Last lesson drew the line: AI makes building cheaper, but it does not create demand. This lesson follows the money that line unlocks. When the cost of building and operating falls, software for a narrower workflow can pay for itself — work that used to need a huge engineering organization first.

Why narrow got viable

The old math was brutal: a product for a few hundred users had to carry the same engineering payroll as a product for a few hundred thousand. So vendors built generic software and told every niche to adapt. The analyst, the property manager, the parish events coordinator — all squeezed into the same grey forms.

The new math: a small team, with AI-assisted building and managed infrastructure, can build and operate software for one narrower workflow without staffing every specialty first. Hosting, auth, scaffolding, tests, and docs — the fixed costs that killed small products — are cheaper or rented. What remains is the part that was always the real work: understanding the workflow deeply enough to be worth switching to.

That last clause matters. The opportunity is not "small market equals easy." It is a focused problem where generic software is clumsy — where the vocabulary, data, timing, and failure cost differ enough that a tailored tool wins.

Five terms, plainly

• Vertical: one industry or profession (independent insurance adjusters, dental labs, local landlords).
• Niche: a slice small enough that the giants ignore it (adjusters who handle storm claims, not all adjusters).
• Workflow: the actual sequence of steps someone repeats — inputs, decisions, outputs, handoffs — not the job title.
• Domain context: the vocabulary, data sources, rules, timing, and failure modes that make this workflow *this* workflow.
• Willingness to switch: whether the pain is sharp and frequent enough that someone changes behavior, pays, and tolerates a v1. Pain × frequency beats market size.

A niche earns its keep when a painful recurring workflow meets a team small enough to serve it profitably. Both halves are required.

Where the niches live

Look for recurring awkward jobs — work people do with spreadsheets, email threads, and quiet dread:

• Analyst research for one kind of analyst. Not "research AI" — the Friday workflow of a junior credit analyst assembling the same packet from the same five sources, checking the same three ratios, dreading the one manual copy step.
• Property-management follow-up for one local market. Not "proptech" — chasing late maintenance confirmations across twelve buildings where the failure cost is a flooded unit and the vocabulary is work orders, not tickets.
• Compliance review for one document type. Not "legal AI" — checking one certificate format against one regulator's checklist, where a miss costs a fine and the timing is always end-of-quarter.
• Event coordination for one community. Not "a social platform" — the Neighborhood Events pattern: one organizer, recurring listings, ownership of who edits what, no feed, no network-effect fantasy.
• Editorial operations for one publisher. Not "content AI" — the weekly pipeline of assignments, drafts, fact-checks, and corrections, where provenance and correction history are the product.

Each example is narrow on purpose. The narrower the workflow, the more the generic tool annoys — and the more a tailored default, adapter, or check feels like relief.

The false niche

Beware the cosmetic niche: changing colors, copy, or the hero image for a group while the workflow underneath stays generic. "Built for dentists!" on top of a generic scheduler is not a niche product. It fails the moment the user hits the step that makes their job distinct — the insurance code, the recall timing, the no-show cost — and finds nothing there.

Test every niche claim against five questions:

If you cannot answer all five, you have a landing page, not a niche.

The domain-context advantage

Here is how depth becomes product. A team that understands the workflow ships four things a generic wrapper cannot copy by rewording a prompt:

1. Better prompts and files. Markdown instructions and JSON handoffs written in the trade's vocabulary, encoding its checks — the Part II and Part X machinery applied to one job. 2. Better data adapters. Connectors, normalizers, and validators for the actual sources: the county records format, the vendor CSV with the weird date column, the regulator's PDF layout. 3. Sensible defaults. The report structure, the reminder schedule, the approval threshold the user would otherwise configure for an hour. Defaults are respect. 4. Useful quality checks. Validations tied to real failure cost: "this certificate number pattern fails every March" beats "please review the output carefully."

Each one is small. Together they make the application feel *built for that user* rather than generic AI pasted onto a form. That feeling is willingness to switch, made tangible.

Exercise: observe one awkward job

Do not brainstorm features. Find one person doing one recurring awkward job — a coworker, a friend, a user you can reach. Watch or interview them for 30 minutes and write a NICHE-WORKFLOW.md:

# NICHE-WORKFLOW.md

## The person and the job
- Who:
- Recurring job and frequency:

## Current workaround
- Steps they take today:

## Time / quality cost
- Minutes per recurrence, error rate, dread factor:

## Existing alternatives (and why they stay clumsy)
- Generic tool 1 + misfit:
- Generic tool 2 + misfit:

## Narrowest useful outcome
- One finished output for one recurrence:

## Niche test (data / vocab / timing / failure / default)
- ...

Finish line: a NICHE-WORKFLOW.md grounded in one real behavior, with the narrowest useful outcome stated as a finishable output.

Verify fast: read the "narrowest useful outcome" to the person you observed. If they say "yes, I would use that Friday," you have a niche. If they say "maybe, if it also did…," you have a brainstorm. Common failure: writing the interview from imagination. One real conversation beats ten personas.

Check your understanding

1. Why does cheaper building make *narrower* products viable, specifically? 2. What is the difference between a vertical and a workflow? 3. Give an example of a false niche and explain which of the five questions it fails. 4. Name two domain-context advantages and how each shows up in the product.

Next

You have a niche worth serving. But an AI capability is not yet a product. Next, Lesson 55.3 shows why the product is the workflow, not the model call — and how to map every non-model step that makes the capability dependable.