Fast Track AI

Decision: AI as amplifier of existing platform, not a parallel system that would fragment user workflows

Rationale: I advocated for integration over isolation from the earliest strategy sessions. Adding AI as a separate product would create two ways to do everything, neither complete, forcing users to context-switch constantly and never build confidence in either path. Instead, I designed AI to use existing modals and patterns wherever possible, so the new capability felt like a natural extension of what partners already knew.

Result: Partners adopted AI alongside existing workflows immediately. When AI creates activities, it populates the same Activity modal users already know, complete with segment targeting, market selection, and timing configuration. New entry point, familiar execution, no retraining required.

Decision: Single AI interface in the main navigation rather than contextual AI buttons scattered throughout the platform

Rationale: I evaluated three options during the strategy phase: a global command palette similar to Spotlight that could be invoked anywhere, contextual AI buttons on every relevant screen, or a dedicated AI space accessed from the main navigation. I chose dedicated space because enterprise users committing real budgets needed clear scope boundaries. Scattered entry points would create confusion about what AI could do in each context.

Result: Users understand AI as a distinct capability with clear boundaries. The welcome screen chips communicate exactly what is possible before users type anything, preventing the frustration that comes from discovering limitations after attempting something.

Decision: Support text, voice dictation, and file uploads as input methods from the start

Rationale: Different users have different preferences and contexts. Some want to type precise queries, others prefer dictating while reviewing other screens, and some need to upload files like sales reports for AI analysis. Restricting to text-only would have limited adoption among users who find typing cumbersome or who work with external data sources.

Result: Voice transcription works reliably for users who prefer dictation, and file uploads enable workflows like generating detailed reports from uploaded PDFs that would be impossible with text-only input.

Decision: Constrain scope aggressively and communicate limits clearly upfront rather than letting users discover them through failure

Rationale: Early prototypes let AI attempt any query, which seemed user-friendly but caused problems in testing. Users were confused and frustrated when AI confidently attempted questions outside its actual capabilities. I redesigned the welcome experience to show specific capabilities through labeled chips (Data Mining & Analysis, Reporting & Visualisation) rather than open-ended "Ask anything" framing.

Result: Usage errors dropped significantly after the redesign. Partners understood what AI could and could not do before investing time, so their attempts were more likely to succeed and their overall impression more positive.

Decision: Errors as dialogue opportunities with clear recovery paths, not dead ends that require starting over

Rationale: Enterprise users committing real budgets need to trust the system even when it fails, because failures are inevitable with AI. I designed comprehensive error states where AI takes responsibility using first-person language ("I will simplify and retry", "This turned into more than I could handle") and always provides a specific recovery action. The tone is collaborative rather than blaming.

Result: Partners reported feeling in control even when things went wrong. Trust was maintained through transparency about what happened and clear guidance on what to do next.