Introduction
It’s late in the year, Christmas is approaching fast, and for many people the annual gift-buying scramble is reaching its peak. Consumers juggle gift preferences, budgets, purchase channels and delivery-date guarantees. For both retailers and shoppers, it’s the busiest, most pressured time of the year.
All the while, around the world, the retail sector is undergoing a revolution – driven not by customers but by machines. As shoppers increasingly lean on AI for discovery and purchase decisions, inbound traffic from generative-AI assistants to UK retail sites has ballooned by 1,200% over the past year (1). Research shows that globally around half of all web traffic now originates from bots rather than humans (2), while surveys of US consumers found that 60% plan to use AI to assist with shopping in the future (3).
For those of us at True who spend our time advising retailers on the latest technology, these facts raised an obvious question: Can a bot do all of my Christmas shopping for me? And if so, what implications would that have for retailers and brands?
We decided to find out…
The Experiment
We created 5 hypothetical personas and gave each of them a profile consisting of demographic features, Christmas values, and gift preferences. The idea was to provide sufficient information to allow an agent to guess what these personas might like for Christmas.
We agreed some guardrails for the search: the gifts must all fall within a total budget (£250), be purchased at their lowest costs, sourced from unique retailers, and be guaranteed to arrive before 23rd December.
Finally, the experiment would test some differing levels of human involvement in the shopping process, ranging from user-led to fully autonomous:
What actually happened -The Results
Level 1: Conversational Shopping (Human-led, AI-assisted)
At the lowest level of automation, the AI behaved most like a capable shopping assistant. It asked sensible clarifying questions, quickly built a picture of taste and constraints, and translated vague preferences into concrete product directions.
Where it worked best was in sense-making. It surfaced products, explained why they might suit a person, and spoke to some trade-offs around price, delivery, reviews, and sustainability. Great.
The main limitation was flow. The AI often narrowed too quickly, moving to specific products and next steps before fully exploring the category or getting my broad feedback. When deeper research was triggered, it sometimes swung too far the other way — taking a long time to return results that were thoughtful and well-justified but not always proportionate to the time spent.
Execution was mixed. Multi-retailer discovery worked well, but product imagery and links were not always reliable. Though the agent was able to identify the potential for discount codes, search for those codes, and test them on the checkout page, which looked, frankly unsettlingly conscious activity.
Overall, Level 1 worked well as a thinking partner. It reduced cognitive load and got me to a good recommendation quickly, but still required a human to manage pacing, verify details, and sequence the task effectively.
Level 2: Semi-Autonomous Agentic Commerce (AI-led, human-in-the-loop)
Level 2 was surprisingly effective. With richer context upfront, a few more clarifying instructions around preferences, and fewer interruptions, the AI handled end-to-end reasoning much more coherently. Broad product category selection, shortlisting, and comparison were generally strong, and it comfortably worked across multiple retailers rather than defaulting to a single platform.
This was also where the system began to show its value as a filter, not just a recommender. It reduced noise, surfaced credible options, and explained trade-offs like it had done in the more conversational-style of the previous experiment. In several cases, the final shortlist was good enough that the human role became one of approval rather than decision-making.
Checkout automation varied by retailer. Links broke. Images didn’t always represent the actual product.
Most interesting to me was the fact that, despite a clear instruction to find the cheapest retailer to buy the recommended product, the agent ignored this request and prompted to purchase from a more ‘reliable’ retailer (£2+ than the cheapest option, for the exact same product). The implications of this for retailers are huge and are explored in the implications section below.
Level 2 felt close to something useful, but not yet reliable. The intelligence was there. The slick execution layer wasn’t.
Level 3: Fully Autonomous, End-to-End Agentic Commerce (AI-led, minimal human involvement)
At Level 3, the experiment removed the safety net of requiring the agent to confirm with me before progressing. The AI was authorised to choose, proceed, and push as far into purchase as possible without asking for any confirmation.
What stood out was how confidently it made decisions. It concocted an elaborate assessment framework which quantitatively assessed products against elements it thought the persona might value. Category choice, product selection, and retailer preference were sensible and conservative, favouring established brands, clear descriptions, and predictable delivery. This wasn’t creative thinking, it was structured risk minimisation.
And then it hit a wall.
With my payments and delivery information preloaded into the prompt, the agent was able to partially complete the check-out fields, but completing the payment transaction still remains the hard boundary.
Identity checks, checkout UX assumptions, and payment authorisation are, for now, still a human-activity. The AI didn’t lack the intent, it knew the ask, but the technical systems it was navigating weren’t designed for non-human actors. Yet.
Level 3 failed in its main purpose; the ‘single prompt purchase’. This is because the surrounding infrastructure still isn’t ready for autonomous transactions. Many of the technical, regulatory, and platform foundations needed to make this possible are being developed rapidly, but they haven’t yet converged into something reliable enough for true end-to-end autonomy. We are confident this will be possible before next year’s festive frenzy.
The Gifts For Each Persona
My Partner Sophie - The Conscious Creative Parent
Gift: Lefrik Scout Stripes Backpack - Olive Green from Lefrik, bought from KJ Beckett
My Mum Mary - The Sentimental Crafter
Gift: Friday Night Pottery workshop from Belfast Ceramics Studio
My Brother Conor - Fashion-Forward Teenager
Gift: Beard Kit from Apothecary 87
My Niece Evie - The Magic-Loving Harry Potter Fan
Gift: Hermione (Harry Potter) Fancy Dress bought from Character.com
My Neighbour Adebayo - The Design-Savvy Modern Homeowner
Gift: Luxury Chocolate Selection Box from Fortnum & Mason
What Worked vs What Broke – Cross-Cutting Patterns
What Worked
Understanding people, not products
Given only light context, the AI was consistently good at inferring taste, lifestyle, and intent, and often explained why something would suit a person, not just what it was.
Reducing cognitive admin
Its biggest practical win was filtering. A long-list of messy options quickly collapsed into short, decision-ready lists and recommendations, saving time even when a human still made the final call.
Retailer-agnostic discovery
The AI naturally searched across lots of brands and retailers, often surfacing credible options beyond the obvious names. Notably, the first results weren’t always the cheapest, suggesting clarity and metadata matter more than price in agent-led discovery.
What Broke
Premature narrowing
Unless explicitly guided, the AI tended to converge too quickly, optimising early choices rather than fully exploring the category landscape.
Fragile retail execution
Broken links, misleading imagery, inconsistent delivery information, and discount-code loops regularly interrupted otherwise strong decision-making.
Checkout as the hard boundary
No matter the level of autonomy, payment and identity remained the point where human intervention is unavoidable, for now.
Implications
For consumers
How to use this technology well, today
Agentic shopping works best when treated as a series of steps, not a single request. Let the AI help you explore, narrow, and compare, but the experiment highlighted that the best results come from being deliberate about the order in which you want to do those things and the ranking of your trade-off preferences. Early framing strongly influences what gets surfaced later.
For now, the sweet spot is partnership. Use AI to reduce effort and improve decision quality, but expect to stay involved along the way (verifying details, sanity-checking claims, and completing checkout). Full automation is coming, but it isn’t the most reliable mode yet.
For brands
How products get discovered will change before where they’re sold
As AI increasingly mediates discovery, having product/service information legible to machines matters as much as being compelling to humans. Products with clearer descriptions, better imagery, and richer contextual signals were consistently easier for AI to reason about and recommend.
More importantly, intent matters more and more. Brands that clearly articulate what problem a product solves — and for whom — are more likely to be surfaced when an agent is trying to complete a mission, not just match keywords. Over time, structured, mission-based product information may become a meaningful source of advantage, even in channels brands don’t own.
For retailers
Execution, not intelligence, is the new battleground
The experiments showed that AI reasoning is already strong, but retail infrastructure is patchy. Checkout complexity, brittle discount logic, inconsistent delivery metadata, and poor or inconsistent product imagery regularly broke otherwise good journeys.
As agents take on more of the shopping process, retailers with simpler, more predictable, and more machine-readable purchase flows will increasingly become default endpoints for AI-driven demand. In this world, being “agent-friendly” may become as important as being mobile-friendly once was.
Closing
Agentic shopping already works well to take some of the stress out of exploring and can help us find gifts we otherwise might not have considered, especially when time is short and Christmas stress is high.
But gift-giving shouldn’t be about optimisation. The value lies in intent, thought, and care – things AI can support with, but hopefully not replace.
The technology will keep improving. As it does, the more interesting question isn’t what bots can buy for us, but what activities we choose to keep human. Christmas feels like a good place to start drawing that line.
Merry Christmas.
True.
Sources: