The savings number is line one of five
The most-shared SaaS story this month is worth knowing about, because it's the rare one with receipts. swyx and the Latent Space team took a real tool they pay $43,000 a year for — a speaker-management platform for events — and ran a public hackathon asking teams to replace it with AI agents in a weekend. Seventy-two submissions. Around 100 billion tokens. Real human judges, real evals, a $10,000 prize pool.
The headlines wrote themselves: SaaS is cooked, agents eat everything. But the more useful signal for a small business is the quieter one the experiment surfaced — the hard part wasn't building the replacement. It's everything that comes after the build.
That maps exactly onto what happens when an owner cancels a subscription and points an agent at the gap. The cancel is real money and the easiest number in the world to compute. It's also the only number anyone computes. The rest of the ledger gets paid in hours nobody records.
The five-line replacement ledger
Run this before you cancel anything. Five lines, real numbers.
Line 1: the subscription you stop paying
The headline. Real, immediate, worth having. For the hackathon's event tool it was $43K a year; for a typical small business it's more often $200–$800 a month of some mid-tier tool. Write it down — then keep writing. The next four lines are paid in hours, and hours you don't count are hours you didn't save.
Line 2: the maintenance retainer you signed yourself into
A licensed tool is maintained by people whose whole job is maintaining it. A replacement you generated is maintained by you. The model under your agent changes monthly. Your evals catch the drift or your staff do. Every feature update you used to get for free now shows up as your own to-do. Price this line like a vendor would: a couple of hours a month, minimum, forever — with a named owner. "Whoever notices" is not an owner.
Line 3: spec and context upkeep
An agent runs a workflow only as well as the process is written down, which means your one-page spec is now part of your product. It goes stale the first week reality drifts from the page — a new pricing rule, a new client exception — and a stale spec is worse than no spec, because it fails confidently. The settings screen used to document your process for free. Now you own that too: check intent before every change, keep the spec current, and never rely on the code to explain itself.
Line 4: the edge-case tax
Every tool you license quietly absorbs your weird cases — the one-off refund path, the client with the odd contract, the March report format accounting suddenly needs. You never noticed, because the vendor handled them. After a replacement, each of those is a support ticket you file to yourself. Price the line before you commit: run the replacement in parallel with the subscription for two to four weeks, log every gap, then decide whether the gaps are tolerable or disqualifying. Discovering them after the cancel is the expensive way to learn what the tool was actually doing all those years.
Line 5: who answers at 2 AM
When the licensed tool breaks, you open a ticket and someone else's on-call rotation deals with it. When your agent breaks the night before your busiest week, you are the support contract. Some businesses can absorb that risk; if yours can't, that's not a failure of nerve. That's line five telling you that part of what you were buying was the SLA — and it's a legitimate thing to keep buying.
When the ledger comes out positive
Replacement wins when the work is boring, well-specified, and high-volume. Same task, over and over, fits on one page, used daily — the maintenance lines stay small and line 1 dominates. Invoice generation, lead intake, quote assembly, report building: that's the win zone. If a hackathon can build the thing in a weekend, a boring version of it is exactly what agents maintain well.
When it silently loses
Three quiet failure patterns, all survivable if you see them coming:
- Low volume. A $200-a-month tool used weekly costs about $25 per use. Two hours of monthly maintenance loses to that on its own. The math only works when volume amortizes the upkeep.
- Compliance weight. If what you're actually buying is the audit trail, the filing, the certificate — the vendor's product is the paperwork, not the workflow. An agent replicates the workflow and inherits none of the weight.
- Edge-case-heavy work. If every job carries a client-specific exception, the spec never converges and you've hired yourself a permanent side job at your own hourly rate.
The 20-minute exercise
Before the next renewal you're tempted to cancel:
- Pull your subscription list. Pick the one you're most confident an agent could replace.
- Write the five lines with real numbers — subscription saved; monthly hours times your actual cost; spec upkeep; one parallel-run discovery month; one honest sentence about who covers breakage at night.
- Add the lines. That total is the honest ledger.
- If it's still positive, run the parallel test before you cancel — never cancel first and discover line 4 live.
What not to do
- Don't cancel on a demo. A weekend clone proves capability, not economics. The build was never the expensive part.
- Don't treat the build as the project. The build is the weekend. The product is the maintenance, the spec, and the pager.
- Don't assume your hourly cost is zero. The most common way this ledger lies is the owner's time entering the math at $0. It isn't.
One line to keep
Canceling the subscription is the first line of the savings, not the whole of them. The businesses that win on replacement priced all five lines before they canceled — and kept paying, happily, wherever the ledger said to.
If you want the full framework for making any agent earn its seat — the permission ladder, the boring first assignment, and the one-page spec format that keeps replacement maintenance cheap — it's all written down in The AI Agent Owner's Playbook. $49 CAD, instant download.
Top comments (0)