By Vihan Patel (Head of AI Solutions) & Sangeeta Vishwanath (Head of Digital) at Mantel
Key takeaways for business leaders
-
AI agents are transforming how customers discover, decide, and purchase products.
-
Analysts project a 25 percent drop in traditional search volume in 2026 as these agents take over the work of finding and comparing.
- AI agents help customers with their purchasing decisions by reasoning over the product data and reducing cognitive load, in owned third-party channels
-
Enterprises must optimise their structured data to ensure their catalogues are legible to generic consumer AI products.
-
This shift extends beyond retail checkouts to affect intake, advice, and service across all regulated industries.
-
Regulators are closely monitoring the security risks and compliance issues associated with autonomous AI actors.
Gartner’s 25% projection for the drop in traditional search volume in 2026, as agents take over the work of finding, comparing and ranking. Discovery, decision and increasingly the act of purchase itself are migrating away from the channels brands have spent two decades optimising.
What is agentic commerce?
Agentic commerce is a digital model where artificial intelligence software acts on behalf of a customer to find, compare, and purchase products. Instead of users manually navigating a store catalogue, these intelligent agents interpret user intent and interact directly with business systems to deliver tailored outcomes.
A generational shift in commerce is underway. Every major payments network, search engine and AI lab is investing in the same architecture: software agents that act on behalf of customers, discovering, deciding, transacting and servicing, across whichever surfaces those customers happen to use. For enterprises willing to engage with what’s actually changing, the opportunity is significant.
Done well, agentic commerce isn’t another channel to manage. It’s a chance to redesign the relationship between a brand and its customer around something the customer has always wanted but rarely been offered: genuine cognitive relief. An agent that remembers their preferences, understands their context, plans on their behalf, and acts only when they want it to. Less effort to be a customer. Better outcomes when they are.
For organisations, that translates into deeper customer relationships, higher lifetime value, richer behavioural data, and entirely new service moments that simply weren’t viable when every interaction had to fight for human attention. The companies that get there first will look back on the last decade of “personalised marketing” the way airlines look at fax confirmations, perfectly serviceable for its time, completely beside the point now.
From seller-mediated to customer-led commerce
For two decades, digital commerce has been seller-mediated. Brands invested in their own channels because that’s where they had control, over experience, over data, over the paths a customer could take through a catalogue. Channel strategy meant dragging traffic to your storefront: paid acquisition, SEO, retention loops, app installs.
Agentic commerce reorders that. The customer, instead of coming straight to your channel, chooses an agent instead – either one you own, or a third party agent like Chat GPT, Google AI mode or a meal planning tool that became a household’s default for weekly grocery shopping. The customer chooses the entry point. The retailer competes to be the answer the agent selects.
When an agent returns three recommendations, those three are the consideration set. Customers don’t discard them and start a manual search. If your catalogue is legible to the agent, you make the shortlist. If it isn’t, you don’t, and you may not even know you’ve been excluded. The boring fundamentals like quality, availability, fulfilment, service all get re-priced.
Where the market sits today
Most “AI in commerce” still serves the seller
It’s worth being honest about where current solutions land. Five categories dominate the conversation, and each is solving a real problem — but none yet holds the buyer’s interest end-to-end.
- Amazon · Alexa
Voice shopping, Subscribe & Save. Locked to Amazon’s ecosystem and optimised for Amazon’s margin – not portable across the customer’s life. - Instacart
AI-powered delivery and search. Logistics-focused; no meal planning, no trust escalation, no continuity beyond the order. - HelloFresh · Gousto
Subscription meal kits. Rigid menus, ~70% annual churn, limited adaptation to how households actually live. - ChatGPT · Perplexity
Generic AI assistants. No commerce integration, no persistent preference memory, no trust model for delegated purchase. - Retailer chatbots
FAQ and product search bolted onto storefronts. Scripted, session-based, designed to defend the funnel rather than represent the customer.
The position that’s still open: an agent that holds the customer’s intent across planning, comparison, execution and service. That is the agentic opportunity worth building toward.
The technical landscape
Technical requirements for AI agents
Your product page was designed for human eyes. In today’s world, the primary user is now a machine. The main question that you need to solve for: can agents see you?
Can agents find me?
Discovery – making your catalogue machine-readable
The web’s discovery layer was built for keyword crawlers and human attention. Agents need structured product, inventory and pricing data, accurate in real time, with a semantic model the agent can reason about. The protocol layer is converging on open standards, but the precondition is still the same boring thing it has always been: clean, structured catalogue data.
Surfaces are emerging where buyers already spend their planning time. ACP is live in ChatGPT today. UCP is forming around Google AI Mode. Klarna’s app has been quietly building agentic shopping for months.
Can agents buy from me?
Payment – settling a transaction the agent initiated
The transaction layer assumed a human at the keyboard, clicking confirm. Agents need credentials they can use without exposing the customer’s underlying payment instrument, with delegation, scope and revocation built in. Every major payments network is now in this race.
The strategic question for sellers is whether your checkout, fraud and settlement stack can accept agent-initiated payments at all without falling back to human verification.
Can my agents negotiate?
Inter-agent – your agents talking to other agents
If you have your own agents for sales, service, underwriting, fulfilment etc, how do they talk to the customer’s agent, or a partner’s agent? Multi-agent protocols are the layer where competitive advantage will compound: your agent representing your customer’s interest against a supplier agent, your service agent resolving issues with a partner agent, your underwriting agent talking to a broker agent.
Most businesses will not need to build for this in 2026. However, almost all will need a credible answer to it by 2028.
Channel strategy
Comparing on canvas and off canvas strategy
Once the customer’s agent picks the entry point, every enterprise ends up dealing with two distinct surfaces.
On-canvas
Your owned agentic experience.
What it is
A branded agent inside your app, site or product, designed by you, deeply integrated with your data, loyalty, fulfilment and policy.
You control
- The brand, tone and feel of the relationship
- The interaction model (beyond chat if you choose)
- The depth of personalisation and proprietary data
- How to upsell, cross-sell and pace the journey
Insurance home health check · Hardware store DIY advisor
Off-canvas
The customer’s agent, on someone else’s surface.
What it is
Consumer AI products – ChatGPT, Claude, Perplexity, Google AI Mode – reading your structured data and presenting you to the customer alongside competitors.
You participate in
- Discovery and shortlisting decisions
- Product comparisons against rivals
- Increasingly, end-to-end checkout (ACP, AP2)
- Whatever conversation the agent decides to have
ChatGPT Shopping · Google AI mode · Perplexity Shopping
The framing is sometimes presented as a choice. It isn’t. Customers are starting their planning, comparison and increasingly purchase journeys inside consumer AI products, and that pattern is accelerating. Off-canvas presence (being legible to ChatGPT, Claude, Perplexity, Google AI Mode) becomes table stakes. If your catalogue, pricing and policies aren’t readable to those agents, you don’t appear in the consideration set, and nothing tells you you’ve been excluded.
On-canvas is where the more interesting strategic question sits. A branded agent earns its keep when it offers something a generic consumer AI cannot, exceptional personalisation grounded in first-party data, interfaces that go beyond chat (interactive canvases, calculators, generative UI), or genuinely immersive experiences like Suncorp’s Haven, where proprietary risk data lets a customer’s home describe its own bushfire and flood risks. Whether to invest there is a real choice. Whether to be readable off-canvas is not.
Industry view
Retail moves first. The cart is the smallest part of the prize.
Retail will lead. The dynamics are visible (search, shortlist, checkout) and the technical layer (ACP, AP2) was designed against this shape. So when most coverage uses the phrase “agentic commerce,” the picture is a cart with an AI on top.
That picture undersells what is actually happening in every other industry. In regulated and complex categories, the cart barely exists. The high-friction moments are intake, advice, comparison and service and agents rewrite all four parts. The strategic mistake right now is reading the retail strategy and assuming it generalises.
Retail & grocery
The visible cart
The risk for incumbents: the agent’s shortlist is the consideration set. Brands optimised for human shopper psychology, such as clever copy, hero imagery, store layout, find that none of those levers reach the new buyer.
Insurance
Intake, not checkout
Insurance has almost no cart. The friction lives in intake, comparison and claims. Agentic AI rewrites all of it:
- Home contents valuation by photo: point a phone at a room, the agent estimates a defensible sum-insured in minutes rather than the survey-and-spreadsheet ritual that no one completes
- Cover comparison against the customer’s actual conditions, dependants and claim history rather than generic personas
- Pre-claim triage that explains what is and isn’t covered before the customer files, collapsing the cycle that drives most service complaints
- For private health insurers: gap-cover lookup, Informed Financial Consent surfaced before specialist appointments, extras optimisation and Medical Gap Scheme guidance
Banking
Aggregation and advice
The agent becomes an advisor and a switcher. Continuous comparison of products against actual usage. Eligibility pre-check before a formal application. Switching support. Intake for loans and cards that doesn’t require the customer to know which product they want.
The bank that is most legible to the customer’s agent and has the clearest products, cleanest APIs, fastest pre-approval, most honest fee disclosure will capture the customer who would never have shopped on their own. The bank that hides behind bundles and opaque pricing will find itself de-shortlisted.
Energy & utilities
Continuous comparison
Plan optimisation moves from an annual chore to a continuous service: agents compare tariffs against real consumption, model time-of-use scenarios, switch when worthwhile, coordinate demand response.
The bigger opportunity sits one layer up. Energy retailers are being asked to guide customers through the cost and complexity of electrification (solar, batteries, EVs, heat pumps, hot water, charging tariffs). A household making the right decision across all of those is doing structured analysis most people can’t or won’t. Agents will absorb that work, and the retailer who can offer the guidance which is grounded in real consumption data and honest about trade-offs will capture the customer at the moment of biggest spend.
“Cart creation is a high effort activity in retail and an obvious target for an agentic solution. Other industries have other kinds of friction points, whether it be selecting the right product or processing complex information. Agentic commerce solutions should focus on solving for those with helpful and novel utilities, not just creating a conversational check-out.”
Vihan PatelHead of AI Solutions | Mantel
The maturity arc
What “personalisation” means when the agent gets smart enough
Most enterprises think of personalisation as recommendations tuned to past behaviour. That is table stakes, and it is largely where the market sits today. The real impact is the agent learning why the customer is asking, in the moment, and adapting its entire operating mode in response. We are seeing three levels forming:
Authenticated access
Knows who you are
Conversational interface over siloed user accounts. Static tool calls. Highly reactive. Zero initiative. The agent waits for instructions and returns results. This is where most “AI shopping assistants” actually live, regardless of how they’re marketed.
Baseline · Common today
Preference memory
Knows what you like
Leverages settings, history and saved defaults. Suggests routines, warns on rules. Familiar and convenient – but rigid. The agent personalises outputs but doesn’t change how it operates. Increases retention; doesn’t change the experience.
Most current solutions
Adaptive intent
Knows why you’re asking
Real-time semantic analysis of intent. The agent dynamically reorders which APIs to call, how to weight constraints, and what tone to use. Same agent. Different operating mode. Chosen by inferred intent, not selected by the user.
6–12 months out
Level 3 is where the agent becomes a revenue-generating partner rather than a cost-saving interface. It is also where the engineering, conversational design and guardrails become substantially harder. Most enterprises are anchored in Level 2 today and assume that’s the destination. It isn’t.
The trust layer
Trust stops being marketing. It becomes engineering.
Agents expand the surface area of what can go wrong. They also make trust observable in ways human-mediated commerce never was. Every recommendation, substitution, payment and service interaction the agent takes is logged, traceable and auditable. That is a gift to the brands that mean it. It is a reckoning for the ones that don’t.
There is a tendency to treat AI risk as ‘just another technology’.
Recently, APRA wrote to every regulated bank, insurer and superannuation trustee in Australia. The letter named agentic AI explicitly, and it should be on every CRO’s desk this morning. Three lines from it matter most:
“Manipulation or misuse of autonomous AI agents” – listed alongside prompt injection and data leakage as a primary new attack pathway
“Identity and access management capabilities have not yet adjusted to nonhuman actors such as AI agents”
Boards “are still developing the technical literacy required to provide effective challenge and oversight” on AI risk
For regulated entities, agentic isn’t a commerce strategy decision. It is a prudential one. The board literacy gap APRA flagged is the same gap that determines whether an agentic deployment ships safely, or doesn’t ship at all.
Outside regulated industries the pressure is gentler but the dynamic is the same. When the agent is the customer’s interface, every action your systems take is a small repeated test of whether you are acting in the customer’s interest. Trust is no longer something a brand claims. It is something an agent verifies, transaction by transaction.
The strategic capabilities worth investing in
Prioritising the right agentic surface.
The temptation is to launch a flagship “agent” and call it a strategy. That is the most expensive way to discover that the entry ticket is the same boring thing it always has been: clean, structured data with real-time inventory.
The infrastructure horizon
Where this is going underneath.
Most of the conversation today is about agents and apps. The deeper shift we’re seeing is in the infrastructure underneath, as well as how the players building it now define how the next decade of commerce works. We’re seeing four layers rising:
Real-time API marketplaces
Commerce moves from static catalogue feeds to live programmatic access to retailer capabilities. Agents query for availability, pricing, delivery windows and services like installation or assembly. Real-time comparison and fulfilment orchestration across retailers becomes routine.
Agent-to-agent commerce
Customer agents interact directly with retailer agents to negotiate and execute transactions. The customer’s agent requests options, negotiates delivery timing, assembles bundles and completes checkout. Retailers compete to be the agent’s preferred supplier.
Personal context infrastructure
Identity-linked data layers let agents access a user’s preferences, history and constraints across services- purchase history, brand preferences, budget, household profile. Highly personalised recommendations and automated repeat purchasing become the default, not the differentiator.
Agent trust & identity
Infrastructure that lets agents transact safely on behalf of users: agent identity and permissions, verified merchant credentials, payment delegation, policy enforcement. This is what enables trusted autonomous purchasing – and what regulators will increasingly require.
None of these layers exists just yet in mature form. But they’re not far off. The opportunity in agentic commerce isn’t to build a flashier chatbot, but to build something genuinely new: a relationship with customers that is more useful, more trusted, and more durable than what current digital channels can support. The organisations that approach it with that ambition, in the right sequence, will define the next decade of how customers and brands meet.
“Trust is earned in increments, through demonstrated value. An agent that eases a customer's decision-making burden becomes the preferred channel.”
Sangeeta VishwanathHead of Digital | Mantel
See how we’re helping businesses scale with AI-first solutions
