From Conversation to Cart
A useful retail agent does not merely answer a shopper. It turns uncertain human intent into a valid, available and governable commercial action. This fifteen-part series follows that journey end to end. Parts 1 to 3 set out the value journey: why the agent is the smallest part of the system, how a conversation becomes a basket, and why a recipe is not a basket. Parts 4 to 6 cover the catalogue and decisions: personalisation as a constraint system, the catalogue as the real knowledge base, and assigning decision rights before building agents. Parts 7 to 12 are the production-engineering arc: load-testing the shopping journey, what concurrency teaches you, tokens as a capacity dependency, model routing and failover, tracing one basket across every dependency, and the readiness and connection-pool details that decide whether the journey works. Parts 13 to 15 cover governance and operations: security controls that break the journey, read-only SRE agents, and a production-readiness checklist. The series follows Ubuntu Grocer, a composite: a fictional South African grocery retailer whose scenario draws on lessons from real enterprise retail-agent delivery. Its people, systems and figures are illustrative and describe no single organisation. Research claims were checked on 5 November 2026.
15 articles
- Part 1The retail agent is the smallest part of the systemA retail agent that writes a good meal plan has done the easy part. Part 1 of From Conversation to Cart argues that value comes from the system around the model: trusted context, bounded decisions, execution against retail systems, recovery, governance and operations. It also introduces the composite architecture the series follows.10 min read
- Part 2From conversation to cart: the basket is the outcomeA meal assistant can write a perfect answer and still fail the shopper. Part 2 of From Conversation to Cart traces the eleven stages from a vague request to a purchasable basket, shows where each one breaks, and argues for measuring the accepted basket rather than the conversation.9 min read
- Part 3A recipe is not a basketRecipes speak in cups and onions; catalogues speak in pack sizes and product codes. Part 3 of From Conversation to Cart shows why the two cannot be matched directly, and sets out the intermediate representation, unit roll-up and substitution rules that turn ingredient lines into a basket that makes sense.7 min read
- Part 4Personalisation is a constraint systemAllergies, dietary rules, budget, household size, brand preferences and the pantry are not flavour text for a prompt. They are constraints of different strengths, and the strongest must be enforced in code at every step that can change a basket. Part 4 of From Conversation to Cart shows how.9 min read
- Part 5The catalogue is the real retail knowledge baseA retail assistant can only recommend what its catalogue can describe, find and prove is available, and catalogue failures arrive looking like model failures. Part 5 of From Conversation to Cart sets out the product record an agent needs, the retrieval pipeline around it and why every field needs an owner.9 min read
- Part 6Design the agent boundary before the agentA retail assistant needs judgement in a few places and none at all in others. Part 6 of From Conversation to Cart compares three ways to orchestrate the shopping journey and assigns decision rights for search, recommendation, substitution, basket changes and checkout, action by action.8 min read
- Part 7Load-test the shopping journey, not an endpointA load test that fires one happy-path prompt at a chat endpoint measures the wrong unit and hides the real bottlenecks. Part 7 of From Conversation to Cart sets out a reusable scenario generator for retail-agent journeys and explains why the choice between open and closed workload models decides what the test can see.9 min read
- Part 8What thousands of concurrent shoppers teach youUnder rising concurrency, latency does not grow in a straight line and the bottleneck does not stay in one place. Part 8 of From Conversation to Cart explains the queueing intuition behind the latency knee, the order in which retail-agent bottlenecks typically surface, and a method for finding and fixing them one at a time.8 min read
- Part 9Tokens are a capacity dependencyTokens are usually discussed as cost. For a retail agent at peak they are also capacity: every journey consumes tokens and requests per minute against limits that are set per model deployment. Part 9 of From Conversation to Cart gives a variable-based capacity model, a worked example with illustrative numbers, and the decisions it should drive.9 min read
- Part 10Model routing without losing the customerA retail agent that routes every step to one model either costs too much or fails when that model does. Part 10 of From Conversation to Cart routes by task, sets priorities and circuit breakers, checks fallback quality and degrades the shopping journey one rung at a time without ever dropping an allergen.8 min read
- Part 11Trace one basket across every dependencyA retail agent's journey crosses a front end, a gateway, several models, the catalogue, a cache, a database and the basket. Part 11 of From Conversation to Cart shows how one journey ID and W3C Trace Context turn that path into a timeline an engineer can read in minutes.8 min read
- Part 12Readiness, connection pools and the hidden architectureThe parts of a retail agent that decide whether it survives payday never appear in the demo: readiness probes, connection pools, client lifetimes, cache limits and configuration. Part 12 of From Conversation to Cart shows how each one fails under load and gives the connection-limit arithmetic to do before the autoscaler does it for you.7 min read
- Part 13When security controls break the customer journeyA web application firewall can fail a retail agent in two directions: a negation on the wrong side quietly opens access, and an over-strict rule blocks a shopper's long, multilingual shopping list. Part 13 of From Conversation to Cart shows how to test both directions, roll rules out in stages, watch blocked legitimate traffic and roll back.7 min read
- Part 14SRE agents need read-only by defaultAn AI assistant that diagnoses a live retail incident is useful long before it should be allowed to fix one. Part 14 of From Conversation to Cart applies the approval-gated SRE loop to a Friday-evening basket failure, and shows why the dangerous remediation in retail is often the one that quietly changes what shoppers are offered.7 min read
- Part 15The production-ready agentic retail checklistA retail agent is ready for shoppers when evidence exists across twelve dimensions, from outcome and data to authority, capacity, security and operational ownership. Part 15 of From Conversation to Cart closes the series with a reusable checklist: what good looks like, the evidence to ask for, and the red flag that should stop a launch.9 min read
Working through any of this in your own organisation?
These articles come out of delivery work rather than theory. If the problems in this series look familiar, a short scoping conversation is usually the quickest way to work out what is worth doing first.
Talk to CloudNala