Escaping AI Slop: How Atlassian Gives AI Teammates Taste, Knowledge, & Workflows, w- Sherif Mansour

Mansour explains how Atlassian has evolved from serving primarily technical teams to having the majority of its 3.5 million AI users in non-technical departments like marketing, HR, and finance. He introduces the concept of AI as a 'virtual teammate' integrated into collaborative workflows rather than just personal productivity tools.

• •Majority of Atlassian's users are now non-technical knowledge workers, not software developers
• •3.5 million users actively using Atlassian's AI capabilities (Rovo)
• •Focus on team-level AI integration rather than individual productivity
• •AI teammates help speed up collaboration across project management and service delivery workflows

Mansour introduces his three-ingredient framework for combating generic AI outputs: Taste (team character and voice), Knowledge (organizational context and documentation), and Workflow (deployment in business processes). He emphasizes that 'open by default' knowledge bases provide significant advantages for AI effectiveness in enterprise contexts.

• •AI slop defined as technically correct but creatively lazy output that lacks differentiation
• •Taste: Apply team's unique voice, tone, and creative preferences to avoid generic results
• •Knowledge: Connect relevant organizational documentation and context (with proper permissions)
• •Workflow: Deploy agents in specific business processes with clear instructions and evaluation criteria
• •Atlassian's 'open by default' philosophy (20+ years) creates better AI outcomes than permission-locked systems
• •Personal memory features automatically capture user preferences over time (e.g., writing style with emojis)

Mansour explains Atlassian's 'Teamwork Graph' - a knowledge structure that maps relationships between users, teams, goals, and work artifacts across multiple systems. This approach handles queries like 'what did my team work on last week' that require traversing structured relationships rather than semantic search.

• •RAG fails for complex enterprise queries requiring traversal of team structures and work relationships
• •Teamwork Graph maps entities: users, teams, goals, Jira tickets, Confluence pages, GitHub PRs, Figma designs, etc.
• •Graph populated both organically (users linking content) and inorganically (system integrations)
• •Hybrid approach: Use RAG for document retrieval ('annual leave process'), use graph traversal for relationship queries ('team status')
• •Collaboration signals (likes, comments, views) provide additional weighting for relevance
• •Permission-aware at field level across all connected systems (Salesforce, SharePoint, GitHub, etc.)

Mansour discusses Atlassian's approach to managing dozens of AI features across their portfolio using an AI gateway that proxies multiple models. He argues that general-purpose models are commoditizing for most knowledge worker use cases, while emphasizing the importance of model stability for agent deployments.

• •70-80+ AI features across Atlassian's product portfolio, all managed through a central AI gateway
• •Teams get free rein in discovery, then optimize with ML team for production deployment
• •Simple tasks (summarization) offloaded to local open-source models; complex tasks use sophisticated cloud models
• •Customer demand for model selection has declined significantly - most care about results, not specific models
• •General-purpose models commoditizing for average knowledge worker use cases (90% of problems)
• •Model stability critical for customers building agents with their own evals and testing frameworks
• •Mix of Claude, GPT, Mistral, and other models used across different use cases

Mansour provides practical guidance on when deterministic code is better than AI, using customer examples of over-applying LLMs to simple string comparison tasks. He emphasizes that customers need to learn when traditional automation is more efficient than AI agents.

• •Many customers initially over-use AI for tasks better solved with deterministic functions
• •Example: Using AI to check if ticket contains specific words vs. simple string comparison
• •Agent framework supports both no-code and code-based 'skills' for hybrid approaches
• •Customers learning to combine deterministic workflows with AI where appropriate
• •Cost and performance both benefit from using right tool for each job
• •Skills can be Java code, regex, or other traditional programming for predictable operations

Drawing parallels to MS-DOS and the command line, Mansour argues that while chat is the universal interface to AI (like terminal to OS), specialized UIs will emerge for specific tasks. He predicts a spike in chat usage followed by proliferation of verticalized AI-native interfaces.

• •Historical parallel: MS-DOS terminal was universal interface, but specialized apps (Word, Photoshop) proved superior for specific tasks
• •Chat is universal interface to LLMs, but will be worst interface for many specialized use cases
• •Any AI feature can be 'faked' with prompts, but specialized UIs provide better user experience
• •Example: Leonardo AI for game sprite generation - chat interface would be terrible for this workflow
• •Humans want predictability in interfaces, not constantly changing AI-generated UIs
• •Design becomes huge differentiator when everyone has access to same AI models
• •Expect proliferation of specialized AI-native interfaces built on conversational backends

Mansour explains the vision behind acquiring The Browser Company, drawing parallels to how business messaging (Slack/Teams) evolved differently from consumer messaging (AOL/ICQ). Resetting assumptions about how knowledge workers use browsers leads to fundamentally different product requirements.

• •Parallel to messaging evolution: consumer tools (AOL, ICQ) vs. business tools (Slack, Teams) serve different needs
• •Knowledge worker browser needs: channels for org structure, integrations with business systems, permissions/control
• •Resetting assumptions about browser usage in work context leads to different product
• •Not competing with consumer browsing (shopping, travel planning)
• •Focus on how people work with variety of SaaS tools throughout the day
• •Josh Miller and team brought strong vision for reimagining browser for work context

Mansour challenges the Silicon Valley notion of the 'one-person unicorn,' arguing that even highly AI-leveraged individuals still orchestrate agents in workflows, face bottlenecks requiring hiring, and must combat AI slop through taste and customization. He emphasizes the shift from doing work to architecting how work gets done.

• •Even 'one-person' startups orchestrate AI agents in automated workflows (e.g., Jira + WordPress content pipeline)
• •High-leverage individuals still become bottlenecks and need to hire as they grow
• •Regulatory and compliance requirements in many industries require human domain expertise
• •AI slop problem persists: typing same prompt in 7 tools yields 80% similar results
• •Differentiation requires making AI 'yours' through taste, voice, tone, and context
• •Shift from humans as doers to humans as architects of AI-driven processes
• •Junior staff can be excellent AI adopters due to behavioral flexibility and AI-native thinking

Mansour discusses how to develop process architecture skills, noting that AI itself is 'arguably the best teacher of all time.' He shares insights on hiring AI-native talent and the challenges of behavioral change, particularly for senior staff accustomed to old workflows.

• •AI is excellent teacher for learning process architecture and workflow design
• •Atlassian deliberately changed hiring profiles to include more 'AI-native' candidates
• •Example: 11-year-old son went from Scratch visual coding straight to AI-assisted development, skipping traditional syntax learning
• •Behavioral change harder for senior staff comfortable with existing workflows
• •Best time to instill new behaviors is when someone is new to a role
• •Junior staff reviewing AI outputs build domain knowledge to become architects over time
• •Growth path: reviewer → architect as people develop taste and understanding of what works

Mansour provides guidance on acquiring AI startups, emphasizing team quality and product vision over traditional tech moats. He notes that in the AI era, technology advantages are temporary, making team capability and product thinking more critical evaluation criteria.

• •Traditional M&A criteria (team, tech, traction) weighted differently in AI era
• •Technology moats are temporary - models and capabilities commoditizing rapidly
• •Team quality and product vision become more important than specific technical implementations
• •Browser Company acquisition driven by team's ability to reset assumptions and reimagine products
• •Traction can be misleading - fast growth may not indicate sustainable differentiation
• •Focus on teams that understand how to apply taste and build specialized experiences
• •Look for teams building for specific contexts (knowledge workers) rather than general solutions