“AI-powered” hides six different products under one marketing label. A pitch deck that reads “AI-powered” is silent on whether the system is one model call per request or a fine-tuned multi-modal agent with a vector index and a tool-call schema. Those systems differ by an order of magnitude in cost, calendar time, and operational burden. A founder who builds the wrong shape in week one ships an MVP that cannot do what the pitch promised, and rebuilding from the wrong primitive costs more than starting from a blank repo.
This piece decomposes “AI-powered” into six technical commitments. It is the architecture companion to the idea validation playbook, a chapter inside the broader idea-to-product manifesto. It pairs with the AI feasibility check and the AI capability map. Feasibility asks can-the-frontier-do-this; the capability map compresses the answer for a board; this piece answers the architecture question between them — if it is feasible, in what shape do you ship it.
Decision Scope
An editorial decomposition for non-engineer founders making a first-architecture call. Numeric ranges are 2026 planning heuristics. Validate against real inputs, real traffic, and current vendor pricing. Not legal, financial, security, or accounting advice.
Why “AI-powered” became a category-confusion phrase
Three shifts turned a useful descriptor into proposal fog. The cost of one LLM call collapsed; one API key produces a working demo in an afternoon, making “we use AI” technically true for almost any product. The tooling ecosystem expanded faster than founder vocabulary — every vendor pitched their layer (vector DBs, agent frameworks, fine-tuning platforms) as the foundation. And the gap between “calls a model” and “operates a model in production” widened.
“AI-powered” now describes six shapes that share almost nothing operationally — different cost, latency, security surface, evals, team. A founder who has not chosen consciously has chosen by default. The vendor default is the most expensive shape that justifies the retainer. Name the six, pick one, defend the choice.
Shape 1: Prompt-only
A single model call per user request. No state, retrieval, tools, or fine-tuning. The smallest valid LLM system.
Requires. A provider account. A versioned prompt. A 30–100 input eval suite. A cost cap per request and per user. Basic logging.
Right for MVP-1 when the task is single-shot, stateless, and bounded — drafting an outbound email, classifying a support ticket, summarizing a transcript, extracting fields from a clause.
Wrong for MVP-1 when the answer depends on data the model has not seen, the task requires coordinated steps with state changes, or the user expects memory of prior turns.
Eval shape. Flat suite of 30–100 examples, pass/fail or rubric-graded. CI gates merges on the score.
Cost shape. A Claude Sonnet 4.6 or GPT-5 call at 2026 prices lands well under a cent for short inputs. A surprising share of B2B SaaS AI features in 2026 are prompt-only — the ones that ship on time and hit margin.
Shape 2: Multi-turn dialog
A conversation thread with state. The model sees prior turns as context. The product is shaped like a chat.
Requires. Shape 1, plus a thread store, a context-window strategy (truncate, summarize, or semantic-rank old turns), drift guardrails, turn-level evaluation. Latency budgets tighten — users feel chat latency more sharply than batch.
Right for MVP-1 when the value is conversational — a coaching tool, a co-pilot, a tutoring system. The user would naturally use back-and-forth to refine the answer.
Wrong for MVP-1 when the task is single-shot but the founder added chat because “chat is what AI looks like.” A chat interface around a classification task adds latency, hides the eval, and trains users to expect fluency the primitive does not have. A form is often the right interface.
Eval shape. Five to ten scripted multi-turn scenarios with checkpoints — “by turn three the agent should have asked for the user’s role.”
Cost shape. 3–10x prompt-only per session because context grows with each turn.
Shape 3: RAG (retrieval-augmented generation)
The model receives, alongside the user prompt, a passage retrieved from your data. The passage is the answer source; the model is the reasoner. RAG is the shape most often miscast as the default — vector-DB vendors have trained the market to assume retrieval is mandatory.
Requires. Embeddings of your corpus. A vector index (Pinecone, Weaviate, pgvector, Turbopuffer, or in-process). A chunking strategy. A retrieval-quality eval. A freshness story.
Right for MVP-1 when value depends on your private data — a sales assistant grounded in product docs, a legal-review tool over a contract corpus, a customer FAQ over support history.
Wrong for MVP-1 when the corpus fits in the prompt directly, the corpus is unselected (retrieval over a junk drawer returns junk), or the source is structured — a SQL query beats vector search for tabular questions.
Eval shape. Two layers. Retrieval scored on recall@k against labeled question-source triples. Generation scored on faithfulness. Founders who only score the final answer miss which layer is broken.
Cost shape. 2–4x prompt-only on tokens, plus a vector index bill that scales with corpus size.
Shape 4: Tool-using agent
The model calls external APIs and uses the results to compose its next step. Model is the planner; tools are the hands. A loop runs until the model returns a final answer.
Requires. A function-calling schema with typed signatures and return schemas. Per-tenant authentication. A step budget. Trace logging of every thought, tool call, and observation. A halting criterion. Replay-able observability.
Right for MVP-1 when value is the action, not the answer. A booking agent that reserves the flight, a research agent pulling from five sources, an outbound agent that finds, qualifies, drafts.
Wrong for MVP-1 when the same outcome could be reached with a deterministic workflow plus one model call per branch. Agents shine when the path is unknown at runtime; when the path is known, a state machine with model calls inside is shorter to build, cheaper to run, easier to debug. The right primitive is often model-in-workflow, not agent-in-loop.
Eval shape. Trajectory eval over 20–50 tasks with known correct trajectories. Score is end-to-end completion: goal reached, steps, cost, no unsafe tools invoked.
Cost shape. 5–20x prompt-only per request. Cost variance is the hazard — identical-looking requests can differ by an order of magnitude in tokens. Cap per-request budget in code or unit economics melt under one bad day.
Shape 5: Fine-tuned model
A base model trained further on your data. The output is a checkpoint that is yours — same weights, different behavior. Lightweight LoRA, full-parameter tunes through provider APIs, or instruction-tuning a small open model.
Requires. A labeled dataset — a few thousand examples for instruction-tuning, far more for behavior shifts. A training pipeline. A held-out eval set. A retrain cadence. Versioned artifacts.
Right for MVP-1 when the domain vocabulary defeats the frontier (specialist medical or legal sub-domains, under-served languages), the latency budget rules out the frontier, or compliance requires the model to live where the frontier does not.
Wrong for MVP-1 when almost any other condition holds. A good prompt plus few-shot examples closes most of the gap. RAG handles “model doesn’t know my data” better. Founders who fine-tune at week one inherit a model they cannot improve, because they have not learned the failure modes.
Eval shape. Hold-out against a set training never saw. The fine-tune must beat the base by a margin worth the cost.
Cost shape. Forward-loaded. Labeled data dominates — typically several thousand dollars of expert labeling. Model at 100x current traffic before committing.
Shape 6: Multi-modal
The model sees, hears, or speaks. Inputs or outputs include images, audio, or video. The 2026 frontier — Gemini 2.5 Pro, GPT-5, Claude Opus 4.8 — accepts image inputs natively; audio and video are first-class.
Requires. A model that supports the modality. A preprocessing pipeline (image resizing, audio chunking, frame extraction). Storage for large inputs. A latency budget that absorbs the modality overhead. Modality-specific privacy handling.
Right for MVP-1 when the input is naturally non-text — a furniture-arrangement assistant that reads room photos, a maintenance app diagnosing equipment from a photo, a transcription product, an accessibility tool that describes images.
Wrong for MVP-1 when the founder added a modality because “vision is the future.” If a structured form delivers the same value as a photo upload, the text version ships in a quarter of the time at a third of the cost. Add the modality in v2.
Eval shape. Modality-specific. Image tasks need labeled outputs that account for visual ambiguity. Audio needs ground-truth transcripts.
Cost shape. Higher per-request than text-only. Latency is the bigger constraint: a multi-modal call can take several seconds before the first text token streams.
The six shapes side by side
| Shape | Right for MVP-1 | Wrong for MVP-1 | Eval | Cost vs. prompt-only |
|---|---|---|---|---|
| Prompt-only | Single-shot, stateless task | Needs private data or external actions | Flat 30–100 input suite | 1x |
| Multi-turn dialog | Value is conversational | A form would do | 5–10 scripted scenarios | 3–10x |
| RAG | Answer must come from your corpus | Corpus fits in prompt or is structured | Recall + faithfulness | 2–4x + index |
| Tool-using agent | Action with an unknown path | Path is known; workflow would do | Trajectory eval over 20–50 tasks | 5–20x |
| Fine-tuned model | Domain, latency, or compliance forces it | Prompt or RAG closes the gap | Hold-out vs. base | Forward-loaded |
| Multi-modal | Non-text input is the value | Text would deliver the same value | Modality-specific set | 1.5–3x + latency |
Decoding your own pitch: four founder questions
Four questions resolve which shape MVP-1 needs.
1. What does the user hand the system? A sentence of text means Shape 1 or 2. “A question plus a reference to data we own” is Shape 3. “A goal and the expectation the system goes and does something” is Shape 4. A photo, audio, or video is Shape 6.
2. Does the system take an action in the world? If yes, Shape 4 enters — but only if the path is unknown at runtime. If the path is known, the architecture is a workflow with model calls at branch points, not an agentic loop. Most founders who think they need an agent need a workflow.
3. Does the answer depend on data the model has not seen? If yes, Shape 3 territory. How much data and how often it changes decides whether a small static corpus fits inline or a true RAG index is required.
4. Is fine-tuning a 2027 question or a 2026 question? Almost always 2027. Ship Shape 1 or 3, run real users for a quarter, collect actual failures, then decide whether a fine-tune closes a gap prompt or retrieval cannot.
The questions resolve in order. The first that resolves is the shape.
The default position
The default shape for MVP-1 in 2026 is Shape 1. Prompt-only. One well-engineered prompt, a 50-row eval, a cost cap, a logging pipe. The smallest valid LLM system — and it covers more viable B2B SaaS AI features than the agency ecosystem admits.
Escalate when data forces it. Shape 2 when value is conversational. Shape 3 when the answer must come from a corpus you own. Shape 4 when the user pays for an action and the path is unknown. Shape 5 when domain, latency, or compliance leaves no option. Shape 6 when input is fundamentally non-text.
The decoding move protects the founder twice. At the pitch — “we are AI-powered” becomes “we are a prompt-only system with a 50-input eval and a per-user cost cap,” and investors trained by the 2024–25 noise weigh that higher. At the build — the team ships the smallest valid system, learns where it breaks, escalates only when data demands. The difference between an MVP shipped in eight weeks and one still in architecture review at month five.
The companion piece on decoding production-ready in AI agency proposals applies the same move downstream. Refuse the marketing phrase, decompose the commitment.
Before week one, decode your own pitch. Build the shape that survives the question, not the shape that wins the demo.
MVP scoping worksheet — pin the six shapes against your pitch in 30 minutes. Get the AI MVP scoping worksheet.
FAQ
Is “AI-powered” still acceptable on a pitch deck in 2026?
As a category marker, yes; as a technical claim, no. A 2026 reader expects the next slide to name the shape — “prompt-only with a 50-row eval” or “RAG over our 12k-doc corpus.” Leaving it undecomposed signals the architecture has not been decided.
Which shape is cheapest to ship?
Shape 1. A well-engineered prompt with a small eval suite ships in days. Multi-turn and RAG add weeks. Agents and fine-tuning add months.
Which shape do most AI agencies push by default?
The most expensive shape the buyer will pay for. Counter-move: ask the agency to defend the choice against the next-cheaper alternative. A generic defense means the shape is over-specified.
When does combining shapes make sense?
Almost any non-trivial production system combines two or more. Name each, ship the smallest combination for MVP-1, treat every extra shape as a later decision.
How does this differ from the AI feasibility check?
Feasibility answers “can the frontier do this at all.” This piece answers “if yes, in what shape.” Feasibility is go / no-go; shape selection is the architecture decision that follows.
What about no-code tools like Lovable, v0, or Bolt?
They almost always produce Shape 1 under the hood — a templated prompt-only app with auth, payments, and a UI bolted on. The misuse is asking a no-code tool to ship Shape 3 or Shape 4 — the primitive is not there.
How much labeled data does a fine-tune need in 2026?
A few thousand examples is the floor for instruction-tuning. Behavior shifts rise into the tens of thousands. Fine-tunes that meaningfully beat a well-engineered prompt over the same data are uncommon for general SaaS use cases.
Does RAG always need a vector database?
No. A small static corpus that fits in the context window can be retrieved inline. A medium corpus with predictable queries can use full-text search. Vector search shines when the corpus is large, queries diverse, latency tight.
Which shape is the riskiest to underestimate?
Tool-using agents. Looks small in a demo, grows by orders of magnitude in production — cost variance, security surface, multi-step failure modes. The most common pattern behind a project that ships a brilliant demo and dies in month one.
Dirk Jan van Veen, PhD
