Contextual AI Layer

Product, segment and users

• Strategy layer: Topic briefing → Story structure → alignment to strategic parameters and personas/target groups.
• Execution layer: Articles → Multi-post → channel-specific versions such as social posts, blog posts, newsletters and press releases.
• Planners and managers define direction, structure, timing and strategic consistency.
• Editors, creators, copywriters and content specialists write, rewrite, adapt, translate and prepare channel-ready content.

Team setup

Scope and timeline

Pain points

• Strategy alignment did not scale: teams manually reviewed multiple screens and depended heavily on individual expertise.
• Channel adaptation was repetitive work: editors transformed similar ideas into channel-specific formats, tones and constraints, increasing inconsistency.
• Generic chat in a B2B product felt disconnected from real work, reducing trust and control.

Root causes

• A complex Topic → Story → Article → Multi-post model distributed context across briefings, personas, content models, strategy parameters and custom fields.
• Dense interfaces risked cognitive overload when AI entry points and guidance were unclear.
• Users could not trust generic outputs when the AI context was invisible.

Success criteria

• Faster alignment checks and improvement actions at Story level, with fewer steps to create channel-ready variants.
• Frequent AI tool usage on key screens and sessions that result in Apply rather than read-only output.
• Improved discoverability, clearer command naming and reduced result overload through progressive disclosure.

Domain and workflow map

• Topic: briefing, strategic parameters and target groups/personas.
• Story: narrative, alignment to Topic and timing planning.
• Article and Multi-post: writing and channel adaptation.

Real-world constraints

• Legacy, data-dense UI; complex permissions and roles; multilingual quality expectations; MVP timelines.
• Apply actions must respect edit rights, while view-only users still need useful read-only suggestions.
• The solution had to handle incomplete metadata, varied content types, custom fields and different configuration setups.

Risk areas

• AI too hidden → low adoption.
• AI too prominent → an overwhelming interface.
• Generic output → reduced trust.
• No clear apply flow → productivity collapses into copy/paste.

Discovery

• Personas and JTBD mapping highlighted two major jobs: aligning content with strategy for planners, and creating or adapting content efficiently for editors.
• Customer feedback showed that adaptation and improvement guided by existing context were more valuable than blank-page generation.
• Competitive research showed that AI was becoming expected; differentiation depended on context and workflow integration.

Hypotheses

• The highest value comes from context-aware alignment checks and recommendations using structured strategy inputs already in the system.
• Adoption improves when AI is embedded where work happens—in objects, fields and forms—not isolated behind a separate chat.
• Trust improves when users can control context inputs and turn output into an action through Apply or Replace.

Iteration loops

• Too many options and unclear navigation → a clickable AI menu, fewer steps and clearer navigation.
• Unclear labels and commands → names based on user intent, with hints and examples.
• Overwhelming or generic-feeling suggestions → summary-first results with details on demand.
• AI disconnected from text fields → inline actions plus Apply or Replace.

Contextual tools instead of one global chat

Adapt, improve and check before generating from scratch

Explicit, controllable context

Apply as a first-class outcome

Progressive disclosure

A scalable interaction model

Where the assistant lives

• A side panel that expands and collapses between narrow and wide modes.
• Tools relevant to the current object: Topic, Story or Article.
• Inline tools inside text fields and forms for quick rewriting and translation.
• Recent chats and new chat for continuity when switching tasks.

Planners: strategy alignment on Story level

• Check Story alignment with the Topic briefing: select the inputs to validate, review a summary, drill into missing information and next steps, then apply fixes in place.
• Suggest good timing for a Story based on relevant real-world events, including regional context.
• Compare a Story with other stories in its Topic to find strategic drift, duplication or inconsistencies.

Editors: channel-specific content generation

• Choose Create post text, select sources such as text, attachments, personas, models, strategy parameters and custom fields, then add tone, goal and constraint instructions.
• Generate text, iterate with another option and apply it through Use this text or Replace current text, with confirmation.

Editors: inline text editing and multilingual work

• Selected text exposes inline rewrite, correct errors, simplify language and translate commands.
• Outputs support replace or add actions, confirmation, and multiple options for controlled iteration.

What worked

• Treating AI as a workflow layer rather than a standalone chat feature.
• Designing for trust and control through context selection, apply options, confirmation states and progressive disclosure.
• Using prototypes and usability testing to make the AI experience concrete and shippable.

Challenges

• Early prototypes risked too many options and unclear navigation; an AI interaction model must stay strict to avoid adding complexity.
• Generic output and multilingual quality expectations required stronger guardrails and clearer context.

What I would do differently

• Define the product-wide AI interaction model earlier: entry points, naming, history, apply and undo patterns.
• Instrument analytics before release—from opening a tool to running, applying and undoing—to capture adoption evidence and iterate post-launch.