Three legacy desktop tools, thirteen years of workarounds. Redesigned into one AI-assisted web workbench for the Safety Data Sheet department.
The department didn't need faster tools. It needed one system that remembers what happened, and a customer request form that stops creating the work in the first place.
The SDS department ran on three isolated desktop tools (Requester, Handler, Entry Tool) with no shared data and no automated handoff.
Every action took 30–60 seconds, so specialists kept three windows open just to keep working.
Nothing was traceable. Teams chat and personal Excel trackers had become the real system of record.
Most duplicates and wrong-language sheets were created upstream, in the customer request form nobody had questioned.
What a specialist opened every morning. Three unrelated applications, none aware of the others, all real production screens.
Seven contextual interviews across Requester, Handler and Entry Tool roles, plus task shadowing and system walkthroughs over 4–5 weeks.
Heuristic evaluation, service blueprint, journey map and a 17-item issue register scored by impact, effort and root cause.
Remote workshops with the department manager, delivery lead and data engineer to converge on one unified architecture.
Mid-fi wireframes, then an AI-assisted lifecycle prototype tested with the same users who were interviewed.
Design system, UI specs, functional-action mapping and a phased implementation roadmap.
Contextual inquiry: users shared their screens and narrated every workaround, so I could map the flow as it happens, not as documented.
| Role | Tools used | Focus |
|---|---|---|
| SDS Department Manager | Requester + Handler | Intake, product creation, customer links |
| SDS Administrator | Requester + Handler | Supplier communication, uploads, assignment |
| SDS Coordinator | Requester + Handler | Product matching, new registrations |
| SDS Support Specialist | Requester + Handler | Supplier search, SDS updates, emails |
| SDS Administrator (key accounts) | Requester + Handler | Large international customers, link maintenance |
| Data Entry Specialist | Entry Tool | Validation and publishing to iChemistry |
“It's 2025, and we still work in tools that look like Windows 98.”
“We have to open three Requesters at once just to work efficiently, one to register, one to match, one to search.”
“If I upload a document, the system doesn't record my action unless I manually requeue it.”
“AI helps, but we still have to manually review every critical section, and it's tiring.”
Work is split across three desktop apps with no shared state. Products are searched twice, SDS files are split manually, and notes written in one tool are invisible in the next.
Design need · One web platform with automated handoffs and built-in document handling.
There is no global search and no filtering, so users scroll thousands of records. Duplicate suppliers and inconsistent naming make even a successful search untrustworthy.
Design need · Global search, smart filters and duplicate governance.
Freezes during upload and matching, 30–60 second waits with no progress feedback, unreliable OCR. Users compensate with multiple windows, manual refreshes and Teams messages.
Design need · Async architecture, visible system status and autosave.
Reviewing customer requests in iChemistry showed unclear product names, unrecognised existing products and missing supplier or language data, the direct source of downstream duplicates.
Design need · Validation and smart suggestions at the customer request form.
| Heuristic | Observed issue | Severity | Design response |
|---|---|---|---|
| Visibility of system status | No loading feedback; users open extra windows to check whether anything is happening. | Critical | Progress states, live status chips, explicit success/failure messages. |
| Error prevention | Nothing validates language or supplier before a wrong SDS is saved. | Major | Auto-detected document language and supplier validation before save. |
| Consistency & standards | Each tool has its own layout and vocabulary for the same object. | Major | One design system and one shared product/supplier terminology. |
| Flexibility & efficiency | Four separate emails sent to the same supplier; data re-entered per tool. | Medium | Batch email with templates, task grouping, inline editing. |
| Help & documentation | No onboarding or in-system guidance; new staff learn workarounds from colleagues. | Medium | Contextual help, empty-state guidance and documented flows. |
Scored with the department manager and the delivery lead, so scope reflected both daily pain and what engineering could realistically build first.
Removes the single largest manual time cost in the Requester workflow.
Daily friction; users currently depend on other teams for small fixes.
Restores accountability and replaces email archaeology.
The root cause of every multi-window workaround.
Cleans the product base and prevents new duplicates.
Ends repeated one-by-one supplier contact.
Requester, Handler and Entry Tool merged into a single role-based workspace: a dashboard with workload and activity timeline, a product queue with inline editing and duplicate flags, a supplier panel with batch email and advanced SDS search, and an entry panel that validates and publishes to iChemistry in the same flow.
Halfway through, the evidence pointed upstream. Duplicates and language mismatches were created when customers submitted requests in iChemistry, where the form offered no suggestions and no validation. Moving error detection to the customer interface prevents work instead of speeding it up.
The final concept spans the whole request-to-publish lifecycle: Intake Queue, Handle Hub, Entry Assistant, Documents and Audit & History. AI parses documents and proposes matches, but every field carries a confidence level and a review state, the specialist stays the decision-maker.
AI output is labelled High or Review at field level. Anything below the threshold blocks publishing until a human confirms it, automation earns trust by being auditable, not by hiding.
Statuses propagate automatically between intake, handling and entry. The old habit of setting a product back to 'queued' by hand disappears with the tools that required it.
Product, supplier, document and request are named identically everywhere, which removed the translation layer users carried between three interfaces.
The audit trail records who did what and when, with notes attached to the product rather than to a chat thread.
Validation at the request form, AI assistance during handling, safe correction (inline edit, delete, merge) afterwards, in that order of priority.
Five moderated think-aloud sessions on the clickable prototype, with specialists from discovery. I measured unaided completion, hesitation and confidence, not raw speed.
| Task | In the legacy tools | Prototype result | What I learned |
|---|---|---|---|
| Find a specific product and open its history | Search in two tools, no history exists | 5/5 unaided | Global search plus the product timeline replaced what used to be an email search. |
| Import a customer Excel and review matches | Manual line-by-line matching | 4/5 unaided | One user missed the match-summary step; a result summary panel was added after upload. |
| Contact three suppliers about missing SDS | Three separate emails, no record | 5/5 unaided | Batch email with templates; participants asked for a visible 'last contacted' date, which I added. |
| Publish an AI-parsed sheet with a low-confidence field | No confidence signal at all | 3/5 unaided | Biggest failure. 'Review' chips read as decoration; I made low-confidence fields block publish and surface a review counter. |
| Explain who last changed a status and when | Not answerable in the system | 5/5 unaided | Audit trail answered it in one click, the change participants reacted to most strongly. |
Interactive prototypes were tested with the same specialists interviewed in discovery, focused on filters, AI flags and navigation between queue, hub and entry.
Feedback drove clearer confidence labelling, a visible match-result summary after Excel upload, and simplified navigation between the five lifecycle areas.
Success metrics agreed for launch: time per matched product, duplicate rate, share of requests auto-validated at intake, and number of supplier emails per product.
Three tools became one role-based workbench with automated handoffs, a product no longer changes system to change status.
Error prevention moved upstream: the request form validates supplier, market and language before work is created.
Every action writes to an audit trail, replacing Teams threads as the record of what happened.
Figures are targets agreed with the department during prioritisation; the platform is in prototype validation.
The most valuable decision was leaving the brief. I was asked to redesign internal tools; the research showed a share of the workload was manufactured upstream by a customer form nobody had questioned.
Working with the data engineer early kept the concept buildable, traceability and matching were designed against the real entity model, not around it.
If I ran it again I would instrument the legacy tools first. The qualitative evidence was overwhelming, but baseline timings would have made the business case immediate.
I'm available for product design roles and consulting projects.
contact@lidatohidi.com
