Building the Assistant Was the Easy Part
Retailers have spent the past year building AI shopping assistants to defend the customer relationship. As AI agents take over more of product discovery, the risk is that a third-party agent slots itself between you and the shopper, compares you on price alone and treats your brand as interchangeable supply. Bain puts the US agentic commerce market at $300 to $500 billion by 2030, between 15 and 25 per cent of online retail. We already know that the customer relationship is the most defensible asset a retailer has. Building an on-site assistant is one of the clearest ways to protect it.
The strategy is right. The problem emerging is what happens after launch. New Australian data shows these assistants failing in plain, avoidable ways and the reason is almost always the same. The assistant has been given a customer-facing job without being connected to the commercial rhythm of the business. Promotions are added and removed, products launch, prices move and the assistant isn’t wired into them. Its ongoing care fell into the gap between the team that built it and the teams that run the business day to day and it was never assigned to be cared for long term, by either.
The Evidence
Inside Retail’s new Online CX Index, research this year by the team at Humii, introduced a new reason shoppers abandon a purchase: chatbot obstruction. Its share of pre-checkout abandonment has climbed from 3.4 per cent at the start of 2026 to nearly 7 per cent six months later. For some retailers it now reaches 35 per cent, high enough that they would have been better off with no assistant at all.
The failures are small and ordinary, which is what makes them serious. An assistant insists it does not sell a product the business promoted in its own email that morning, then links to the item while denying it exists. Another cannot answer a straight question about a discount running that week. A third falls over on a simple spelling mistake. None of these is a model failure. Each is a maintenance failure. The promotion changed, the product launched, the price moved and nobody updated the assistant to match.
That gap has a direct cost. Mirakl finds that almost half of shoppers abandon a purchase when product information is missing or inconsistent. An assistant working from a stale knowledge base is a machine for producing exactly that inconsistency, at the moment of highest intent, in fluent and confident sentences. The sought-after product that does not exist is not a funny one-off. It is a data problem the assistant is now broadcasting to every customer who asks.
Why Ownership Falls Through the Cracks
Humii names the cause as ownership. The most important question they raise is who is accountable for it once it is live and changing every day? That is a structural question. The assistant is usually built by a digital or technology team as a project with a launch date. The knowledge it depends on, this week’s promotions, yesterday’s launches, the current price, lives with merchandising and marketing and changes every day. The build team ships and moves on. The commercial teams never saw the assistant as theirs. So, the ongoing care of it, the part that decides whether it is any good, belongs to no one. They never decided who does it.
From Project to Product
Fixing that starts with a change of category. Retailers have been running the assistant as a project and retail runs good projects. A finite budget, a timeline, a PMO to report to, a team that ships and moves on. That works for a website or a campaign. It doesn’t work for something that has to stay accurate every day after it goes live. A product is run differently: someone is accountable for how it performs in market, with authority across both the technical and commercial sides, resourced to keep going rather than to finish. Run as a project, maintenance, testing and monitoring have no home once the launch is done. Run as a product, they are the job.
Three Parts of Ownership
From there, ongoing ownership has three parts and each maps to where failures are most likely.
Maintenance. Running on business events rather than simply the calendar. The assistant fails right after a promotion goes live or a product drops, because that is the moment its knowledge is out of date. So it has to update on the same triggers that update the website, the email, your store teams. If marketing can push a campaign live, the assistant should know about it at the same moment, from the same source of truth. A separate knowledge base, refreshed on a monthly cycle, will be wrong at least every week.
Quality Assurance. Made a standing function rather than a one-off launch gate. Signing the assistant off once before go-live tells you nothing about how it behaves three campaigns later. It needs adversarial testing after every material change. Ask it about today’s promotion, use a typo, challenge it, the exact things Humii do when they stress-test these systems and watch them come apart.
Instrumentation. Ensures you have visibility of what you have built and how it shows up in the world. The failures in the research were found by outside mystery shoppers rather than the retailers’ own monitoring, because most retailers have little other way of knowing what their assistant is actually saying to customers. An owned assistant is instrumented, so obstruction can be measured as its own signal rather than read about later in someone else’s index. The retailers who treat that data as a live signal will fix problems in days. The ones who do not will hear about them from their customers, or worse, from an index like this one.
Run together, the three compound. A fault the QA process catches gets fixed once, because instrumentation makes sure it’s actually seen, and stays fixed. Maintenance keeps the assistant’s knowledge close enough to the business that adversarial testing keeps finding genuinely new gaps instead of the same ones each time. Run this way for six months and the assistant gets measurably more accurate, because every fix compounds instead of resetting.
Woolworths has built standing QA into the product itself. When it made its Olive assistant agentic this year, it identified eight distinct capabilities Olive has to demonstrate consistently, from pricing accuracy to compliance to actually doing what the customer asked. It then built a specialist “agentic judge” for each one, which runs against every response before it reaches a customer, as a continuous test of whether Olive is still doing its job. Run as a retail technology project, the quality check would have signed off at go-live and nothing after it. Woolworths made checking a permanent function with people assigned to it, tying it to the commercial reality the assistant has to keep pace with. That decision is the ownership.
Bain finds shoppers currently trust a retailer’s own agent around three times more than a third-party one and that lead is narrowing as consumers get used to outside agents. Every failed interaction spends down the one advantage you started with and funds the case for shopping through the very platforms your assistant was built to hold off.
So the questions for any leader running a customer-facing assistant are straightforward. Who owns it? Who updates it when a promotion goes live? Who tested it this week and who saw the result? If the answers are a shrug or a vendor’s name, the assistant is not defending the customer relationship. It is quietly spending it.
Building the assistant was always the (relatively) easy part. The hard part is the ownership through its life. A named owner, maintenance tied to the business, testing that never stops and monitoring that shows what the thing is doing. That work is unglamorous, but it is also where the whole advantage sits. The customer relationship is still the most valuable thing a retailer has and every conversation with your assistant is a small moment where a customer decides whether to keep choosing you. An assistant nobody owns turns those moments into reasons to choose someone else.
References
Bain & Company, “Agentic AI in Retail: How Autonomous Shopping Is Redefining the Customer Journey,” 13 November 2025. bain.com
Inside Retail Australia / Humii, “Online CX Index special edition: The rise of chatbot obstruction,” Mareile Osthus, 23 July 2026. insideretail.com.au
iTnews, “Woolworths gives agentic-powered Olive chatbot to its 200,000 staff,” 28 April 2026. itnews.com.au
Mirakl, “Why product content optimisation is critical for conversions,” February 2026, mirakl.com. Note: Mirakl’s own figure is qualitative (“almost half”), not a precise percentage; represented as such in the piece.