Less Email, Better Thinking: Grammarly’s CEO on AI That Augments Communication

Jacob Effron opens by introducing Rahul, the CEO of Grammarly, and framing the company as one of the most established players in AI writing. He notes that Grammarly has more than 30 million daily active users, has raised over $400 million, and was most recently valued at $13 billion. Jacob says Rahul was a fitting guest because Grammarly has been building AI productivity tooling since long before the current generative AI wave.

Asked how AI-written communication will affect human-to-human communication in ten years, Rahul says the goal should be more meaningful connection, not more automation for its own sake. He imagines AI removing drudgery so people can focus on creativity, synthesis, and connecting with other humans. In that future, people would send fewer emails and read and write fewer documents, not more.

Rahul argues the point is not to create more content or volume but to make communication more memorable, evocative, and precise. He cites the average worker’s 1,200 context switches per workday as evidence that current work patterns are already beyond human limits. AI should help people stay in flow and focus on what they do best, he says.

Rahul acknowledges the opposite future is possible: if the cost of content goes to zero, people could drown in even more material, with AI generating content for other AI to consume. He calls that world dystopian and says writing and communicating are part of what it means to be human and how he thinks. Humans, he argues, have agency and should push toward the better outcome rather than outsourcing that human activity.

Rahul says Grammarly launched in 2009 and has been around for 15 years, riding several technology waves. It began as a mostly rules-based natural language processing product, moved into deep learning models, and now uses LLMs and generative AI. The company’s pattern, he says, is to identify user use cases first and then bring the best available technology to bear on them.

He offers a simple four-stage model of communication: ideation and conceptualization, composition, revision and polishing, and comprehension on the receiving end. For most of its history, Grammarly focused primarily on revision—helping users get words correct, follow style guides, strike the right tone, and be concise. LLMs, Rahul says, let the company turbocharge that value and expand beyond revision.

One direction is tying communication to business outcomes. Rahul says Grammarly will increasingly auto-apply correctness and polish so it can focus attention on strategic suggestions. In an event email, for example, Grammarly might notice the writer wants to drum up enthusiasm and suggest explaining why people should attend, such as free food; in a board email, it might notice a missing call to action.

The second direction is being present through the entire communication lifecycle: ideation, composition, revision, and comprehension. Rahul describes calling up Grammarly to summarize a long email thread, extract the main points, and identify whether there is an action item for the user. That moves the product from a writing checker to a broader communication assistant.

Jacob asks how much of this works by simply throwing an LLM at the problem versus requiring deeper engineering. Rahul says LLMs are powerful but early, and Grammarly cannot just throw model output over the wall. The use cases are important—people are writing emails that matter—and false positives or sensitive-text failures have real consequences.

Rahul says Grammarly fine-tunes models for its use cases and runs quality and safety evals before shipping. The company has a responsibility to shape and customize a new model until it is confident the product is consistent with its promise to users. That work sits between a raw model release and a user-facing feature.

The quality bar is set by users as much as by internal evals. Grammarly tracks whether users accept or reject suggestions and whether they engage with features, creating a continuous feedback loop. Rahul says the bar is context-dependent: there is no one-size-fits-all accuracy number, and a feature can pass internal evals yet fail when users apply it to unseen use cases.

He gives the tone detector as an example of learning from edge cases. A police department told Grammarly it used the product to type reports of serious crimes and asked the company to stop suggesting 'sound more positive' in that context. Grammarly learned to suppress certain suggestions when text is sensitive and to decide what is helpful to show in those scenarios.

Rahul says the most effective Grammarly users treat communication as a strategic imperative. They connect writing to personal and organizational goals, rather than using AI because it is novel. He cites ModMed, a healthcare company that deployed Grammarly organization-wide to improve sensitive documents, patient reports, and patient outcomes.

The lesson for enterprise AI is to start with the goal, not the feature. Rahul repeats the common pattern that AI is overhyped in the short term and underhyped in the long term, and says simply sticking AI into everything does not solve user problems. Jacob adds that evals are hard for the same reason: teams need to define what success means before they can measure it.

Jacob mentions an FTC document saying product descriptions that use the word AI must actually include AI; Rahul agrees with the rule. The exchange reinforces his point that AI should be tied to a real capability and a real outcome. Grammarly’s focus, he says, is on communication outcomes rather than AI branding.

Rahul describes ChatGPT as a watershed moment, but says Grammarly was taken aback by the speed and scale of change. ChatGPT became the fastest-growing consumer product in history, and model quality improved quickly. Grammarly saw the new stack as an enabler for its mission: it understands communication and users, and it will harness whatever technology is best at the moment.

Rahul recalls that when GPT-3 launched, Grammarly ran side-by-side comparisons and found rules-based NLP had far greater precision. The early LLMs hallucinated a lot and quality was not great. Over a couple of years, however, quality improved phenomenally; today he says edge-case deltas largely remain, but the old gap has mostly disappeared.

Grammarly cannot punt safety to a future model release. If models are unsafe, Rahul says, the company must post-process and fine-tune them and run safety evals customized to its use cases. He also sees smaller, more efficient models enabling more inference at the edge, which could help with security, privacy, cost, latency, and user experience—especially for compute-intensive prompts with large context windows.

Looking to the next generation of models, Rahul is most excited about multi-step reasoning and agentic workflows. A board email about a product launch is not one prompt: it requires context about the launch, marketing, engineering, product, cost structure, PR plan, board expectations, brevity, and tone. Grammarly wants to orchestrate that complex communication flow and remove the figurative cutting and pasting of context from place to place.

Rahul says he once assumed models would become plug-and-play, but they remain idiosyncratic 'special snowflakes' that respond differently to prompts and have different safety parameters. Grammarly runs about half a dozen production models, a mix of closed and open source, most fine-tuned on user data. It distills them to the smallest and most efficient model that preserves quality for each use case, rather than chasing the smallest model possible.

Jacob asks how Grammarly tailors output to a company or individual. Rahul points to Grammarly’s scale: it processes 75 billion user events every day, giving it fresh, high-quality contextual data for fine-tuning and training across use cases. He calls that user data one of Grammarly’s unique advantages in a fast-changing AI landscape.

For individuals, Grammarly lets users shape how they sound, though Rahul says the voice customization is more manual than the company would like and should become more automated over time. For organizations, it enforces style guides, brand tones, corporate values, and compliance across internal and external communication. The goal is consistency with the organization’s identity wherever employees write.

Asked whether personalization is retrieval or fine-tuning, Rahul says Grammarly typically fine-tunes on use cases and separately ingests organization-specific knowledge such as style guides and brand tones. He cites a loan officer example: strict rules may sit in a 25,000-rule document, and Grammarly can automate those rules in the flow of communication. That turns manual, out-of-flow guidance into in-flow assistance.

Rahul describes Grammarly’s model evaluation as multi-dimensional rather than a single magic number. The company looks at external benchmarks closest to its use cases, runs safety evals informed by real false positives in the field, and uses expert side-by-side comparisons against idealized human output. It then ships to a small percentage of users and watches whether they accept or reject suggestions; sometimes a model that passed offline evals fails in the real world.

Jacob asks about competition from ChatGPT, Apple Intelligence, Notion, Google Docs, and others. Rahul welcomes it: competition brings attention to the problem space, and launches like Apple Intelligence remind people what Grammarly does. He says the market is massive and Grammarly should stay laser-focused on solving its use cases well.

Rahul names two differentiators. One is the user-data loop that improves product quality; no advantage is infinite, he says, but this one is large. The other is distribution across a fragmented tool landscape: Grammarly itself uses 200 B2B SaaS apps, and Grammarly sits inside Gmail, Microsoft Word, Slack, Salesforce, Greenhouse, and other line-of-business tools to improve the investments customers already have.

On AI org design, Rahul says Grammarly does both central research and embedded teams. A core research group looks 18 to 24 months out at capabilities such as on-device inference and the data-collection infrastructure needed to support them. AI engineers are embedded in full-stack product and feature teams, because a launch includes front end, back end, and AI capabilities working together.

Rahul frames enterprise AI as a multi-year transformation journey, like moving from on-prem to cloud, not a set-and-forget deployment. Enterprises need to select vendors they trust for several years, and Grammarly wants to be that partner. He says productivity gains are still elusive outside a few areas such as software engineering and code generation, so Grammarly emphasizes measurable value: the average organizational user saves 19 days per year.

In education, Rahul compares AI to calculators and looking up code snippets: the tool can be misused, but banning it is not the answer. He says educators are eager to engage and equip graduates with AI skills, and Grammarly launched citation of AI use so students can show where they used it. The company is also launching Authorship, which gives provenance for each part of a document—manually written, cut and pasted, paraphrased, or AI-generated—so institutions can set their own guard rails.

Asked what a Chrome team might build with AI, Rahul says people will search for and consume content very differently, with AI synthesizing information, remembering what they have seen, and surfacing it at the right moment; he jokes AI might finally solve too many tabs. In the quickfire round, he calls chat interfaces overhyped and hopes they fade into the background, while saying AI is underhyped as a tool for upskilling people worldwide. He says the tone feature has resonated more than expected—people stop him in the street to describe how it saved a difficult conversation. He has changed his mind on consumer versus enterprise, now seeing the distinction as artificial and treating adoption as one journey from free to premium to team licenses to enterprise sales. He is excited about AI in healthcare, citing AlphaFold and ModMed, and points listeners to grammarly.com and Grammarly’s social channels.