Chat Is the Universal Interface—and Often the Worst One

Atlassian is a $40 billion company, sitting among the world's top hundred technology firms by market cap, ahead of household names like eBay, Reddit, SoFi, and Expedia. Most people still picture Jira and software teams. Sherif Mansour, head of AI, offers a different census: the vast majority of users now come from non-technical departments—HR, finance, marketing, legal—people who still run projects and still provide a service to the rest of the business, they just do not write code. He has been at the company for 16 years. Before joining he was a customer, watching the tools change how a large Australian telecom worked. Recording this episode, he and host Nathan Labenz were 16 time zones apart. Another number he carries: close to three and a half million people already using Atlassian's AI capabilities.

Mansour's working picture of AI at work is not a personal chatbot. It is a virtual teammate dropped into a team's context, helping move a job along collaboratively. That raises an awkward design problem Atlassian has a value for—*build with heart and balance*. Nobody wants a robotic agent, and nobody wants one that tries too hard to be human and feels fake. In a business setting the path to authenticity, he argues, runs through trust, and trust runs through transparency: citations (Rovo has a team dedicated to them), visible instructions, a named creator, and a log of the last ten things the agent did. He is not arguing that agents should feel like human colleagues. He wants them to feel like collaborators who can take on a team's character—the in-jokes, the voice, the way a bank's customer-service squad shows up versus a youth-marketing campaign.

That character is also how he would fight slop. Plenty of people use the word for wrong answers and hallucinations. His looser definition is different: the output is technically correct, but the creativity is lazy, and everyone is getting similar results. In a world where the same frontier models are available to everyone, the differentiator is applying a team's soul—what it stands for, how it talks, what it likes and does not. He boils the practical recipe down to three ingredients. Taste is the opinion you inject, by prompting or by context, so you do not end up with the same website, the same product, the same service as everyone else. Knowledge is what the agent should access or be trained on—and those are not the same thing. Workflow is where you deploy it: a rigid Jira ticket, a Confluence page, some other automation, it almost does not matter as long as it lives in the work.

A procurement team he had just spoken with had all three. They kept Confluence pages on what to watch for in vendor contracts, connected SharePoint for extra material, and wired a Jira flow so that when a new item arrived the agent read the attached contract, drafted a first response, and handed the work to a human to revise. When customers hunt for a killer use case, he tells them to stop and narrate what they already do; the three ingredients usually fall out. He is Egyptian, so he thinks in pyramids: layer one is ask-a-question, get-an-answer; layer two is generate an artifact—a proposal, show notes, an intro; layer three is packaging knowledge and instruction into a repeatable workflow that scales quality or efficiency. Most people stall on the first step and conclude that AI is "just okay."

Labenz's own taste pipeline is a PDF of past intro essays plus the current transcript and a short prompt. The examples do the work; Claude had always been best at capturing his voice, until Gemini 3 produced one of the strongest drafts he had ever gotten. Mansour's observation is that most customers type a prompt, skip the prior art, and report that the result was fine, nothing special. The line he hears constantly is that the first 80% was okay and the last 20% took another five or six hours. Measuring taste is hard. Atlassian ships thumbs-up signals, reuse of the agent's output, and rephrase rates, plus LLM-as-judge evals on the 20 or 30 agents it ships out of the box. Customers still have to define success for the teammates they create, the way you would brief a new hire on a good answer versus a bad one.

Two Atlassian accidents of history now look like AI features. Rovo Chat builds personal memory: Mansour likes emoji before headings, and after enough writing assistance the model just does it. The phrase he uses is *user proximity*—how close someone already is to the AI while they work. Coders populate memory in the IDE without thinking; Confluence writers do it by saying make this shorter, you are too formal. The second accident is open by default. For more than 20 years, a new Confluence space or Jira project has been public unless someone restricts it. Join a company tomorrow, ask AI a question, and in a closed-by-default estate a new hire sees almost nothing. In an open knowledge base they get answers they would otherwise have had to message 50 people to find, or at least a pointer to the right person. AI still cannot invent taste nobody wrote down. Effort in, outcome out.

Under the hood, enterprise memory is not one trick. RAG gets brutally hard once permissions go to the field: the same GitHub issue or Salesforce record can show Mansour fields that Labenz cannot see, and the semantic index has to stay permission-aware and fast. The other half is the Teamwork Graph—users, teams, goals, and the work that attaches to them, from Jira items and pull requests to Confluence pages, Trello cards, Figma files, Google Docs, and Amplitude dashboards. Customers populate it inorganically by connecting GitHub or Figma, and organically by pasting links as they work. "Give me a status update on everything my team did last week" is a loaded sentence: what counts as a status update, who is the team, which week. RAG would be a terrible solution, he says—top five documents, then a summary. Graph traversal across the squad's tickets, pages, and designs is the right one. "How do I take annual leave?" is the reverse: RAG, not a graph walk. Serious enterprise AI needs both. Labenz heard HippoRAG in the description: entity linking, then expand one to three hops and search inside that neighborhood. Mansour agrees that multi-hop traversal is the point, and Atlassian layers a collaboration-signal graph on top—likes, comments, shares, views—as extra weights on people-and-work questions.

Remembering is only half the problem. Open-by-default also means 20 years of pages whose relevance nobody can swear to. Early in Rovo, Mansour asked the system to rewrite a team charter from current goals; it surfaced a goal from four or five years earlier that was still marked active because he had never archived it. They now decay content with time, interaction, and collaboration activity, but customers still have to think before they connect a source. One wanted to pipe all of SharePoint into Rovo. How big? About a terabyte. Did they need it? They did not know. Garbage in, garbage out has not been repealed. Systems should fade old objects—and users should archive the ones that are still labeled living.

Labenz used to tell people to always use the best model and optimize cost later. That advice aged; GPT-5 Pro is overpowered for many jobs, if only on latency. Internally, Atlassian still gives teams free rein in discovery—find product-market fit, test the riskiest assumption first, do not gold-plate a feature that is not useful. Across the main apps he has lost count, but he puts the portfolio at well over 70 or 80 AI features, written as if AI will eventually be always on. After discovery, teams hit an ML group and an AI gateway: a proxy over locally hosted open-source models and cloud models, so they can swap, compare, and shunt simple summarization onto local boxes. Costs fall, and the savings get passed to customers, without slowing teams down while they are still learning.

Customers have the opposite problem. The fastest-growing agent use case is people dropping agents into the millions of Jira workflows they already have—triage this ticket, draft that reply. Some of them overuse the model. A ticket arrives; you only need to know whether certain words appear in it. The automation framework already has deterministic string nodes that inspect content and branch. Switching that step to an LLM can get a similar outcome at a ridiculous token bill. One customer put it in a phrase Mansour has kept: she had a rocket launcher, and she was swatting mosquitoes with it. Labenz's own mantra rhymes with hers: anything that can be done with traditional code should be, faster, cheaper, more reliable. Have the model write the regular expression; do not have it re-run the match every time.

The agent framework lets people attach no-code or code skills, so an agent can call a Java function that checks a bill or does the math instead of improvising. Customers usually have to discover that the hard way: amazement, then a case where they needed two exact steps and the model wandered. Asked for a token breakdown by model, Mansour pulled up a dashboard and shrugged—ridiculous millions, probably billions, a number that has become pointless to stare at. What has changed is customer attention. Two years ago every buyer wanted to pick the model. Now the care factor has collapsed; it feels like insisting the company use Postgres instead of MySQL. The stack is a mix of GPT, Claude, Mistral, and whatever else fits the job, with docs on which major features use which model.

His advice to software teams is blunt: it pays to build a model gateway. If you believe that layer is being commoditized, you need a way to learn, switch, and cost-optimize without rewriting the product. He says it has paid off by orders of magnitude. He is on the commoditization side of the debate for general-purpose models and general-purpose desk work—finance, HR, marketing, customer success—where those models already solve 80 or 90% of the problem. Vertical models, a DNA-sequencing specialist, are a different story. The important asterisk is for teams who build their own agents and evals: API stability matters. Swap the model overnight and tomorrow's results are a different product. Models commoditize. Control does not.

Labenz's theory is that we build UIs because software needs a human's judgment, and that the best interface might be no interface: abstract the screen into a task for an agent. Mansour reaches for tech history. MS-DOS and the terminal were the universal interface to the operating system. You could do math, type a document, even draw ASCII art. People learned quickly that the terminal was a terrible interface for many of those jobs, and they spent decades building vertical apps on top of it—word processors, spreadsheets, image tools, the dedicated recorder the two of them were using. Translate that to AI. The universal interface to a large language model is conversation, because transformers predict the next token and chat is the shape that exploits that. Universal is not the same as good. Chat usage will spike. Then specialist experiences will grow on top of it, the way GUIs grew on top of the shell. He still opens a terminal now and then. He does not live there.

The test he puts to internal teams is rude and useful: for any AI feature, nine times out of ten you could ship the poor man's version as a prompt with no UX. That is the worst experience, because you cannot expect people to type those prompts, or to stuff structured data into a blank box. The fashionable rebuttal is that AI will dynamically generate the right interface every time. He is unconvinced. Filling a form in chat is a disaster—validation, conditional fields, the whole mess. Models can generate a new layout on each turn; humans do not want software that will not sit still. His son was building tower-defense sprites in Leonardo AI, with predictable characters, hand positions, weapons, and size variants. Could you do that in ChatGPT? Totally. Is it the worst interface for dedicated, repeatable game art? Also yes. Design becomes the differentiator in a world where everyone has the same models.

That argument folds into how AI changes SaaS. Jira's value was never a to-do column you could vibe-code in a weekend. It was that Coca-Cola, or whoever, made the tool theirs by modeling how they build products, handle incidents, and onboard employees. Agents do not retire that problem. If anything, employers want more control over how virtual teammates behave, or they slide back into slop—every company running the same process. "Everyone goes from doing the thing to architecting the thing," he says. Specialist legal AI is the existence proof: his brother, an M&A lawyer, works in tools that cite prior cases and legislation by country. A prompt can gesture at the same job. It cannot run it at scale. Dynamically generated UI, pushed far enough, just becomes vertical SaaS with extra steps.

Labenz presses the one-person unicorn: if models keep getting better, maybe a kid with Gemini 4 and light guidance outruns an enterprise that boxed its agents into prescribed workflows. Mansour is skeptical on four counts. He had just talked to a five-person sales-AI startup; their leverage came from agents in a Jira workflow publishing thought leadership to WordPress on a schedule, with humans reviewing. Even the supposedly tiny teams are orchestrating, not sitting in a prompt box. Second, those founders are the bottleneck and they keep hiring—more ideas than people and agents to build them. Third, whole markets cannot move without a domain expert's signature. Fourth is slop again: the first person to ship a unicorn-shaped product with generic model output will be copied, 80% alike, the moment someone else types the same prompt into seven tabs. Making it yours—voice, taste, context—does not go away. He does not see another solve.

If the scarce skill is process architecture, Labenz wants to know how you teach it, and whether that means you stop hiring juniors. Mansour says people forget that AI is arguably the best teacher available. Atlassian has already changed product-management hiring toward people who are AI-native—the joke is they want more of the kids who were "cheating" in college. His 11-year-old went from Scratch, a visual language, straight to vibe coding, skipping the middle years of syntax. When Roblox shipped an AI in the IDE, the boy ran into the room and started instructing it; when it missed, he learned to ask the model to teach him how to specify what he wanted. The wrong conclusion is hire only seniors. Behavior change is hard, and many seniors still do the old thing because it is fastest. New people in a role are the easiest place to install a new habit. The procurement reviewers who are not yet architecting the workflow are still building taste about what good looks like, and they become architects over time.

Atlassian bought The Browser Company to build a browser for knowledge workers, not to compete with shopping and trip-planning. Mansour's analogy is consumer chat: AOL Messenger and ICQ worked okay at the office and never really took off there. Reset the assumptions for work and you get channels that mirror org structure, integrations, permissions—Slack and Teams, not a clone of ICQ. Reset the browser the same way: a surface that already has per-user permissions, the teamwork graph, and a team plus its virtual teammates moving across many SaaS tools. On buying AI startups more generally, he tells teams to define long-term as 12 months or they drown in analysis. Ask about *workflow proximity* on both the user side and the buyer side: who already lives in that job every day, and where does the startup sit on that job's Maslow pyramid? The closer to the base of the stack, the easier it is to eat upward. A calendar-AI from a nobody is two hops from winning; sitting at the fifth step of a four-step job is a dangerous place unless the niche is truly one the platform vendors will ignore.

Will cheaper software mean fewer developers? Labenz offers a spectrum: free dental work would not make him buy more of it; free massages might make him consume a hundred times more. Mansour wants the question sliced. Consumer software—entertainment, leisure, home renovation, cooking—he cannot see shrinking, even as fabricated content blurs. In business, AI looks better the more tools it can call, and humans still have to build or review those tools. Granularity explodes: a calendaring app is one tool to a person and something like 50 to an agent (find a slot across time zones, block busy time, pull public holidays). Designing software for AI is more complicated, not less. Vendors will shift from UIs toward tool interfaces—Apple is already pushing app intents—but he does not see the industry contracting. The useful debate, he says, is by domain: automotive software may not move like compliance-document software.

On adoption, leading by example only works if the stories are better than "I summarized a doc and drafted an email." Tech waves that hit personal life first hit work harder: e-commerce, then mobile. He landscaped a backyard with AI, built a math-homework agent that knows the curriculum and will not give the answer (Fortnite jokes included), watches his daughter make AI music, and after an MCL injury added a daily rehab assistant that also comments on what he ate. In interviews his favorite question is how the candidate uses AI personally. He does not believe in mandates; most of the workforce hears a threat. The tactic that worked at Atlassian was an AI builders week: a thousand product managers, designers, engineering leaders, and researchers blocking time together so nobody felt guilty for tinkering, celebrating learning rather than a shippable outcome. When customers hunt for a greenfield killer use case, he tells them to throw it out. Pick a team, write the steps of a task they already do, and next to each step note the knowledge, instruction, and action a human uses. The agent-shaped holes appear by themselves. His closing encouragement, repeated in different words all hour: ignore the buzzwords and tinker on a tap washer, a homework problem, a sore knee. Behavior at work follows what you actually use when nobody is watching.