AI can’t do your Christmas shopping just yet – but next year might be different
- Written by Jon Whittle, Director, Data61
I’m a computer scientist and a bad Christmas shopper. Over the weekend, I wondered whether AI systems might be able to help me out.
Could I just prompt ChatGPT to pick a personalised gift for my cousin Johnny and have it shipped in time to reach him? Alas, the cheerful chatbot couldn’t help, telling me it “can’t make purchases or handle shipping directly”.
In the two years since ChatGPT launched, we have seen wave after wave of AI products and features promising to save us from mundane tasks. Yet so far, gift shopping is beyond them – apart from the odd built-in chatbot on a shopping site or app.
However, things may be different by next Christmas. One thing many experts expect in 2025 is the rise of AI agents: bots that can take actions on your behalf in the real world.
Agents are already here
An AI agent can do more than just suggest where you can get a Santa suit. It can buy it for you and have it delivered to your door.
And the vision for “agentic AI” is that teams of AI agents will work together. You would give your team of agents a prompt:
I’m cooking Christmas dinner this year. Find my closest Facebook friends, send them invites, make sure one of them is a chef and tell them to bring the turkey.
The agents would sort it all, without you ever having to lift a finger. Crucially, AI agents should have the ability to coordinate across multiple websites.
In fact, limited AI agents are already here. A report[1] by AI developer Langchain claims 51% of respondents to its survey already use AI agents in production.
In 2024, venture funds invested an estimated US$1.8 billion[2] in AI agent projects. Deloitte’s latest Global Predictions Report[3] argues 25% of companies that use generative AI will launch agentic AI projects in 2025.
Research firm Gartner predicts[4] that by 2028, 15% of day-to-day work decisions will be made by AI agents.
Agents for everybody?
We are also seeing agents for consumers beginning to emerge. These are systems that could potentially automate many browser-based tasks (including shopping).
In October, Anthropic – the company behind the popular Claude generative AI bot – released a “computer use[5]” feature that allows the AI to take over a user’s mouse and keyboard to browse and take actions on any website.
Education expert Leon Furze created a demo using computer use[6] to automatically browse to a learning management system, open the page for an assignment, create text for the assignment, and click the submit button. All done automatically from a single text prompt.
More recently, Google Deepmind released its own version, Project Mariner[7], which similarly allows an AI to autonomously navigate and carry out actions in the Chrome browser.
Both these systems are still early versions, with Project Mariner only available to a trusted set of testers. But they hint at what’s to come.
You can’t use either of these tools today to automate your Christmas shopping – at least, not easily. So what would be needed to make a truly useful Christmas shopping AI agent?
The technology exists
The technology side of a shopping agent is relatively straightforward. As a user, I might want to give a prompt such as
Send photo gifts to my family in England. Select some fun family photos from my phone, search for a website that does photo gifts, order appropriate gifts for each family member, and send using my address book.
Executing this would require multiple AI agents: one to find the photos, one to find the shopping sites, one to personalise the gifts, a credit card agent to buy them, and an address-finding agent.
Whether through computer use, Project Mariner, or some other AI agent platform, there is no technological reason why this can’t be done today.
The trust problem
However, there are two significant barriers to making AI agents useful.
First, and most obvious, is trust. Would you trust an AI agent with your credit card details?
Despite two years of advances in AI since ChatGPT, hallucinations – where the AI doesn’t know an answer and so simply makes something up – are still a problem.
A recent study[8] showed that even in AI programming – one of the most popular and valuable uses of AI – 52% of AI-generated answers to coding questions contained errors.
It only takes one error from the AI to send Aunty Molly’s gift to Uncle Joe. And let’s just hope it’s a harmless error such as poor gift matching, not leaking your bank account details.
What agents need to know
The second and less obvious barrier is that for AI agents to be useful, they need to understand context. Even with something relatively simple like buying gifts, context is everything.
I have years of knowledge about what my mother likes. I won’t always get it right, but I’ll do a lot better than a generic AI response. This knowledge is usually tacit and there’s simply no way ChatGPT can have access to the rich history of human interactions that lead to that perfect gift.
Having said that, AI bots are already recording information about their users. To prove this, just ask ChatGPT, “What do you know about me?” Depending on your settings, you might be surprised by the answer.
Perhaps at some point the AI systems we use regularly will know enough about us and our family that Christmas shopping can be fully automated.
But this year, I will still have to attend to it myself. Bah humbug!
References
- ^ report (www.langchain.com)
- ^ an estimated US$1.8 billion (www.globaldata.com)
- ^ Global Predictions Report (www.deloitte.com)
- ^ predicts (www.gartner.com)
- ^ computer use (www.anthropic.com)
- ^ demo using computer use (leonfurze.com)
- ^ Project Mariner (deepmind.google)
- ^ recent study (dl.acm.org)
