Slack is now offering its first generative AI features after announcing them last fall.
By clicking Slack’s star-shaped AI button, users can get a written summary of everything that happened in a channel over a specified date range. Slack AI can also summarize conversation threads, and it’ll attempt to answer searches with written summaries and suggested follow-up questions. Third-party Slack apps are now adding generative AI features as well, and routine digests of channel activity are coming soon.
The new AI features will cost extra, though the exact pricing is unclear. Slack Enterprise customers can upgrade now by contacting their sales reps, and the company is promising more information for other plans in the future.
Noah Weiss, Slack’s chief product officer, says the AI features have already changed how people use the platform in the company’s early testing, and he made clear in an interview that Slack is leaning hard into AI as the future of its business. It might have to, as rivals such as Microsoft and Google bring AI to their own workplace chat tools.
“This age of generative AI technology is obviously really interesting in the consumer world, but the possibilities in the world of enterprise software are completely mind-blowing,” he says.
Quicker catch-ups
When you ask Slack AI for a channel or thread summary, you’ll get a list of headlines for each discussion topic along with brief, AI-written description. Each blurb has a “read more” button, which brings up extra details, plus citations linking to the original conversation.
For now, there’s no way to further refine the summaries, which means you’re trusting the AI to make sure didn’t miss anything important. Support for follow-up questions, which Microsoft has already built into Copilot for Teams, may come later.
“Right now we’re really gearing toward things that are accurate rather than comprehensive,” Weiss says.
Slack is also applying AI to search, but in a somewhat different way. While search results will still show snippets from relevant conversations, they’ll also include AI summaries at the top along with suggested follow-ups, akin to what you’d find in a Google search.
It’s designed for those times where you search for something broad—like the name of some company initiative or obscure industry acronym—and aren’t sure what else to ask for. By synthesizing an answer, Slack AI can gather up the necessary context along with ways to explore further.
Weiss says that getting search right is the “holy grail” of productivity, citing a decade-old McKinsey study that found the average knowledge worker spends 20% of their time searching for information or find a colleague with the right answers.
“What [Slack AI] winds up doing is synthesizing, from all this messy conversational data, actual answers to questions that you have,” he says.
Playing it safe
Adding AI summaries to Slack wasn’t as simple as just setting a large language model loose on everyone’s workspaces, Weiss says. He doesn’t get into specifics on the models that Slack is using, but says the company trained a “variety of models” to better-understand the freewheeling nature of text chat.
“You can’t just treat the whole channel like it’s a long web page,” he says. “That produces horrible results. You have to add the notion of structure to something that’s fundamentally unstructured.”
Perhaps more importantly, Slack trained its models to err on the side of caution with its responses. In a demo, Weiss showed how if a user searched for a nonexistent coworker, Slack AI would say that it didn’t have enough information to answer.
“We literally tell the models, don’t try to come up with answers that you don’t have corroborating evidence for,” Weiss says.
The AI chat arms race
Slack isn’t alone in building into workplace chat, though its take is a bit different from its rivals, at least for now. Both Copilot in Microsoft Teams and Duet AI in Google Chat are offering AI in a more conversational way, with a persistent chat box where users can ask questions about their conversations and documents.
Slack’s approach is narrower in scope, but that might not be a bad thing. Weiss says the company designed Slack AI to be transparent about where it’s getting its information; clicking a button to summarize a specific amount of data could feel like less of a leap than asking a chatbot to go out and retrieve that info for you.
“We don’t want people to have to blindly trust this stuff out of the box,” Weiss says. “If you’ve been using a lot of AI tools you know how potentially risky that can be.”
Even so, Weiss says there’s a lot more to come.
“There has never been a moment like this, where there’s this huge step change in what’s technologically possible with AI, and frankly I think no one really knows what’s possible yet in the ways you can apply it,” he says. “We’re really just scratching the surface.”
Slack is now offering its first generative AI features after announcing them last fall.
By clicking Slack’s star-shaped AI button, users can get a written summary of everything that happened in a channel over a specified date range. Slack AI can also summarize conversation threads, and it’ll attempt to answer searches with written summaries and suggested follow-up questions. Third-party Slack apps are now adding generative AI features as well, and routine digests of channel activity are coming soon.
The new AI features will cost extra, though the exact pricing is unclear. Slack Enterprise customers can upgrade now by contacting their sales reps, and the company is promising more information for other plans in the future.
Noah Weiss, Slack’s chief product officer, says the AI features have already changed how people use the platform in the company’s early testing, and he made clear in an interview that Slack is leaning hard into AI as the future of its business. It might have to, as rivals such as Microsoft and Google bring AI to their own workplace chat tools.
“This age of generative AI technology is obviously really interesting in the consumer world, but the possibilities in the world of enterprise software are completely mind-blowing,” he says.
Quicker catch-ups
When you ask Slack AI for a channel or thread summary, you’ll get a list of headlines for each discussion topic along with brief, AI-written description. Each blurb has a “read more” button, which brings up extra details, plus citations linking to the original conversation.
For now, there’s no way to further refine the summaries, which means you’re trusting the AI to make sure didn’t miss anything important. Support for follow-up questions, which Microsoft has already built into Copilot for Teams, may come later.
“Right now we’re really gearing toward things that are accurate rather than comprehensive,” Weiss says.
Slack is also applying AI to search, but in a somewhat different way. While search results will still show snippets from relevant conversations, they’ll also include AI summaries at the top along with suggested follow-ups, akin to what you’d find in a Google search.
It’s designed for those times where you search for something broad—like the name of some company initiative or obscure industry acronym—and aren’t sure what else to ask for. By synthesizing an answer, Slack AI can gather up the necessary context along with ways to explore further.
Weiss says that getting search right is the “holy grail” of productivity, citing a decade-old McKinsey study that found the average knowledge worker spends 20% of their time searching for information or find a colleague with the right answers.
“What [Slack AI] winds up doing is synthesizing, from all this messy conversational data, actual answers to questions that you have,” he says.
Playing it safe
Adding AI summaries to Slack wasn’t as simple as just setting a large language model loose on everyone’s workspaces, Weiss says. He doesn’t get into specifics on the models that Slack is using, but says the company trained a “variety of models” to better-understand the freewheeling nature of text chat.
“You can’t just treat the whole channel like it’s a long web page,” he says. “That produces horrible results. You have to add the notion of structure to something that’s fundamentally unstructured.”
Perhaps more importantly, Slack trained its models to err on the side of caution with its responses. In a demo, Weiss showed how if a user searched for a nonexistent coworker, Slack AI would say that it didn’t have enough information to answer.
“We literally tell the models, don’t try to come up with answers that you don’t have corroborating evidence for,” Weiss says.
The AI chat arms race
Slack isn’t alone in building into workplace chat, though its take is a bit different from its rivals, at least for now. Both Copilot in Microsoft Teams and Duet AI in Google Chat are offering AI in a more conversational way, with a persistent chat box where users can ask questions about their conversations and documents.
Slack’s approach is narrower in scope, but that might not be a bad thing. Weiss says the company designed Slack AI to be transparent about where it’s getting its information; clicking a button to summarize a specific amount of data could feel like less of a leap than asking a chatbot to go out and retrieve that info for you.
“We don’t want people to have to blindly trust this stuff out of the box,” Weiss says. “If you’ve been using a lot of AI tools you know how potentially risky that can be.”
Even so, Weiss says there’s a lot more to come.
“There has never been a moment like this, where there’s this huge step change in what’s technologically possible with AI, and frankly I think no one really knows what’s possible yet in the ways you can apply it,” he says. “We’re really just scratching the surface.”