Building for Human Customers and AI Agents

Written By Emily Wilkinson

Short on time?
Search is becoming conversational and answer‑led, AI agents are starting to shop on behalf of people, and stores are rich in signals but poor in orchestration. For Product teams, that means three priorities this quarter:

  1. tune your catalog and APIs for agents as well as humans.

  2. start with clearly defined business pain and measurable outcomes (not tools).

  3. wire store and ops signals back into the product cycle so we act on what matters.

Source: Sora

Discovery has moved from keywords to conversation. Design your product surface accordingly

Traditional, keyword‑driven search volume is projected to drop ~25% by 2026 as people and assistants turn to AI chatbots and virtual agents for answers. That’s not a vibe; it’s a forecast Product teams should plan for.

At the same time, AI‑driven referrals to retail sites have exploded, with Adobe reporting huge year‑over‑year spikes during peak season and analysts warn that “answer layers” can intercept intent before a click ever reaches you.

If your product pages aren’t machine‑readable and your key facts aren’t easy for models to parse, you risk becoming invisible even if you “rank.”

What to do

  • Ship a “Bot‑Readable” Product Profile for your top SKUs/flows: unambiguous specs, availability, delivery windows, returns, warranties, sustainability claims exposed in structured data (schema.org), JSON endpoints, and documentation that LLMs can quote without guessing. Source.

  • Instrument answers, not just clicks. Track questions you see in support, search logs and chat. Then close content gaps (FAQs, comparison tables, sizing/compatibility guidance) so humans and agents can resolve the same jobs quickly.

Bots are becoming customers. Treat “agent readiness” like a first‑class requirement

AI shopping agents aren’t sci‑fi anymore. McKinsey expects AI agents to mediate trillions in global consumer commerce by 2030, and the industry is blending around open protocols that let agents read, negotiate and transact safely. In parallel, platform players have started announcing standards and partnerships to make agentic checkout real. For Product, that’s a signal: your site and services must be open, verifiable, and transactable for software clients and not just browsers.

Agent‑readiness checklist (Product & Engineering):

  • Catalog API that’s precise and resilient (ids, variants, stock status, delivery SLAs, promotions) with rate limits and fallbacks an agent can understand.

  • Transparent policies (returns, price‑match, data usage) in machine‑readable form removes ambiguity for agent decisions.

  • Consider emerging protocols as they stabilise. Track how your platform partners are implementing them, and design adapters rather than bespoke one‑offs.

  • Guardrails and observability: log agent behaviours, permissions and tool access; define what an agent may do (quote, reserve, pay) and when to escalate to a human.

Start with business pain, not tool excitement

I’m a big believer that good Product work begins with a crisp “so what?”. In practice, that means linking every AI/agentic idea to a specific, measurable problem and aligning teams through OKRs and tight feedback loops. (I’ve written about how OKR coaching pushes Product teams from activity to outcomes and why cadence beats intention.)

A simple framing that works

  • Problem to solve: “40% of ‘Where is my order?’ contacts happen even though tracking exists.”

  • Who’s affected / impact:
    Contact centre load, repeat customer effort, trust in the brand, NPS, cost‑to‑serve.

  • Smallest useful experiment:
    Proactive order status updates + clearer delivery milestones in My Account + one‑click access to returns and support from the order timeline.

  • Evidence of success:
    Weekly reduction in WISMO contacts, faster return completion, improved My Account NPS, and fewer customers dropping out to contact support.

Stores are over‑instrumented, under‑orchestrated

On the shop floor, you’ve got sensors, cameras, and Radio Frequency Identification (RFID) everywhere; what we often lack is orchestration. Turning those signals into actions for design, merchandising and product experience. RFID adoption is now mainstream among large retailers, and smart fitting rooms show how try‑on/abandon data can inform size curves, content, and availability decisions. Build that loop back into Product, not just Ops.

Use the signals you already have

  • Try‑on but not purchased? Trigger a backlog item: improve imagery, add fit notes, or adjust variant defaults.

  • Frequent size swaps? Feed sizing guidance or recommendation logic on the PDP and in agent‑readable attributes.

  • Stock accuracy gaps? Prioritise RFID/real‑time inventory for omnichannel promises (ship‑from‑store).

Data minimalism

More data doesn’t mean better decisions.

Product teams should test whether a data set actually changes the forecast or the decision and look to strip out noise that distracts teams from acting. Focus on the smallest set of signals that predict the outcome you care about (demand, conversion, returns), then iterate. (This mindset is echoed by multiple industry analyses on agentic commerce and by my own OKR practice: evidence over vanity.)

A 30/60/90‑day plan for Product leaders

Days 1–30: Baseline & quick wins

  • Run a discovery audit: How do we appear in conversational queries and AI overviews? Where are facts missing or ambiguous? Prioritise top 50 products/flows.

  • Ship “bot‑readable” facts for one high‑volume category.

  • Define two outcome‑based OKRs tied to conversion/returns/search support, with weekly check‑ins.

Days 31–60: Open up and observe

  • Expose a read‑only catalog & availability API to partner agents; add explicit usage/permissions.

  • Instrument agent traffic separately and set alerts for behaviours that need escalation.

  • Pilot a fitting‑room → PDP feedback loop (abandon reasons → content & size guidance updates).

Days 61–90: Scale what works

  • Add agent‑safe actions (reserve, apply discount, complete checkout) for one category; design fallbacks to human.

  • Extend structured data and APIs across your next two categories; harden SLAs and error messaging.

  • Review OKRs; keep what moved the metric, drop what didn’t.

Why this matters now

  • Search is fragmenting and answer layers are rising. Waiting means losing surface area in discovery.

  • Agentic commerce is maturing, with protocols and ecosystem momentum you can hook into rather than build from scratch.

  • The data advantage lives in your store signals. Connect them to Product decisions and you’ll out‑learn competitors who only optimise ad spend.

 
Sources, for the curious
Emily Wilkinson https://www.emyilwilkinson.co.uk
Next

What OKR Coaching Taught Me About Better Product Thinking (and Better Teams)