The cohort spreadsheet is open on my second monitor. Forty-one columns. 2,200 rows. Three months of public ARR signals scraped from Crunchbase, Latka, and earnings calls. The bottom row reads: $240B at risk inside 24 months. I do not love that number. The data does not care.

Here is the thesis in one breath. The seat-based SaaS layer was a UX bet. The bet was that humans would rather click through a purpose-built interface than write a prompt. That bet held from 2010 to 2024. In May 2026, with Claude 4.5 Sonnet running multi-tool agent loops at sub-second latency and ChatGPT Agent shipping Canvas-native spreadsheets, it does not hold anymore.

I am not predicting bankruptcies. I am predicting margin collapse, multiple compression, and the slow death-by-renewal that public SaaS investors will recognize. The companies will still exist. The pricing power will not.

The 12, ranked by ARR exposure

Order is rough. Conviction is high on the top six. The bottom six are where my model gets noisier and where I want pushback. Numbers are 2025 ARR estimates from Latka, Sacra, and company filings. Treat the dollar figures as ranges, not precision.

Add the column. Round down for double-counting. You get something in the $7B-$9B ARR vicinity sitting directly in the prompt-replacement crosshairs by end of 2027. Multiply by typical SaaS revenue multiples and that is the $240B exposure number. The math is rough. The direction is not.

Why transcription dies first

Otter raised at a $500M valuation in 2022. Fireflies hit the Inc 5000 last year. Both are betting their next round on the assumption that a dedicated transcription company is better than a horizontal agent that happens to transcribe. The bet is wrong. ChatGPT Voice already returns better diarization than Otter on noisy three-person calls, and it does it inside the same window where you ask follow-up questions. The product moat was always thin. The pricing moat is gone.

Why DeepL is the biggest single casualty

DeepL’s pitch was that they were better than Google Translate on European languages and that their enterprise tier shipped glossaries and TM integration. That was a real moat in 2022. In 2026, Claude 4.5 Sonnet ties or beats DeepL on five of the seven languages I tested (German, French, Japanese, Korean, Brazilian Portuguese), at lower latency, with structured-output mode that returns translation + tone notes + back-translation in one call. The enterprise glossary feature now ships as a 12-line system prompt. DeepL just hit $1B ARR. They will not see $1.5B.

The mechanism: agent calls eat seat licenses

The compression has a specific shape. Watch.

A 40-person company in 2024 paid Otter ($16.99/seat/mo), Calendly ($10/seat/mo), DeepL ($25/seat/mo for one localization person), Typeform ($25 for one ops person), and Canva Pro ($15/seat). That same company in 2026 pays for one ChatGPT Business seat ($25/seat/mo) and a Claude API budget for agents. The spreadsheet I ran on three real companies last month: average seat-SaaS spend dropped 47% in the past 12 months. Agent API spend is up 380%, off a small base. Net: down 31% in horizontal SaaS spend per FTE.

That is not a recession. That is a substitution event. Seat-based SaaS revenue is migrating to token-based AI lab revenue. OpenAI’s API revenue ran past $11B annualized last quarter. Anthropic crossed $7B. The money does not vanish. It just leaves the SaaS P&Ls and shows up on Sam and Dario’s.

Three calls we could be wrong on

I owe you my blind spots. The cohort model says these 12 are exposed. Three of them I would not bet aggressively on.

1. Canva (#8 on the list)

Canva has a real distribution moat. 200M+ monthly users. They already shipped Magic Studio. Their template library is a content asset, not just software. The free tier is a habit. I have Canva on the list because the underlying generation engine is becoming commoditized, but the brand and the user habit could survive that just fine. If I had to remove one, this is it.

2. Intercom Fin (#11)

Intercom pivoted hard into AI agents in 2024 and Fin is generating real revenue. They have the wedge: their inbox owns the customer history that any replacement agent would need. The risk is that Decagon and Sierra are pulling Fortune 500 logos at speed. If Intercom can’t defend mid-market against vertical agent players, this becomes the worst comp on the list. But they could also become the platform.

3. HubSpot Starter (#12)

HubSpot’s data graph is the moat. The CRM record is just the front end. Even if every UI gets replaced by chat, you still want a database that has the email history, the deal stages, the marketing automation logs. HubSpot is at risk on the low end where Attio and Day.ai are eating, but the durable Hub will be just fine. Maybe better than fine.

What survives

The SaaS that survives owns one of three things. A system of record (Salesforce, Snowflake, GitHub). A regulated workflow (Veeva, ServiceTitan, Workday). Or a multiplayer network (Figma, Notion, Linear). Everything else is a UI on top of a database, and a chat interface plus an agent loop is a strictly better UI for a growing share of workflows.

The cohort spreadsheet is messy and full of judgment calls. If you run a category on this list and you want to argue, mail tony@minstants.com with your numbers. I will publish the best rebuttals. The honest version of this piece is that I am modeling an industry-wide phase change with 90 days of clean data and a lot of vibes. The vibes are loud.