck. Don't buy a dedicated AI logo tool unless the company has no usable identity and only needs a placeholder.

Prototypes and mockups

Product concepts, app flows, wireframes, click-through demos, design exploration. The slot belongs to prompt-to-interface and AI prototyping tools: Claude Design for teams already paying for Claude, Figma Make when designers or engineers will inherit the work. Uizard and Visily are simpler UI-mockup options when you only need a single screen rather than a system. Figma Professional starts at $16 per month with higher AI credits than Starter; Claude Pro at $17 per month annual billing unlocks Claude Design's research preview.

What ships clean: product concepts, onboarding flows, dashboard mockups, mobile-first feature mockups, clickable prototypes for customer interviews. The output is something users can react to, not just an aesthetic. The ceiling is production logic. A prototype can imply functionality that does not exist, ignore edge cases, mishandle accessibility, and hide data-state complexity. Signup flows, payments, permissions, and admin workflows are especially vulnerable.

If you start this week, generate one clickable prototype for a single user job, such as "approve an invoice," "compare two plans," or "create a campaign." Use Claude Design or Figma Make depending on where the work will go next. Ship a testable prototype, not production code. Move on when three target users complete the task without you narrating the flow.

Working app prototypes

Demo apps, founder MVPs, single-purpose tools, fake-door product tests, front-end concept handoffs. The slot belongs to prompt-to-app builders: Lovable, Bolt, and v0. They are powerful, and they are not the default design stack. They earn their slot at the edge, where the demo has to behave enough like software to test the idea. Lovable's Pro tier sits around $25 per month; Bolt builds with team components and design-system imports; v0's natural handoff is to a React or Vercel engineering workflow at $20 per month plus usage.

What ships clean: product demos, founder prototypes, design explorations with real interactions, lightweight proof-of-concepts. The ceiling is production. AI coding tools create hidden fragility: authentication mistakes, state bugs, insecure APIs, weak data models, inaccessible UI, and dependency sprawl. A polished demo can hide a bad system.

If you start this week, use Lovable, Bolt, or v0 only after you have a specific interaction to test, and build it with fake or sandbox data, not production customer data, payments, or any internal system of record. Move to engineering review the moment any real user, real data, or paid workflow enters the prototype.

Internal tool UIs

Admin dashboards, ops tools, approval queues, partner portals, customer portals, spreadsheet replacements. The slot belongs to governed internal app builders: Retool for tools touching production systems, Softr for portals, Glide for spreadsheet-driven apps. This is the surface where teams most reliably overestimate AI. Internal tools touch real data, customer records, operational decisions, approvals, payments, or employee workflows. A pretty generated UI is not enough; the tool needs permissions, data integrity, auditability, and maintainability. Softr's plans run from $49 to $269 per month; Retool and Glide depend on builder mix, data sources, and internal-user count.

What ships clean: read-only dashboards, approval queues, simple CRM views, project intake forms, customer portals, spreadsheet replacements. The ceiling appears at security, permissions, data mutation, workflow complexity, and audit trails. Don't build tools touching payment authorization, regulated records, HR decisions, or customer-facing account state without engineering and compliance review.

If you start this week, pick one internal workflow currently run in a spreadsheet. Make the first version read-only or low-risk. For production data, use Retool; for a partner or customer portal, use Softr; for a spreadsheet-driven workflow, use Glide. Ship a read-only dashboard or intake tool by end of week. Move on only when permissions are correct and the daily user prefers it. Don't vibe-code the workflow in Lovable or Bolt; use governed builders for operations.

Documents with diagrams or charts

Strategy memos, SOPs, GTM plans, explainers, process docs, reports, decks with visual ideas. The slot belongs to text-to-diagram and deterministic diagram tools: Napkin AI for visual explanation in docs and decks, Mermaid for diagrams that need to be deterministic, versionable, or embedded in technical docs. Claude and ChatGPT cover charts when the source is a spreadsheet or dataset. Napkin's free plan ships 500 AI credits per week with watermarking; Plus at $9 per month adds PPT and SVG export and removes branding. Mermaid pricing tiers free use against paid collaboration features.

What ships clean: strategy maps, workflow diagrams, process visuals, funnel diagrams, comparison charts, simple data charts, visual explainers embedded in docs or decks. The ceiling is accuracy and editability. A diagram that looks clean but misrepresents the process is worse than text. For charts, the ceiling is data correctness: axis choices, aggregation, missing values, implied causality. AI can make the chart; you own the interpretation.

If you start this week, take one existing memo or SOP. Generate three visuals in Napkin: one flow, one comparison, one system map. Export one to a deck and one to a doc. Use Mermaid only for diagrams that will live in engineering docs or need text-based version control. Move on when the visual reduces explanation time in a real meeting. Don't use general image generators for business diagrams; they create pictures of diagrams, not reliable diagrams.

Where the designer or engineer comes back in

The boundary is not "when it looks bad." The boundary is when the work carries system risk. AI tools produce default-pretty across every surface in the stack. Default-pretty is fine for social posts and a campaign page. It fails at the surfaces where taste, brand, accessibility, or production logic is the differentiator.

Designer returns when

Brand-system fidelity matters across many surfaces. When the company needs logo, typography, ad templates, event graphics, web pages, decks, social assets, product UI, and partner materials to feel coherent, AI-generated assets cannot hold the system. A designer cleans or creates it.

The output sits in high-trust sales: investor decks, enterprise sales decks, category-defining landing pages, flagship launch pages. AI can draft any of them; it shouldn't finish them.

Accessibility is part of what you ship. Complex signup flows, healthcare or finance pages, public-sector workflows, internal tools used by diverse staff. AI outputs often look modern while missing contrast, keyboard flow, screen-reader logic, error handling, and focus states.

Multi-locale or multi-market use begins. AI produces variants quickly, but brand, language, layout, and claims break across markets. A designer and a localization reviewer return when the company scales beyond one language or one region.

Engineer returns when

Real data enters the build whenever a workflow touches customer data, internal financial data, production databases, employee records, or anything regulated.

Authentication, permissions, or payments appear in the build, and prompt-to-app tools generate plausible flows that aren't safe.

The app has more than one meaningful state: loading, empty, error, success, permission denied, partial data, stale data, duplicate submission, rollback, and audit. AI prototypes underbuild every one of them.

The internal tool becomes a system of record. The moment a dashboard writes to source systems, triggers workflows, approves payments, updates customer records, or sends notifications, engineering owns the architecture.

The generated code must be maintained, which is where prototypes break: Lovable, Bolt, and v0 ship useful demos, but somebody has to own the dependencies, tests, security updates, deployment, logging, and incident response after the demo.

Cost calculus and coexistence

Free tiers are useful for evaluation, not for running the business. Custom domains, serious exports, branding removal, AI usage allowances, and team features sit on paid plans across the stack. The first paid stack that earns its keep for a solo founder is small: one general AI assistant (ChatGPT Plus or Claude Pro at $17 to $20 per month), one editable design suite (Canva paid only when the free tier blocks the workflow), one deck and doc tool (Gamma paid only when branding or exports require it), one diagram tool (Napkin Plus at $9 per month), and one site publisher (Framer or Webflow at $14 to $39 per month). That's five subscriptions, one per surface, totaling around $80 to $120 per month.

The mid-paid tier sits between $40 and $150 per user per month: Beautiful.ai Team, Webflow CMS or Business, Adobe Express plus Firefly, Softr Professional. It earns when deck governance, CMS capacity, Adobe-native workflow, or a real portal is the bottleneck. It doesn't earn when the lower tier already covers the surface.

The designer-replacement tier above $200 per month is the one to slow down on. Claude Max starts at $100 per month and rises to $200 per month for higher usage. Enterprise internal-tool stacks push higher with seats and AI credits. At that spend, compare the stack against actual specialist time. The clean spend is template production, deck drafting, diagramming, campaign variants, and simple site publishing. The weak spend is buying taste, product architecture, brand strategy, accessibility, or engineering safety through more subscriptions.

PowerPoint and Google Slides do not disappear; many buyers and investors still expect them as the final deliverable, so deck-tool exports are load-bearing. Microsoft 365 Copilot and Google Workspace Gemini are baseline utilities for document-native drafting, not best-in-class production. Figma stays when handoff matters; Webflow stays if the site already lives there; Canva coexists with everything because it is the asset factory; Adobe stays when the team already runs Adobe.

Pitfalls and anti-patterns

Treating AI design as one category

The same prompt that ships a campaign page is reckless for an investor deck. The same generator that nails a logo concept fails across the brand system. The category is fragmented; the tool list is not. Teams that reach for "the AI design tool" buy overlap and ship nothing.

Subscribing before mapping the surface

Buying three tools before defining the surface map produces overlap, four learning curves, and nothing shipped. Map first, subscribe per surface, learn one tool deeply before adding the next.

Letting a prototype become production

Letting a prototype become production is the most expensive failure in this stack. Lovable, Bolt, and v0 produce useful demos but not maintenance-grade systems, and the moment real customer data, payment flow, or customer-facing logic enters the prototype, an engineer owns it or it owns you. "It's just a prototype" is how brittle systems become production by momentum.

Optimizing for AI taste rather than brand fidelity

AI tools produce default-pretty. Default-pretty is fine for social posts and lead magnets. It fails on flagship sites, investor decks, and brand systems where taste is the differentiator. Use AI to draft; bring a designer for the surfaces that carry the brand.

What to validate before paying for the stack

• Map the surfaces you actually ship. If the company won't produce slide decks for two more quarters, don't pay for a deck tool.
• Run each tool's free tier through one real piece of work, not a generic demo prompt. The free tier reveals whether the export, the brand context, and the editing flow match your actual workflow.
• Confirm the export path matches where the work lives: decks need PPTX or Google Slides, pages need a custom domain, diagrams need PPT, SVG, or PDF. A tool that can't export to where the file actually lives isn't a tool.
• Set the diminishing-return threshold before the third paid subscription. If the lower tier already covers a surface, the next subscription is duplicate spend.
• Confirm someone on the team can act on the output. AI-generated assets only matter if a person on the team can edit them, route them, publish them, and replace them when the brand or claims change.

There is no AI design tool. The surface is the stack.