Escaping AI Slop: How Atlassian Gives AI Teammates Taste, Knowledge, & Workflows, w- Sherif Mansour
Summary
Mansour’s enterprise-AI framing emphasizes that permissioned workflow context around a “virtual teammate” matters more than any single model. Atlassian was framed as a $40 billion, top-100 technology business whose users now skew toward nontechnical teams, with roughly 3.5 million people using its AI capabilities. The strategic shift is from personal chat productivity to agents participating in shared projects and service workflows.
The operating recipe for escaping generic AI output is “taste, knowledge, and workflows.” Taste gives the agent a team’s voice and judgment; knowledge grounds it in organizational material; workflow determines where it acts and where humans review. Mansour’s procurement example joins Confluence and SharePoint guidance to a Jira process in which an agent reviews an incoming contract, drafts a response, and hands it to a human.
Permission-aware RAG alone cannot answer many of the enterprise questions that matter. A query such as “What did my team work on last week?” requires Atlassian’s teamwork graph to traverse people, teams, Jira items, Confluence pages, Figma designs, GitHub pull requests, goals, and collaboration signals; returning the top five semantically similar documents would be inadequate. Conversely, “What is the annual-leave process?” remains a natural RAG problem.
Mansour expects general-purpose models to commoditize across “80–90%” of ordinary knowledge-work use cases, making routing infrastructure and model stability more valuable than brand selection. Atlassian uses a gateway spanning hosted and local models, preserving expensive capability for difficult work while offloading simpler tasks such as summarization. Customers also need to recognize when AI is “a rocket launcher” being used to swat mosquitoes and substitute deterministic rules or coded skills.
Chat may be AI’s universal interface, but Mansour argues it will often be “the worst interface in the long term.” His MS-DOS analogy is that command lines remained universal while specialized applications won individual jobs; similarly, predictable vertical interfaces should sit above conversational models. That implies design and domain-specific UX remain defensible even as virtually any feature can first be prototyped as a prompt.
AI should increase individual leverage without automatically producing the much-discussed one-person unicorn. Even a five-person startup Mansour observed used orchestrated agents, recurring workflows, and human review; its people remained bottlenecks with more ideas than capacity, while regulation and domain expertise still demanded humans. His deeper objection is differentiation: without sustained human taste, multiple AI-native companies converge on “80% the same” output.
Atlassian’s acquisition of The Browser Company reflects a wager that work browsing will diverge from consumer browsing just as business messaging developed differently from consumer messaging. A browser aware of user permissions, organizational knowledge, the teamwork graph, SaaS tools, and virtual teammates can be rebuilt around knowledge work. For AI M&A generally, Mansour favors vertical products that own the valuable last 20% and possess “workflow proximity” to users or buyers.
Enterprise adoption depends less on mandates than on synchronized, safe experimentation. Atlassian blocked time for an AI builder week involving about 1,000 product managers, designers, engineering leaders, and researchers, defining “the outcome” as learning rather than a shipped product. Mansour’s practical prescription is to map an existing task into its knowledge, instructions, and actions—not invent a speculative “killer use case”—while using AI personally enough that new behavior becomes natural.
Deep dive
1. Atlassian sees AI moving from personal assistant to team participant
Mansour describes Atlassian as a collaboration company rather than merely a software-development vendor. Its historical base was software and IT teams, but most users now come from HR, finance, marketing, legal, and other nontechnical departments that either run projects or provide internal services.
The AI user base matters to the thesis: roughly 3.5 million users were already using Atlassian’s AI capabilities. Mansour distinguishes personal chat productivity from putting “a team agent in their context” so it can help multiple people move shared work through a collaborative process.
Labenz stresses why a drop-in knowledge worker is attractive: software-like agents can operate 24/7, scale to several parallel instances, switch off when unnecessary, and be copied. His challenge is that their rough edges still prevent organizations from treating them exactly like human teammates.
Mansour’s answer is deliberately balanced. Nobody wants something “incredibly robotic,” but something trying too hard to be human also feels inauthentic; Atlassian therefore speaks of virtual teammates and collaborators without pretending they are people.
2. Trust and team character are the antidotes to AI slop
In a business setting, authenticity resolves into trust, and trust depends heavily on transparency. Mansour wants users to see an agent’s instructions, creator, and recent actions; Atlassian even has a team devoted to improving Rovo’s citations so answers reliably expose their underlying sources.
A team agent should also reflect the group it joins. Every team has “character”: its in-jokes, vocabulary, tone, creative instincts, and norms. A bank team serving customers should not sound like a digital agency marketing to young people, even if both access the same underlying models.
Mansour defines AI slop narrowly as output that is technically correct but whose “creativity is lazy”—different users receive essentially the same result. His defense is to add “your team’s character, your team’s soul,” preserving differentiation without making the agent falsely imitate a human identity.
Labenz’s introduction demonstrates the point. For the last two years, going back to Claude 2, he had supplied Claude with prior essays and the current episode’s transcript, then extensively edited the draft; Gemini 3 produced one of his best-ever first drafts and seemed to “get” him from the writing samples. The model mattered, but curated examples embodied the taste.
3. Taste, knowledge and workflow turn a model into a business system
Mansour reduces useful enterprise deployment to three ingredients: “taste, knowledge, and workflows.” Taste is the organization’s opinion about how something should look, sound, and work; without it, companies risk producing the same websites, products, and services from the same broadly available models.
Knowledge means deciding what the agent can access or be trained on—two related but distinct choices. The grounding material may span Confluence, SharePoint, GitHub, Salesforce, and other systems, subject to each user’s permissions rather than a single organization-wide pool.
His best specimen is procurement. A team documents its contract-review criteria in Confluence and keeps supplementary material in SharePoint; when a Jira work item arrives with a vendor contract, an agent reviews it, drafts a response, and passes the ticket to a human for critique and iteration.
The practical discovery method begins with current work, not an imagined killer application. Mansour asks a team to explain what it already does, then identifies where its taste enters, what knowledge it consults, and where an agent can be inserted into an existing workflow.
4. Adoption climbs a pyramid, but context and evaluation remain local
Mansour pictures AI maturity as an Egyptian-inspired pyramid. At the base, users ask questions; next they generate artifacts such as proposals or show notes; at the third layer, they combine team knowledge and instructions into a repeatable workflow that improves quality, effectiveness, efficiency, or idea generation.
Personal memory helps users climb without constructing every context package manually. As people use AI-assisted writing in Confluence, corrections such as “make it shorter” or “you’re too formal” arise near the work itself; Rovo applied, for example, that Mansour likes headings prefixed with appropriate emoji.
Atlassian’s “open by default” heritage can bootstrap organizational context. A new employee in a closed document environment may receive almost no useful answers until access is granted; in an open knowledge base, the same person can retrieve institutional information that otherwise might require messaging 50 colleagues.
Yet Mansour rejects magical personalization: “You have to put in effort to get outcome.” The familiar last 20% can consume another five or six hours. Product signals include whether the user used the output, thumbs ratings, and repeated rephrasing, while creators of specialized agents need to define their own success criteria and can add examples of good and bad responses.
5. Enterprise retrieval needs permission-aware RAG and a teamwork graph
Labenz frames the technical problem as the space between context-window working memory and knowledge baked into model weights. Mansour splits that space between personal memory, permission-aware enterprise retrieval, and structured organizational relationships rather than betting that an ever-larger context window will solve everything.
Enterprise RAG is unusually difficult because visibility can vary down to individual fields. One user may see parts of a Salesforce record or GitHub issue that another cannot, so a semantic index spanning Salesforce, SharePoint, GitHub, and other systems must enforce permissions while remaining fast.
Atlassian’s teamwork graph maps users, teams, goals, work items, documents, designs, pull requests, and their relationships. Jira often acts as a system of record that tracks where work stands and links to where the work itself occurred; ordinary use also populates the graph when people paste related links.
“Give me a status update about everything my team did last week” exposes RAG’s limit. The system must identify “my team,” traverse its members and connected Jira items, Confluence pages, and Figma designs, then apply the organization’s status-update format; summarizing five blue-link results would be inadequate.
6. Graph traversal and semantic retrieval solve different questions
Mansour does not claim the graph replaces RAG. A request for the company’s annual-leave process is likely best answered by retrieving the relevant policy, whereas a team-status question needs broad traversal through explicitly modeled people, work, and time relationships.
Labenz compares the architecture to HippoRAG: recognize and reconcile entities, expand outward through one or more hops, then search within the resulting universe. Mansour agrees multihop traversal matters but adds that Atlassian can overlay a separate collaboration graph beyond document relationships alone.
Likes, comments, shares, views, and repeated collaboration can weight relevant nodes. If Labenz frequently comments on Mansour’s pages, that relationship may guide a people-and-work query; the system can also distinguish material already viewed from unseen content when the user’s request makes that distinction useful.
7. Useful organizational memory must learn to forget
Open-by-default knowledge creates a countervailing problem: some customers possess 20 years of accessible content, and an old page may no longer describe reality. Enterprise memory therefore needs recency and activity signals as well as permission checks and semantic similarity.
Mansour encountered the failure personally when Rovo rewrote a team charter using an objective from four or five years earlier. The goal remained marked active because he had never archived it, so the system reasonably treated stale source data as current organizational truth.
The system tries to apply forms of decay based on time, activity, collaboration, and object state, but Mansour preserves the customer’s responsibility. When someone proposed connecting an entire roughly 1 TB SharePoint instance, his response was effectively: can you, yes; do you need it, unclear.
The governing principle remains “garbage in, garbage out.” Buyers evaluating an enterprise-AI system should ask whether connected sources express lifecycle state, whether old information can lose gravity, and whether the system can still retrieve historical material when the question genuinely requires it.
8. Model routing and deterministic code beat indiscriminate inference
Internally, Atlassian first attacks the “riskiest assumptions”: establish that an AI feature is valuable before obsessing over inference cost. Controlled rollouts can manage costs during discovery; only after fit becomes plausible do teams optimize latency, model choice, and expense.
With “well over 70 or 80” AI capabilities across major applications, recurring patterns now justify infrastructure. An AI gateway lets teams swap and test many cloud and locally hosted models, sending simple workloads such as summarization to cheaper options while reserving sophisticated models for harder tasks.
Customer workflows create a different optimization problem. Existing Jira processes contain millions of workflows and potential agent insertion points, but a deterministic string check may classify a ticket faster, more cheaply, and more reliably than an LLM. One customer realized they had “a rocket launcher” and were “swatting mosquitoes with it.”
Atlassian’s agent framework therefore supports no-code and coded skills: the model can call a precise function for math, bill checking, or another business rule. Labenz agrees that anything traditional code can perform should generally remain traditional code, even if AI writes the regular expression or function.
9. General models commoditize, but behavioral stability still matters
Customers once insisted, “I want to pick the model”; Mansour now sees that concern fading toward the importance level of PostgreSQL versus MySQL. Atlassian uses a broad mix—including GPT, Claude, and Mistral variants—and publishes model information for major features rather than treating one provider as universal.
His recommendation to software builders is structural: create a gateway because models, costs, and comparative strengths keep changing. The ability to proxy, switch, test, and reroute workloads has paid off “orders of magnitude” for Atlassian’s development speed and operating economics.
Mansour places an explicit asterisk on commoditization. General-purpose models can “easily” solve roughly 80–90% of ordinary desk-based knowledge-work problems, but vertical models—for example, one specialized in DNA sequencing or healthcare—may retain differentiated value.
Model interchangeability also stops at the evaluation boundary. An agent creator may tune instructions and tests against one behavior profile; swapping the underlying model overnight can materially change outputs. Stability therefore resembles API stability even when model brands cease to be an end-user selling point.
10. Chat is AI’s universal interface—and often its worst one
Labenz proposes that software displays a UI only when it needs human intelligence, taste, or judgment; if AI can supply those inputs, the ideal product may become software users barely touch. He asks whether today’s SaaS interfaces should be progressively abstracted into agent tasks.
Mansour answers with MS-DOS. The terminal was the operating system’s universal interface and could support writing, math, or ASCII art, yet dedicated word-processing, image-generation, spreadsheet, audio, and podcast-recording tools emerged because the universal interface was often poorly suited to the specific job.
The same pattern should recur: “Chat is the universal interface, but it’s the worst interface in the long term.” Mansour asks product teams whether a feature could first be faked as a “poor man’s version” in a prompt; if so, the product’s work is packaging data, controls, and interaction into a better experience.
He remains skeptical that endlessly generated interfaces eliminate design. Dynamic forms may work where inputs are predictable, but users still need validation, conditional fields, and stable learned behavior. Labenz’s son’s use of Leonardo AI to build game sprites illustrates why a dedicated experience for predictable sprite variants would beat repeatedly describing every variation in a generic chat box.
11. AI does not erase SaaS; it raises the value of orchestration
Labenz presses the stronger counterargument: once AI can answer tickets, perhaps ticket triage disappears—the Elon Musk-style principle that “the best step is no step.” Agents might ultimately organize work differently from humans, making workflows encoded for human limitations an incumbency trap.
Mansour concedes that agents will not necessarily work as people do, but their actions still arrive as tools or skills called within some sequence. What survives is the need to decide which actions agents take, which humans retain, and how people and agents divide work and handle handoffs.
Jira’s value, in this account, is not a generic “to do, in progress, done” board. Customers make it their own by encoding how they build products, handle incidents, serve customers, onboard employees, and respond to change—the organization’s taste expressed as workflow.
His conclusion is categorical: “agent orchestration with human workflows” becomes pivotal, and “everyone goes from doing the thing to architecting the thing.” Even dynamically generated specialist software must eventually become predictable and scalable; sophisticated legal interfaces for citations, jurisdictions, and precedent illustrate why vertical tooling persists.
12. AI leverage changes team shape, not the need for people or judgment
Against the one-person-unicorn thesis, Mansour cites a five-person sales-AI startup whose agents publish recurring thought-leadership content to WordPress—or whatever its blog is—from a Jira workflow. Humans still review and critique the drafts: high leverage came from orchestration, not from one founder casually entering a few prompts.
In the lean companies he observes, the people running them remain bottlenecks and keep hiring because they have more ideas than people and agents can execute. Regulation, compliance, and domain expertise further constrain sectors where work cannot proceed without qualified human intervention.
His deeper objection is creative convergence. Seven tools given the same prompt can yield results that are “80% the same”; the first company might monetize that slop first, but followers would resemble it. Building a lasting business still requires making AI “yours” through taste, voice, context, and deliberate process.
Nor does architecture imply hiring only senior staff. Mansour jokes that Atlassian wants more students who were “cheating” with AI because they arrive AI-native; behavioral change may be harder for experienced workers. Reviewers also learn domain judgment from assessing AI output and can become workflow architects over time.
13. A work-native browser and AI M&A both reward proximity
Atlassian’s acquisition of The Browser Company starts from resetting assumptions, not competing for ordinary consumer browsing. Mansour compares the opportunity with AOL Messenger and ICQ: business messaging eventually developed channels, integrations, permissions, and workflow constructs that made products such as Slack and Microsoft Teams fundamentally different from consumer messaging products.
A knowledge-work browser can assume AI is always available across SaaS tools while respecting per-user permissions and drawing on organizational knowledge and the teamwork graph. That creates a collaborative environment for people and virtual teammates rather than a browser optimized mainly for shopping, travel, or casual web use.
For acquisitions, Mansour warns of an “AI fog” around what can be reproduced overnight. His stabilizing rule is to define the long term as 12 months, judge the likely trajectory, and ask whether a target can move customers from the generic model’s 80% to a valuable last 20% they will pay to obtain.
His second filter is “workflow proximity,” split between users and buyers. A calendar platform begins near the base of scheduling’s hierarchy of needs and can climb upward; a startup delivering only the fifth step after four upstream systems is vulnerable unless its niche is large and specialized enough.
14. Cheaper creation should multiply software and atomize its tools
Labenz frames demand elasticity with a spectrum: he would consume no extra dental work if it became free, but perhaps 100 times more massages. The open industry question is whether dramatically cheaper software creation resembles a fixed necessity or a service with vast latent demand.
Mansour expects consumer software to keep expanding across entertainment, leisure, cooking, and home projects, while warning that increasingly fabricated content blurs reality. In business, he sees no contraction because AI becomes more capable as developers give it more tools.
A calendaring application may appear to a human as one product but decompose for an agent into perhaps 50 narrow tools: find mutual availability, handle time zones, block busy periods, retrieve public holidays, and more. Designing these reliable capabilities can be more granular and sophisticated than presenting one UI.
Vendors may shift value from interfaces toward callable tools—Mansour points to Apple encouraging developers to build App Intents—but software does not disappear. The precise outcome will vary by market, from automotive systems to compliance and document-signing workflows, rather than following one universal headcount curve.
15. Adoption accelerates when leaders make experimentation safe and concrete
Mansour agrees leaders should model AI use, but summarizing a document or drafting an email is insufficient inspiration. His richer personal examples include visualizing a backyard renovation, building a curriculum-aware math tutor with Fortnite jokes, helping his daughter make music, and managing daily MCL-recovery exercises.
He pushes back on mandates because many employees hear “use AI” as “do this or get out.” Atlassian instead ran an AI builder week for about 1,000 product managers, designers, engineering leaders, and researchers, blocking time synchronously so experimentation did not feel like neglecting teammates or normal duties.
The event celebrated learning rather than a shipped outcome: “the outcome is the learning.” That framing matters because adopting a new method can initially take longer and fail, while employees under ordinary delivery pressure naturally revert to the fastest familiar behavior.
Mansour’s implementation template is brutally concrete: choose a task already performed, list every step, then identify the knowledge, instructions, and actions used at each point. If nobody currently performs the proposed work, “throw it in the bin yesterday”; personal tinkering—from homework to identifying a tap washer—is what eventually changes workplace instinct.