Are we entering the age of the AI ‘copilot’?

The impact of generative AI was hard to miss in 2023.

In martech alone, Scott Brinker and Frans Riemersma note in their Martech for 2024 report a net increase in tools and applications of 18.5% between May and November 2023, with “generative AI … responsible for at least 73% of the increase.” The authors add that this is “in addition to all of the generative AI features that have been embedded into existing martech products this year.”

Thanks to the overwhelming popularity of ChatGPT, many of these generative AI features take the form of virtual assistants, or ‘copilots’. Microsoft’s CTO and EVP of AI has even declared that copilots will “become an expectation for how all software works” in the future.

How did the copilot trend sweep across the tech landscape so quickly? In this briefing, we’ll look at the key developments that gave rise to the boom, before asking whether AI copilots are as transformational as many believe – or are we facing a repeat of a familiar fad?

The emergence of the AI copilot

The release of an official API for ChatGPT in March last year finally made it possible for developers to officially integrate ChatGPT into their tools and software. The launch was heralded by integrations with platforms like Snap, which introduced a chatbot called My AI into its premium subscription, Snapchat+; Quizlet, which introduced an AI tutor called Q-chat; and Shopify, which introduced an AI-powered shopping assistant.

The appeal of integrating ChatGPT and its capabilities within products and software had become evident almost immediately, and developers were already using unofficial access points to do so – but the resulting tools were also largely unofficial. The API release made it possible to integrate ChatGPT and its capabilities formally, for a fee based on usage.

To add to this, the significantly more capable GPT-4 was announced mere weeks later, no doubt encouraging many businesses who might have been on the fence about ChatGPT’s reliability to take the plunge in incorporating it into their products. Perhaps unsurprisingly, customer service companies were early adopters of generative AI tools to facilitate their work, but sectors from travel to healthcare to beauty and cosmetics are all finding viable use cases.

Parallel to this, generative AI capabilities are increasingly being built into and shipped with workplace tools offered by giants like Microsoft, Google, Salesforce, and Adobe, often in the form of conversational assistants, which have become known as ‘copilots’.

Microsoft, as Open AI’s biggest partner and investor, has instigated and popularised the term, first announcing its generative AI enhancements to Bing and Edge as “your copilot for the web” in February 2023.

It has since built generative AI ‘copilots’ into Microsoft 365, Windows 11, and Power BI, among others, and brought Bing Chat and Bing Enterprise together under the umbrella of ‘Microsoft Copilot’. Alongside AI-powered artistic tools, Adobe has incorporated an assistant called Adobe Sensei, while Salesforce has unveiled Einstein Copilot, “Bringing a conversational AI assistant to every CRM application and customer experience” in its own words.

Demonstrating the assistant’s capabilities at the Festival of Marketing in October, Jonathan Beeston, Project Marketing Director, EMEA, at Salesforce Marketing Cloud showed how Einstein Copilot enables marketers to use natural language to accomplish tasks such as creating a customer segment. They can then edit the output and converse with the copilot to understand its reasoning.

Notably, Microsoft and Salesforce are also enabling developers to build their own ‘copilots’ through tools like Microsoft’s Copilot Studio and Azure AI Studio, and Salesforce’s Einstein Copilot Studio. So bullish is Microsoft’s CTO and EVP of AI, Kevin Scott, on the future of ‘copilots’ in software, that he has declared, “…over the coming years, [copilots] will become an expectation for how all software works.”

The problem with the copilot craze

Does every user experience need a conversational assistant? While assistants and copilots are enabling a variety of fascinating and transformative use cases, it’s important to consider whether their inclusion adds a real benefit to the product or service that couldn’t be achieved by another means.

This sentiment was voiced by GZERO Media’s Scott Nover as he trialled a selection of ‘GPTs’, custom iterations of ChatGPT, from the recently-released GPT Store. This included a hiking bot from AllTrails, which prompted Nover to ask: “When I engaged with the hiking chatbot, I wondered why it needed to be a chatbot. Wouldn’t a map do better?”

When I engaged with the [AllTrails] hiking chatbot, I wondered why it needed to be a chatbot. Wouldn’t a map do better?

– Scott Nover, GZERO

This is a position we’ve been in before. Seven or eight years ago, we saw the age of ‘peak chatbot’, in which companies were integrating conversational bots into products and services left, right and centre. Then, as now, the charge was being led by a major tech company (Facebook) and its provision of easy-to-use chatbot-building tools (within Facebook Messenger).

Looking back on reporting from the ‘peak chatbot’ era, many of the sentiments then could have been written today: The Register’s Sonia Cuff wrote in September 2017, “…the people who are the most hyped up about chatbots seem to be the people that sell platforms to build them on. As chatbots seep out of the consumer space and into our internal workplace tools, are we at risk of losing our job because a bot can do it better?”

Many would argue that the failings of chatbots during the ‘original’ hype don’t apply to LLM-powered assistants today. The chatbots of the mid-2010s failed to capture the public imagination in large part because of their clunky, error-ridden user experience, which was a far cry from pseudo-human interactions and answers to every conceivable query.

But generative AI copilots have a much more advanced conversational facility – so, isn’t the promise of that era finally being realised?

The major issues with this are that first, generative AI copilots do still have issues with producing an accurate response to queries, but instead of an error message, these issues manifest in the form of ‘hallucinations’: false information, confidently stated. While trialling Bard Extensions, which connects Google’s Bard assistant to a variety of Google-owned tools like Gmail, Google Drive, and YouTube, The New York Times’ Kevin Roose wrote that, “Bard succeeded at some simpler tasks … But it also told me about emails that weren’t in my inbox”.

Bard succeeded at some simpler tasks … But it also told me about emails that weren’t in my inbox.

– Kevin Roose, The New York Times

In response, Google’s Jack Krawczyk told Roose that, “Trial and error is still definitely required at this point” with Bard Extensions. Bard’s interface also prominently displays a warning that, “Bard may display inaccurate or offensive information that doesn’t represent Google’s views”, and there is a ‘Google it’ button designed to make it easy to double-check Bard’s responses. These points of friction, however, are not exactly part of the dream conversational copilot experience.

The other issue is that an unsuitable UX is still an unsuitable UX, and as GZERO’s Nover demonstrated, can leave the user baffled as to why the tool exists at all. Brands need to avoid falling prey to ‘copilot mania’ (no matter how optimistic the think-pieces are) and give careful consideration to what, if anything, a conversational UX would add to their product.

Does this feature need to take the form of a conversation with an ‘assistant’? If not, can it usefully take a different form – even if this might mean passing up a chance to jump on the generative AI bandwagon?