I built a scanner to find a SaaS worth cloning. Its best answer was “don’t.”

The claim: one person with AI agents can now clone legacy SaaS

Everyone has the same take: AI collapsed the cost of building software, so the fat, neglected vertical SaaS layer is finally disruptable by one engineer and a coding agent.

Fine. Which company, specifically, should that person clone starting Monday? I built the list, scored every company on it, then scraped all of them to check the one thing the argument assumes.

The answer was none of them, and the reason is better than a yes would have been. Every number here comes out of a scan I can re-run, and every company is named.

Where to look: a holdco that owns ~1,000 neglected products

Constellation Software has spent thirty years buying small vertical software businesses — dental labs, municipal utility billing, tutoring scheduling, funeral homes, transit fare collection — and holding them forever. Hold the product steady, raise prices annually, optimize for retention. Customers stay because it's the only serious software in their industry.

A thousand aging products in a thousand niches, and nobody has the list: published portfolio maps top out around 262 companies, Tracxn has 31, PitchBook 89. If the list is hard to build, nobody has picked it over. And these aren't founders' life work — they're holdco assets run for cash.

One thing I didn't think hard enough about, and it turns out to be the punchline: Constellation doesn't buy companies at random. It buys a type.

Step 1: 1,074 companies, half of them from an annual report

No single source has the CSI universe, so the scanner sweeps four and reconciles them:

• Portfolio pages across every operating group → ~646 companies.
• CSI's 2024 Annual Information Form, parsed → +468 companies no portfolio page lists. This one source roughly doubled the universe.
• Press releases, for tuck-in acquisitions too small to hit any list.
• DNS and HTTP probing to attach a real website to each name.

That's 1,074 active companies, 989 with verified domains. The hard part wasn't crawling, it was deciding when two rows are the same company: "Ecotime," "Ecotime by HBS," a bare domain and a nameless row are one business, and until they merge you can't even count.

Step 2: scoring every company for rebuildability, for $8

An LLM enriched each company from the web — vertical, workflow, segment, size, pricing model, competitors — and a second model scored the question that matters: is there an accessible niche here a solo founder could attack? Not "can you replace the incumbent," which is always no.

Fifty-five companies at 8/10 and a clean top fifteen — Teachworks (tutoring ops), Magic Pulse (salon software), Commerce Sync (POS to accounting), Conasys (builder warranties), EZFacility, FunctionFox, Top Producer, SoftChalk, TORCHx, Vizergy and the rest. The bottom is what you'd expect: safety-critical, hardware-coupled, government.

The project looked finished. Ranked list, top candidates, an attack angle for each. Then the actual constraint showed up.

Step 3: the question I never scored — can a stranger buy it?

My real brief was: which of these should I build, given that I will not do sales calls? That's a distribution constraint, and the scanner had no opinion about distribution. It had spent a thousand LLM calls on "could I rebuild this?" and never asked "could I sell it without getting on the phone?"

For a no-call motion, five things have to be true at once:

• The buyer is one person who can expense a tool — no committee, no procurement.
• The price is card-swipe range, roughly $15–200/mo: enough to matter, not enough to trigger a demo.
• Value is provable in one session — no migration, no onboarding services.
• An existing pull channel exists: a marketplace shelf, intent-heavy search, a niche community.
• Nothing regulated or safety-critical that forces a compliance conversation.

Most of the top fifteen died on the spot. Alteva is VoIP provisioning. Vizergy, 97Display and TORCHx are marketing retainers dressed as software. SoftChalk sells to institutions. EZFacility and Top Producer need multi-seat migrations, and Top Producer needs MLS board approval on top. FunctionFox is a crowded per-seat market with $9/user entrants. Notuleerservice Nederland turned out to be humans taking minutes, not software.

Two survived: Commerce Sync and Teachworks.

The two candidates I recommended, then killed

Commerce Sync was the perfect shape: $24.95–59.95/mo published, self-install from the Square marketplace with a card, MVP is pure data plumbing. I recommended it. Then I looked at the market instead of the product, and Intuit ships the same connector free — QuickBooks Online's own Square sync imports sales, fees, taxes and tips daily, for $0, and Amaka gives one away too. Your prospect's default action costs nothing, so the pitch becomes edge-case correctness, which is the hardest thing to convey without a conversation. The marketplace isn't distribution either: it's a shelf ranked by installs and reviews, held by Synder (~$3.8M ARR), Bookkeep, A2X, Webgility, Shogo and Connex. The one warm channel left is bookkeepers with dozens of merchant clients — webinars, calls, conference booths. Exactly the motion I was avoiding.

The candidate that replaced it lasted six hours. Re-ranking offline surfaced an add-on for auto repair shops: daily KPIs and declined-service follow-up on top of Tekmetric and Shopmonkey. Strictly better on the axes that mattered — those platforms publish $199–499/mo, so a $49–99/mo add-on is an impulse buy; no free substitute exists in automotive; no migration, so no seasonal switching window. Then diligence: AutoRx already ships exactly that, $50/mo, self-serve, no credit card, "set up in under 5 minutes," plus the coach-affiliate program I thought was my wedge. PULSE, HiBeam and Steer cover the variants. And Tekmetric's API is partner-gated — a discretionary 2–3 week approval, i.e. a sales call before I'm allowed to sell anything.

Both picks were invisible on the incumbent's website and obvious one search past it, where the swarm of $50/mo AI-era entrants lives. Reading the incumbent tells you about the incumbent. It tells you nothing about who is already eating the niche.

Step 4: replacing my judgment with one HTTP request per company

Two retractions in, my recommendation rested on my own prose, which is not a system. So I turned the lens into code: a deterministic collector that reads each company's own pricing page. One HTTP fetch per company, no model tokens, cheap enough to run across the entire universe.

Why obsess over a published price? It's the cheapest machine-readable proxy for whether a stranger is allowed to transact. A hidden price usually means four things at once: the price varies per deal, so someone negotiates it; the contract has to be big enough to pay that someone; there's an onboarding fee; and the buyer is a committee. None of that is visible from a homepage. All of it shows up as "contact us for pricing."

It also gates the plan. A no-call motion runs on comparison content — "the incumbent is $15–187/mo, here's what we charge." You can't write that page against a vendor who won't say, and the reader can't self-qualify, so every visitor arrives needing a conversation.

The result: 1 in 100 companies publishes a price

442 companies scored; the rest had dead or unusable domains. Nine clear 4.5/10 on the composite and two clear 6. I had budgeted the expensive research stage for ~150 survivors. That gap is the most informative number in the project.

"1% publish a price" and "my crawler can't find pricing pages" produce identical output, so before believing it I re-checked a random 45 companies with a deliberately different implementation: follow each site's own navigation instead of guessing URLs, six pages per company instead of one, annual and "from $X" prices included. Still one published price in 45. Restricted to the 267 companies that own their own domain rather than redirecting to an operating group's corporate site, the base rates barely move: 2% publish a price, 31% say "contact sales," 9% have a pricing page at all.

Any measurement whose failure mode correlates with the thing you're measuring will lie to you in the flattering direction. If you build one of these, build the second implementation too.

Why: Constellation buys the exact companies I can't attack

1% of a portfolio publishing a price is not a scraping artifact. It's what Constellation buys: phone-sold, contract-renewing, switching-cost-heavy software with high retention. A product a stranger could buy unattended for $49/mo would be a bad CSI acquisition — nobody has to call to cancel it, and its moat isn't made of migration pain.

So the list I spent weeks building was pre-filtered against me. The acquirer's selection criteria are almost the exact inverse of the solo-founder-with-no-sales-team criteria.

Which kills the take I started with. "AI makes it cheap to rebuild SaaS" is true and mostly irrelevant: building was never the moat in vertical software. The moat is the sales motion, and AI has not made it cheaper to be a stranger asking a municipal utility to change billing systems. The cost floor that collapsed was already the smallest line item.

The two survivors are the same shape: published price, self-serve signup, under $200/mo, sold to an owner-operator. Teachworks ($15–187.99/mo, 4.6 across 64 Capterra reviews) is the one I'd have built — its whole comparable set is paid and self-serve (TutorBird $14.95/mo, TutorCruncher $30/mo plus card fees, MyMusicStaff $14.95/mo), so there's no free default to lose to, competitors visibly fund comparison content, and the pain is page-shaped: unpredictable per-lesson bills, a card fee that's a calculator waiting to be written. Lightning Payroll (Australian payroll, from $15.50/mo, self-serve) is the other.

Two out of 1,074 is not a pipeline. Starting over, I'd change the seed list, not the code: point the same machinery at a universe that's self-serve by construction — the Shopify, QuickBooks or Atlassian app ecosystems, or the sub-$100/mo vertical long tail — where published pricing is near universal instead of 1%.

The uncomfortable version: the cheapest way to own phone-sold, high-retention vertical software is not to rebuild it with AI. It's to buy it. Which is precisely the business Constellation is in, and precisely why its portfolio was the wrong place to look.

So am I building any of them? No

Teachworks is a genuinely good target and I'm still not doing it, because the honest version of the plan has a sales motion in it. Not cold calling — a migration to land, a December–January switching window, an affiliate deal to negotiate, an importer to hand-hold. Quiet, but a sales motion, and "no SaaS sales" was the constraint I started with. The scan is what it took to learn the constraint eliminates the category, not just the candidates.

So the deliverable is this post, not a product. The salvage, if you're running the same play: the screen cost one HTTP request per company, it would have eliminated both of my picks in an afternoon, and I should have run it before the thousand LLM calls instead of after.

Five things I'd tell the next person running this play

• Score the constraint you're actually optimizing. Every candidate that looked great and then died, died on the axis I hadn't scored.
• A free platform-native substitute is a hard kill. You're priced against what the platform gives away, not against the incumbent.
• Read one search past the incumbent's pricing page. That's where the $50/mo swarm lives, and where both of my picks died.
• Never let missing data become a passing score. "No pricing page" is not "no sales gate."
• A 1% base rate is a statement about your seed list. When an entire universe fails one screen, someone else selected that universe, for their own purposes.