Context, Not Models: Where the Real AI War Will Be Won or Lost
Forget parameter counts. Move past model leaderboard to contextual intelligence: multimodel interfaces, agentic workflows, and invisible AI that just works.
TL;DR: Win With Context, Not Models
As LLMs and open-source models race toward parity, one truth is emerging: Intelligence is becoming cheap. Interfaces, workflows and adoption are the new moats.
Retailers & brands don't need smarter AI. They need simpler, more embedded, less interruptive AI which intuitively enhances the shopper experience while simultaneously improving operational efficiencies. This means efficient, more contextual workflows and better user interfaces. The best AI won’t feel like software. It will feel like magic, showing up in workflows and interfaces so seamlessly that you forget the model even exists.
From Model Performance to Usability: A Quiet but Massive Shift
For much of the past year, the AI conversation has been a leaderboard game: parameter counts, benchmark scores, token windows. GPT-5 vs. Claude 3 vs. Gemini 2.5. Actually, what started as just a short list has now grown into a long list of thousands of models, each vying for your attention and claiming to be the best.
But in the real world- on store floors, in merchandising cycles, in fulfilment centres, none of that matters. Retailers and brands don’t need “smarter AI.” They need simpler, more embedded, less interruptive AI that fits into the way people already work.
What matters on the ground is intuitive user interfaces and contextual workflows. A model can be powerful, but if the store associate can’t use it or if your shopper won’t wait for it, then what’s the point?
That means the focus has shifted to
Workflows that anticipate actions instead of adding steps.
Interfaces that feel invisible, not like another tool.
Infrastructure that silently optimises in the background.
This quiet shift from model performance to usability is now defining how retailers and brands are adopting AI in their respective organisations. This post explores how the AI advantage in real-world settings won’t come from better models but from efficient workflows embedded in better interfaces
Shift to Contextual Intelligence
In the real world, retailers don’t succeed because they picked the “smartest” model. They succeed when AI shows up in context, inside the flow of work, at the exact moment a decision is made, with just enough guidance to remove friction. That’s contextual intelligence. Not “How clever is your model?” but “Where does it live, what does it change, and will people use it without thinking about it?”
Three important factors are bringing this context to life.
Voice & Multimodal: the post-dashboard era for non-desk work.
Agentic Co-Pilots in Workflows: assistants that nudge the next best action right inside POS/OMS/CDP/CRM.
Invisible AI Infrastructure: intelligence woven into the stack so the workflow just… happens.
Let’s unpack each of these through a retail lens.
1. Voice & Multimodal: The Post-Dashboard Era
Typing is slow, and no one likes sitting at the desk looking at dashboards. If your associate is juggling customers, carts, and a radio, they’re not logging into a BI tool. The new interface is voice + vision: you talk, you show, you gesture, and the system understands.
How it lands on the ground
Store ops: A floor lead says, “Create a 2 pm facing task for Aisle 7; we’re out of 2-litre orange.” The assistant confirms, assigns, and pings replenishment. No app hunting, no forms.
Planogram & merchandising: A category manager points a camera at the shelf, asks “What’s off-plan and why?” and gets a visual map with suggested fixes: facings, price mismatch, missing label, poor adjacency.
Shopper UX: “Show me breathable running shoes under $120 that fit like my last purchase” returns a curated set along with the size guidance from prior returns and gait data captured in-store.
Retail associate demos: Show an item to a camera “What fertiliser for this plant?”, get an answer plus one-tap add-to-basket for BOPIS or delivery.
Why this matters: Non-desk workflows dominate retail. The fastest path from intent to action is not via a screen; it’s now with voice or gestures as multimodality is moving AI out of the back office and into the aisle.
2) Agentic Co-Pilots: Embedded Where Work Happens
The second layer is deceptively simple: Instead of “AI tools”, start shipping assistance where the work is happening.
Agentic co-pilots sit inside POS, OMS, CDP, or CRM. They perceive context (inventory, velocity, weather, promo calendar, ticket backlog), propose the next best action, and often execute with a human in or over the loop.
How it lands on the ground
Category & pricing: The AI Agent flags a pricing gap vs. local competitors, simulates promo uplift, and proposes a 7-day markdown ladder in a few clicks, then pushes changes to the price engine and shelf labels.
Growth & CRM: Inside the CDP, the agentic copilot highlights an emerging audience (“high-margin, low-frequency, coastal stores”), drafts a 3-step journey, and wires an A/B test to MarTech with holdouts and guardrails.
Service & Returns: A returns agent sees a summary of a shopper’s journey (purchases, chats, store visits) and the recommendation: approve refund + issue size-fit voucher. It's a one-click, policy-compliant action.
Catalogue Enrichment: Onboarding products, creating engaging descriptions, and enriching catalogues, all via multi-agent workflows that integrate LLMs directly inside existing content systems.
The above use cases are now getting prioritised amongst retailers & brands for agentic copilot applications. The majority of these agents are now integrated into the respective functional applications like POS, Order management, promotion engines, etc.
Why this matters: Co-pilots reduce switching costs. No new logins, no parallel systems. Just a right nudge, in the right system, at the right moment, creates a major impact.
3) Invisible AI Infrastructure: Workflows That Just… Happen
The third layer is the least visible and the most powerful: intelligence in the foundational pillars. No prompts. No UI. Just systems that learn and adapt, shifting from rules-based to signal-based execution.
How it lands on the ground
Predictive inventory: The network quietly rebalances stock based on store-level velocity, substitute risk, and last-mile constraints, much before a human would flag an issue.
Computer vision at the edge: Shelf-gap detection and price-label verification trigger tasks, not tickets. Compliance becomes continuous.
Fulfilment & last mile: OMS allocates orders to stores that can pick cleanly and deliver reliably based on live capacity, traffic, and SLA risk.
Ops reliability: Equipment-failure prediction schedules maintenance windows to minimise peak disruption; staffing plans adjust automatically to demand signals.
Why this matters: This is contextual intelligence at its purest: no training or no adoption campaigns. It’s just in the stack and makes the system smarter.
All this is great, but how do you build trust with your customers?
Models can be opaque, but interfaces are where trust is earned. That’s true for shoppers (recommendations, promotions, refunds) and for employees (pricing, tasking, performance cues).
Few things we can do to build trust.
Show your work (why AI is recommending this price/promo/product), give control (clear personalisation toggles and easy opt-outs), set visible guardrails (what the AI will and won’t do), and make escalation effortless with human hand-off plus context. Do this for shoppers (recs, refunds, promos) and employees (pricing, tasking, performance cues) in the exact moment of action.
Treat ‘tone’ as part of the product, and design for suggest-and-approve by default in high-stakes flows. If people can see the rationale, override gracefully, and reach a human without repeating themselves, they’ll use it again. In essence, trust becomes a feature, not a policy PDF.
However, this requires a lot of unseen work under the hood. You can't win this trust without orchestrating the logistics behind it. It requires plumbing: clean identity and user consent, high-quality product and store data, and composable event streams from POS, e-comm, sensors, and last-mile collating on a shared data platform, which the agent can “see.”
Wrapping this into a callable policy engine (pricing/refund/compliance), with a lightweight explainability service a (couple of lines per decision), and observability (latency, overrides, policy blocks, SLA breaches) further strengthens this foundational layer.
When you get this right, your AI feels consistent and safe at scale, and your brand is seen as respectful & reliable in every moment that matters.
So, What Does This Mean for the Three Stakeholders
For Retailers: You must stop asking, “Which model?” and start asking, “Which workflow?” Where will this sit? What will it replace? How many steps does it remove? What policy will it respect? How will we measure trust?
For SaaS vendors: Your moat isn’t the model; it is integrated inside existing systems: POS hooks, OMS overrides, CDP audiences, policy calls. Offer suggest-and-approve modes and enterprise-grade explainability out of the box.
For Investors: Models are getting commoditised. Value creation happens when there is usage & adoption. Look for products with embedded, policy-aware agents and quantifiable evidence of time saved and decisions improved. Interface defensibility > model bragging rights.
Contextual Intelligence as Brand
As interfaces take centre stage, they start to carry something deeper: your brand. The tone of the voice assistant, the helpfulness, the clarity of an explanation- they shape how customers and employees feel about you.
A brand used to be colour, copy, and campaigns. Now, it’s also the experience of your AI: how it speaks, what it knows, when it steps back, and how it earns trust. That is not a “feature.” That’s how customers will see you at any given moment
And the real AI war won’t be won on parameter counts, it will be won in moments: a cashier line that never forms, a promotion approved in seconds, a shopper who feels known (and respected), a shelf that never looks empty, an associate who ends a shift less tired.
These moments happen when intelligence is contextual, embedded inside the workflow, through interfaces that feel natural, and infrastructure that simply works. When that’s true, the model fades to the background, and the experience takes the lead.
What are your thoughts on this shift? How are you seeing the "interface war" play out in your industry? Share your insights in the comments below.