Grammarly Was the App | App to Stack

01Grammarly did not fail

Grammarly is one of the cleanest examples of the transition from apps to stacks precisely because Grammarly did not fail. It built an extraordinary company. It attracted more than 40 million daily users. It became embedded in browsers, email systems, word processors, and corporate workflows, with integrations the company counts at over a million apps and websites. By May 2025 it was disclosing more than $700 million in annual revenue. It had brand recognition, distribution, paying customers, and a product people understood in one word.

Then artificial intelligence made Grammarly’s central capability ordinary.

That distinction matters. AI did not make grammar correction less useful. It made grammar correction too easy to remain an independent destination.

02The SaaS bargain

The software-as-a-service model was built around a simple bargain. A software company identified something people needed to do repeatedly. It packaged that capability inside an application, designed an interface around it, and charged for access.

Grammarly packaged grammatical correction.

The user wrote something, opened Grammarly, watched for colored underlines, reviewed the recommendations, and accepted or rejected each change. Eventually Grammarly expanded beyond spelling and punctuation into tone, clarity, style, and rewriting. But the structure never changed: the user attended to Grammarly while Grammarly helped improve the writing.

Grammarly owned the capability, but it also demanded attention.

This was normal. Every application demanded attention. You opened Salesforce to update a customer record. You opened Asana to manage a project. You opened Calendly to schedule a meeting. You opened Canva to create an image. You opened Grammarly to improve a sentence.

The application was the place where the work happened.

03The capability escaped the application

Generative AI changed the unit of software.

“Correct the grammar in this paragraph” is no longer a reason to open an application. It is an incidental instruction inside a much larger request.

Sentence structure is included. Tone is included. Formatting is included. The title, metadata, links, image selection, and publication can all be included.

Grammar has not disappeared. It has been absorbed.

Figure 01 Where grammar went

What the person says

“Write and publish the article.”

↓   absorbed   ↓

What the coworker does, unwatched

• research
• draft
• grammar
• tone
• structure
• format
• title
• metadata
• links
• cover image
• schedule
• publish

This is more than automation. Automation would make the Grammarly process faster. Absorption removes Grammarly from conscious attention. The grammar is still corrected, but the person no longer performs — or even monitors — the correction.

The capability remains. The application disappears from view.

04Grammarly understood what happened

Grammarly’s response reveals how seriously it understood the threat, and how early.

In February 2024 the company eliminated 230 positions — roughly a quarter of a workforce it had just described as a thousand people. Its chief executive told staff the cuts were not about cost; the company’s finances were “strong.” It needed, in his words, “a different mix of capabilities and skillsets” for an AI-enabled workplace.

In December 2024 it announced the acquisition of Coda, the collaborative document and workflow platform, and installed Coda’s founder, Shishir Mehrotra, as Grammarly’s CEO. In May 2025 it secured $1 billion from General Catalyst — not an equity round but non-dilutive financing from the firm’s Customer Value Fund, to be repaid from the revenue it generated, earmarked for sales, marketing, and acquisitions. Five weeks later it agreed to buy Superhuman, the premium email application. Terms were not disclosed.

Then Grammarly did something remarkable.

On October 29, 2025, it stopped calling the company Grammarly.

Figure 02 Twenty-nine months, one conclusion

1. Feb 2024230 positions cut

   About a quarter of staff. Stated reason: a different skill mix for an AI-centered workplace, not cost.

2. Dec 2024Acquires Coda; Coda’s founder becomes CEO

   Rahul Roy-Chowdhury steps aside. Shishir Mehrotra, ex-YouTube and Microsoft Office, takes over.

3. May 2025$1 billion from General Catalyst

   Non-dilutive, repaid from revenue. Same announcement discloses $700M+ revenue and 40M+ daily users.

4. Jul 2025Acquires Superhuman Mail

   Email becomes the “communication surface.” Price undisclosed.

5. Oct 2025Company renamed Superhuman

   Grammarly survives as a product inside a suite: Grammarly, Coda, Mail, and Superhuman Go, an assistant that works across apps.

6. Feb 2026The market prices the thesis

   Anthropic ships legal, finance, and marketing plugins for Claude Cowork. Software stocks fall worldwide. Superhuman, the same month, buys the spreadsheet startup Rows.

7. Jul 2026Coda becomes Superhuman Docs

   The last acquired brand is folded into the parent’s name.

The parent company became Superhuman. Grammarly remained as a writing product, placed inside a broader suite containing Superhuman Mail, the workspace that had been Coda, and Superhuman Go, an assistant intended to work across applications. In July 2026 Coda itself was renamed Superhuman Docs, and the last of the acquired names disappeared into the parent’s.

The company did not abandon grammar. It abandoned the belief that grammar was large enough to define the company.

That is the case study.

Grammarly was not reacting to another grammar company. It was reacting to Claude, Gemini, Grok, ChatGPT, and the emerging class of AI coworkers capable of treating grammar as one invisible operation within a much larger assignment.

05The market understood it too

If you want evidence that this is not one company’s idiosyncratic story, look at the first week of February 2026.

Anthropic released a set of plugins for Claude Cowork that adapted its workplace assistant to legal, finance, and product-marketing work. The plugins were not products in the SaaS sense. They were bundles of instructions and connections that let a general coworker do what a category of applications had done. Within days, Thomson Reuters and LegalZoom each fell more than 15 percent in a session; RELX, the parent of LexisNexis, and FactSet took double-digit hits; Salesforce and Workday declined. By the end of that week an S&P index of software stocks was down 17 percent for the year, with Salesforce off 25 percent and Intuit off 31 percent.

Analysts split on whether the fear was warranted. One at Wedbush argued that enterprises would not rip out systems of record because of a strong model. A Gartner analyst put it differently: investors were not surprised that this would happen, only that it was happening in 2026 instead of 2027.

Grammarly’s board reached in 2024 the conclusion that public markets reached in 2026. The lag was the only difference.

Note what the selloff was pricing. Not that Claude had built a better legal research tool than LexisNexis. That the person with a legal question would no longer open a legal research tool at all.

06Central capabilities become ordinary ones

There are a lot of SaaS companies in Grammarly’s position.

Each company believes it has a product. From the perspective of an AI coworker, many of them possess only a capability.

Figure 03 The sentence that absorbs the product

The application	Its central capability	The request that swallows it
Grammarly	Grammar, tone, and clarity correction	“Write and publish the article.”
Calendly	Scheduling	“Organize lunch with the team.”
Asana	Project management	“Launch the new website.”
Salesforce	Customer-relationship management	“Manage this account.”
DocuSign	Electronic signature	“Complete the agreement.”
QuickBooks	Bookkeeping	“Keep the books and tell me if anything is wrong.”
Canva	Image and layout creation	“Make the announcement look good.”

Scheduling is Calendly’s central capability. To an AI coworker, scheduling is one step inside “organize lunch with the team.” Project management is Asana’s central capability. To an AI coworker, it is one step inside “launch the new website.” Customer-relationship management is the organizing capability behind an enormous category of software. To an AI coworker with access to email, calls, calendars, and customer records, updating the CRM is background bookkeeping performed while it manages the relationship.

The central capability of the SaaS company becomes one of hundreds of capabilities available to the AI coworker.

That is the inversion.

During the application era, the person selected the software and operated its capabilities. During the stack era, the person states the desired artifact and the AI coworker assembles whatever capabilities are necessary to produce it.

Whether that coworker is Claude Cowork, Gemini Spark, Grok Bot, ChatGPT Work, or something not yet named is almost irrelevant. These products are early versions of a new operating layer. They do not merely help people operate software. They increasingly decide when software should be used at all.

07A larger suite is not a stack

Grammarly’s strategy is understandable. If grammar has become too small, assemble a larger collection of capabilities: documents, email, spreadsheets, workflows, agents, and writing assistance. Superhuman Go now advertises an agent store with partner agents from Box, Gamma, Jira, Confluence, and Google Workspace, and it works, the company says, across more than a million apps and websites.

But a collection of applications is not automatically a stack.

A stack is not a bundle purchased from the same company. It is the accumulated operating environment surrounding an AI coworker: its instructions, permissions, memory, relationships, tools, data access, and understanding of how its person works.

Superhuman is attempting to occupy more of that environment. It faces the same fundamental threat Grammarly faced. Claude Cowork or ChatGPT Work can treat Superhuman Mail, Superhuman Docs, and Grammarly as underlying tools. The general AI does not care whether those capabilities come from one company or four.

Grammarly may have escaped the grammar application only to become a larger collection of applications waiting to be absorbed.

The company understands the direction correctly. The unanswered question is whether it can become the stack before the frontier AI companies do.

08What survives

This does not mean every SaaS company disappears.

Some applications control important systems of record. Some possess proprietary data. Some execute regulated transactions. Some connect to physical infrastructure, move money, administer contracts, or maintain legally significant records. Those functions cannot be replaced by an eloquent response from a language model.

But the relationship changes.

1. Salesforce may survive

   But fewer people will spend their days manually operating Salesforce. It becomes the trusted place where the coworker reads and writes the account.

2. DocuSign may survive

   But the person may never visit DocuSign. The signature becomes one absorbed step inside “complete the agreement.”

3. QuickBooks may survive

   As an accounting system of record. But bookkeeping no longer needs to occupy the owner’s attention. The coworker keeps the books and surfaces only an exception.

4. The interface does not

   Reliability, permissions, data, and the ability to complete an action are what get invoked. The screen the human used to stare at is what gets skipped.

Many SaaS companies will remain alive underneath the stack while disappearing from the human experience.

That is not a small change. Most SaaS economics were built around seats, engagement, usage, interfaces, training, and switching costs. The customer paid because employees spent time inside the application. The emerging model rewards the opposite. The best capability is the one that can be trusted and then forgotten.

09Adding AI does not solve the problem

The predictable response from SaaS companies is to add an AI button.

They place a chatbot beside the dashboard. They add a text-generation window. They rename automation as agents. They continue asking the customer to open the application, study the interface, configure the workflow, and monitor the results.

That improves the application. It does not address the transition.

The threat is not that another application has a better chatbot. The threat is that the user’s AI coworker no longer needs the application’s interface.

Grammarly could have added increasingly sophisticated writing features forever. It did, in fact, ship a resume builder, a citation finder, an AI grader, and a “reader reactions” agent. None of that changes the fact that a person can ask ChatGPT Work or Claude Cowork to write the finished document correctly in the first place.

The SaaS company must decide what it actually owns.

If it owns a durable capability, it should make that capability callable by AI coworkers. If it owns a system of record, it should become the trusted place where AI reads and writes consequential information. If it owns neither, then it owns an interface — and interfaces are becoming extraordinarily vulnerable.

10The ledger

The claim, stated so it can be checked: most SaaS companies whose central capability is a task a general AI coworker can perform will, within a few years, either become callable infrastructure beneath that coworker or be absorbed outright.

Already true

A $700 million company with 40 million daily users concluded its category was too small and renamed itself. A plugin release for one AI coworker moved the market capitalization of legal, financial-data, and enterprise software companies by double digits in days. Coworker products from Anthropic, OpenAI, Google, and xAI now ship with connectors to the very applications this piece names.

What has to happen

Coworkers need reliable, permissioned write access to systems of record, not just read access — and businesses have to trust it. Seat-based revenue at the large SaaS vendors has to visibly decline, not merely stop growing. At least one well-known application company has to report that most of its usage arrives through agents rather than through its own interface.

Where I am probably wrong

Regulated enterprises move slowly, and a Wedbush analyst was right that nobody rips out a core system because a model got better. If the coworkers stay unreliable at consequential actions — money, contracts, records — the applications keep their interfaces for longer than I expect, and “absorption” stays a consumer story. And Superhuman may prove me wrong the other way: a suite that already sits in 40 million people’s browsers has a real chance of becoming the stack rather than being eaten by one.

11The Grammarly lesson

Grammarly began as a successful application built around a valuable capability.

Generative AI preserved the value of the capability while eliminating its scarcity. Grammarly responded by cutting a quarter of its staff, acquiring adjacent products, changing leadership, borrowing a billion dollars against future revenue, developing agents, and ultimately surrendering the corporate identity it had spent sixteen years building.

That is how powerful this transition is.

AI will not eliminate software. It will reorganize software beneath the user’s attention.

The old question was: Which application should I use?

The new question is: What do I want completed?

Once the desired artifact can be stated directly, grammar correction, scheduling, project management, customer follow-up, formatting, publishing, and hundreds of other activities become internal operations of the stack.

Grammarly was once the application.

Grammar is now merely one of the things the coworker knows how to do.