Design assistant conversations and relevant experiences people can understand, correct and control. The engagement connects investigation to implementation, with a baseline, reviewed changes and clear acceptance criteria.
What’s included.
Three steps.
Understand
Agree the task, baseline, source systems and delivery boundary.
Build
Develop and review the code, configuration and integrations in scope.
Verify
Test representative cases, compare with the baseline and hand over.
Recognise the challenge?
Start with the problem you see. Explore what we can change, then open the implementation notes for the detail.
01AI assistant and chatbot usability
What you may be seeing
- Customers do not know what the chatbot can help with.
- Answers are too long or difficult to compare.
- Customers cannot act on a useful recommendation.
- The conversation loses context or makes editing needs difficult.
- It is hard to reach staff when the assistant fails.
We design an assistant people can understand, correct and leave when they need a person instead.
Implementation & verification
We redesign the assistant interface around supported customer decisions and tasks. We code clearer starting prompts, concise evidence-backed answers, structured product/result cards and useful next actions. We add editing, recovery, loading and handover behavior that matches the assistant's actual capabilities. We test full conversations including wrong assumptions, failed tools and a customer request for human support.
How we verify the work
- Make supported jobs and evidence visible before asking customers to converse freely.
- Use structured comparisons/actions when they are clearer than long generated prose.
- Test correcting needs, tool failure and human transfer with preserved verified context.
- Measure useful task completion, correction effort and successful handover rather than chat length.
Tools considered for this work
Delivery: conversational UX + frontend/backend integration development. Design: Figma conversation states and realistic task scripts. Implementation: existing assistant UI, tool APIs (controlled connections between applications) and permitted commerce/account actions. Support option: Intercom Fin workflow handover or custom ticket integration. QA: conversation evaluation set, keyboard/mobile tasks and action-outcome events.
02Personalized website experiences
What you may be seeing
- Different customer needs receive the same generic page.
- Visitors cannot quickly find information relevant to their situation.
- Returning customers cannot continue a saved task or reorder easily.
- Personalization shows irrelevant products or the wrong region.
- We cannot tell whether tailored content performs better than a useful default.
We tailor useful parts of the experience to the customer context, with clear rules and a way to measure whether the change helps.
Implementation & verification
We define useful page adaptations from stated needs, permitted session context or reliably known account information. We code segment rules, approved content variants and eligible recommendations while providing a useful default for unknown visitors. We synchronize account-aware next actions with real saved tasks, product availability and service/location rules. We test rule accuracy and responsiveness, then use a randomized holdout where feasible to evaluate incremental business outcomes.
How we verify the work
- Start with distinct needs and reliable signals; prefer explicit customer selections when appropriate.
- Separate display targeting from authorization and verify every account-specific lookup.
- Check default/fallback behavior, cache boundaries, flicker and wrong-segment assignments.
- Measure incremental qualified completion/revenue versus a holdout plus relevance and performance guardrails.
Tools considered for this work
Delivery: frontend/backend + targeting/analytics development; platform configuration where sufficient. Rules/delivery: GrowthBook/PostHog flags or custom server-rendered targeting. Recommendations: Shopify Search & Discovery/API (a controlled connection between applications) or a suitable existing search engine. Content: current CMS with approved segment-specific variants. Measurement: exposure/outcome events and business records; default-experience holdout.
Useful AI begins with the experience.
We design the questions, evidence, correction controls and handover behind an assistant. Tests include uncertainty, missing knowledge and requests the system should not carry out.
Explore the technical toolkit
Selected around your existing stack, licences, permissions and agreed scope.
Define what better means.
Completion
Can people finish the intended task correctly?
Clarity
Do customers understand the offer and the next step?
Inclusion
Verified keyboard, screen-reader and mobile task behaviour.
Targets follow the baseline. These are measurement areas, not promised results.
Engineering terms, explained
- Grounded AI / RAG
- An assistant retrieves relevant approved evidence before generating an answer. Access controls and evaluations still need to be engineered and tested.
- Idempotency
- Retrying the same operation does not create an extra order, payment or record.
- Controlled experiment
- A concurrent comparison of assigned variants, with a defined outcome, observation window and decision rule.
- Core Web Vitals
- Measures of main-content loading, responsiveness to interaction and visual stability.
Technical references: Web Vitals · Azure document access · Structured data guidelines.
Quick answers.
Do you deliver designs or working software?
Both can be included. We can research and prototype, then implement the approved interfaces in your application with your engineering team. The proposal specifies the delivery boundary.
Will you redesign the whole product?
Only if the evidence supports it. We can improve one important journey and extend reusable patterns from there.
Do you work beyond ecommerce?
Yes. This work applies to software products, customer portals, service businesses, booking systems, internal tools and digital commerce.