With the release of Coda 4.0 on Wednesday, Coda has joined the parade of productivity software vendors embedding generative AI features in their products.

Part of the Coda app’s new release, the Coda AI has three main components:

A “knowledge assistant” chatbot that responds in natural language to user queries, with the ability to access data from Coda docs and data made available from third-party app integrations, from calendars to sales data.

A writing assistant that can generate drafts of content such as emails or blog posts — a feature that is fast becoming table stakes in productivity applications.

“It’s kind of the new spell check,” said Coda CEO Shishir Mehrotra. “It would be weird not to have it.”

A “task assistant,” where the Coda AI is applied to workflow automations, essentially allowing it to “take action in the world,” said Mehrotra. For example, a user that has synced Salesforce data could set up a workflow that, subject to a trigger action, automatically pulls in customer data, generates an email, and then sends it to recipients.

Coda AI can automate user tasks.

Coda launched its all-in-one productivity app — which combines elements of documents, spreadsheets, work management, and customer relationship management — in 2019. It’s now used by 40,000 customers, including at firms such as Uber, DoorDash, and The New York Times.

Now, amid an office software market increasingly crowded with generative AI features, Mehrotra believes Coda has some important points of difference.

One, he said, lies in the Coda’s connectivity to a variety of different data sources, such Salesforce CRM data. This is done in a “permission-compliant” way that Mehrotra claims will ensure the AI has access to only the information that users have pulled into a Coda doc, rather than training on large corporate data sets.

“That’s a very hard thing to do in a generic enterprise AI solution. I don’t think anybody has done a good job of that yet,” he said.

Another differentiator has been less expected, said Mehrotra. Unlike many others in the market, Coda is adding generative AI features to its application at no extra cost to those on paid subscriptions — a departure from the trend toward charging additional fees for access to generative AI.

“When we started designing this, we didn’t think pricing would be a big differentiator, but it’s turning out to be,” he said.

The Coda AI is available on all its paid products — Pro, Team, and Enterprise — and free tier users will be given “credits” to try out the AI features at no cost. This bucks the trend somewhat. Generative AI is a new and potentially transformative technology, but it remains to be seen how much customers are willing to pay for access to these features and what returns they will see on their investment.

Some vendors are pitching their generative AI assistants as premium features, with pricing to match. For large business customers, Microsoft’s M365 Copilot and Google’s Duet AI are priced at $30 per user each month, in addition to existing subscription payments. Notion, which sells an all-in-one productivity platform with some similarities to Coda, was one of the first productivity app vendors to market with generative AI features, charging an additional $10 per user each month.

“Right now, software vendors are trying to figure out the best way to monetize generative AI,” said J.P. Gownder, vice president and principal analyst on Forrester’s Future of Work team.

For many software vendors, said Gownder, generative AI will become a feature of their suites or even a brand-new user interface. “The question then becomes how to monetize it: Do they simply embed it in their core product and hope for higher adoption, or can they charge for it as an additional charge?”

Mehrotra believes that charging an additional fee for generative AI features is bad for customers — not because the technology fails to deliver value, he said, but rather that it could lead to some employees being given preferential status within an organization: the AI haves and have nots.

“I think it’s very confusing for customers ... they have to decide who gets Microsoft Word with AI and who gets Microsoft Word without AI,” said Mehrotra.

For vendors, it means supporting two product platforms. “Then for the product teams, it’s impossible to build a product that way, because you can never build AI into the core of your product when you have to support a version that doesn’t have it,” he said. “We would never ship a version of Coda without mobile, as an example. And so we just thought that was kind of antithetical to the way we thought about AI.”

Coda AI includes a writing assistant that can both generate text and offer suggestions for improvement.

The Coda AI uses OpenAI's ChatGPT 4 and a "fine-tuned" version of ChatGPT 3.5, as well as its own models that run on its Amazon Web Services servers.

Customer data isn’t used to train Coda or another vendor’s models, said Mehrotra. “Nothing in Coda ever trains on your data,” he said. “We’re a very business-focused tool and so that’s sort of a given for us — you can’t train on customer data.”

As with all office software vendors that are incorporating generative AI into their products, hallucinations — a term for inaccurate information that AI language models generate at times — are a concern for Coda too.

“It’s definitely a real issue,” said Mehrotra. “I don’t think there’s a generic solution to it.” He points out that hallucinations haven’t had too much of an impact on Coda’s users so far “for the simple reason that the types of tasks that people have gravitated towards are ones where it’s just less likely to happen.”

For example, when the AI summarizes notes, it’s unlikely to create inaccurate information, he said, or invent customers. In the cases where it does present more of a challenge — for example, if a user asks the AI to generate a feature list for a product and insufficient information is provided in the prompt for example — users are advised to be cautious.

This is consistent with the approach taken by other vendors in the market. “It feels like a really well-read, know-it-all intern, so you have to treat it appropriately,” he said.

“Today, the best answer to that question would be to be very careful what tasks you give. I don’t think that’s a very satisfactory answer; over time we’re going to need better answers for that. I know there’s a lot of people working on it, but in our current deployments, I haven’t seen it be a primary issue,” said Mehrotra. 

A greater challenge, he said, is how businesses can deploy AI safely, without compromising corporate data. "Very few people have found good answers to one of the reasons I think we’ve been successful Coda AI so far,” said Mehrotra.

Coda AI creates answers to user queries based on data pulled from Coda docs and other approved company data sources.

In the long run, Mehrotra sees little doubt that generative AI will reap significant benefits when applied to workplace productivity tools.

“The value that AI provides is very high. I don’t think there’s a question that, when you get the right product with AI — especially when you move past generic applications of AI and you get to things that really understand your work — the [AI tools] that take action on your behalf are very easy to quantify in terms of benefit,” he said. “It’s not just helping you write faster, it’s removing a task you used to do every week — you can measure how long those tasks used to take for you, so I think that is quite valuable.”

Making direct comparisons between products is often difficult, and each has its own value proposition. As just one example, Microsoft has promised to provide financial support to businesses that face copyright infringement lawsuits due to outputs from its Copilot — a position that other vendors may struggle to take. At the same time, customers will differ in their priorities when it comes to selecting a tool.

Ultimately, said Forrester’s Gownder, the question is whether generative AI tools are able to provide a productivity boost that justifies a customer’s investment. With products still coming to market, relatively few have had the chance to find this out so far.

“If [employees] save time, become more effective in their jobs, and give the products good reviews, there’s a good chance that enterprises who are early adopters will continue to buy those add-on licenses into the future,” he said. “If they fail to deliver, expect a price spiral downward.”