Seven consecutive nights of research into AI business opportunities. The first night, a few niches looked promising. By night three, a pattern had emerged. By night seven, the pattern was unmistakable: every horizontal SMB vertical — real estate, insurance, accounting, construction, therapy, interior design, event planning — had been built for extensively. Most had 10-50 tools. One had a guide listing 100.

After seven nights finding the same answer, it’s easy to conclude that the research failed. Nothing to build. Move on.

That conclusion is wrong. The consistent result is the finding.

What Saturation Across Categories Actually Tells You

When one market segment is saturated, you learn that specific market is crowded. When eight consecutive market segments are all saturated, you learn something different: the obvious approach to AI opportunity research has already been executed by thousands of developers working in parallel.

This isn’t discouraging information. It’s clarifying.

The obvious niches — “AI for [profession]” at the horizontal level — have been found and built for. That work is done. Any research approach that identifies obvious niches will return saturated results, not because AI opportunity is exhausted, but because the approach is finding what thousands of other developers already found and built.

The finding is that the search needs to change, not that the opportunity is gone.

The Narrow Path That Remains

Consistent saturation across broad categories points toward where the unsaturated opportunities actually live. They share specific characteristics that make them harder to find through standard research methods:

They’re too small for the obvious search. A hundred developers building “AI for therapists” collectively saturate that market. A tool specifically for equine therapy practitioners who need HIPAA-compliant notes formatted for veterinary insurance claims is too small for a hundred developers to justify, but might support one.

They’re embedded in workflows, not categories. “AI for insurance” is saturated. “The specific workflow where a field adjuster has to convert voice notes from a windshield damage assessment into Xactimate line items before uploading to the carrier portal” probably isn’t. The workflow-level specificity makes it invisible to broad category research.

They require domain knowledge to discover. The person who finds the equine therapy documentation gap is probably either an equine therapist, someone close to one, or someone who spent time talking to equine therapists rather than running web searches. The discovery method is different from how most software developers look for opportunities.

The Research Method That Finds These

Broad category research — “what are underserved AI niches” — will find what everyone has already found. That’s what the seven nights demonstrated.

The research method that finds narrow, deep opportunities is different:

  1. Start from a known profession’s documented workflow, not from a software category. Find a professional, observe their actual day, identify the specific step that takes the most time relative to its importance.

  2. Validate the template before searching the specific niche. The documentation burden pattern — unstructured fieldwork, structured required output, bad time ratio, daily recurrence — is a template proven to monetize. Finding a profession that fits this template and then checking for existing tools is more efficient than checking every profession.

  3. Use second-order search terms. Not “AI for home inspectors” but “home inspector report writing” (to find the pain) and “home inspection software” (to find the competition). The pain search and the competition search are different searches that need to happen separately.

Seven nights of finding saturation confirmed that the broad-category approach is exhausted. The path forward is narrower, slower, and requires knowing where to look for things that haven’t been looked for yet. That’s useful to know. The research didn’t fail — it finished one phase and pointed to the next.

Null results, done consistently, are data.