Why Does Your AI Feel Soulless? The World's Top Tech Companies Are All Doing One Thing: Injecting Taste into AI | AI Practitioners

Contributor | Sherif Mansour Content Editor | pippobei Produced by | AI Nao

Intro

AI Nao is launching a new column: AI Practitioners

We focus on AI builders — people using new paradigms and fresh perspectives to solve real problems in the AI space. We believe that in a period of rapid technological change, experience matters just as much as opinion.

This is our second issue.

Over the past six months, AI founders we've spoken with at AI Nao have been bringing up one word with unusual frequency: taste. There's a growing consensus that taste determines whether an Agent product can be competitive.

But what exactly is AI taste? And how do you instill it?

For this issue, we've edited and excerpted an interview with Sherif Mansour, Atlassian's Head of AI, from the well-known video podcast Cognitive Revolution.

• Sherif Mansour

Cognitive Revolution is a program exploring how artificial intelligence will reshape society over the next decade. Its host, Nathan Labenz, was previously founder and CEO of Waymark, a company featured by OpenAI as a case study for "generative AI content creation platforms."

The interview subjects on Cognitive Revolution closely resemble those in our own AI Practitioners — hands-on practitioners focused on the how, not model specs or theoretical debates.

This issue's case study comes from Atlassian, a globally recognized software company whose flagship product, Jira, is an enterprise-grade task tracking and project management tool. Founded in 2022 and headquartered in Australia, the company serves 300,000 customers.

Atlassian's Head of AI, Sherif Mansour (hereafter "Sherif"), is not your typical AI engineer. He's a product person who's spent 15 years at Atlassian and took charge of the company's AI strategic transformation in 2024. The AI products under his leadership are already used by millions.

Atlassian's products have two distinctive characteristics:

First, a large share of non-technical users — HR, marketing, operations, and other teams.

Second, a substantial technical user base — software engineers and the like.

This gives Sherif an unusually broad perspective: how to deeply integrate AI with the core workflows of an organization? At the same time, he faces one of the hardest AI scenarios in the world: multi-team, multi-role, multi-person collaboration; unstructured knowledge; and constantly shifting workflows.

To make AI effective in such a complex environment, Sherif believes the fundamental problem right now isn't that models "make mistakes" or "hallucinate frequently" — it's that AI output is broadly mediocre.

He describes the result as "AI slop" — generic, cheap, soulless, student-homework-grade waste material.

For example, when users ask AI to write content, it's always "disturbingly normal." Ask it to write a PRD, and it can't capture the style your team actually wants. The core reason:

Large models generate the average, but what humans call style is fundamentally preference.

From this, he proposes a key metric for Agent products today: Taste. "Make it truly a colleague who shares your taste, not just a tool," Sherif says. And instilling taste in AI "absolutely does not come from data — it comes from the user."

This feature is adapted from How Atlassian Gives AI Teammates Taste, Knowledge, & Workflows. The original video runs roughly 100 minutes; we've selected the sections where Sherif discusses "taste" and reorganized them into a more readable, broadly applicable format.

Method 1: Learn from User Choices

Most AI products ask users to write examples, provide sample texts, or upload documents to teach the model a writing style.

But Sherif discovered: what users write isn't necessarily what they want — but what users choose is always real.

What matters is: which paragraphs users keep, which they delete, which get edited the most, where they add a data point, which sections they treat as unimportant — these reveal users' truest preferences.

Sherif offers a simple but powerful example:

Why do PMs always delete AI-written background sections?

When users write PRDs in Confluence, Atlassian's core product:

AI-generated background introductions in the first paragraph get deleted by users 80% of the time. But metric breakdowns in those same documents tend to be preserved. This means users don't need AI to write background for them — they need it to help PMs break down problems clearly.

This is an extraordinarily granular yet valuable insight.

Sherif says that from this point on, they designed a "retention rate" metric and used it to train their models.

This is the first layer of where AI taste comes from.

Method 2: Learn from Team Consensus

The second layer of AI taste comes from team habits. Every team has its own customs, but they're rarely documented.

For instance:

• The legal team starts every piece of content with "This does not constitute legal advice";
• The marketing team likes to add CTA summaries in copy;
• The engineering team always includes an architecture diagram in technical RFCs;
• The product team lists "user problems" before "solutions";

These are things models can't guess.

Sherif recommends making your team's entire collaboration process the AI's default style.

For example, when you ask AI to write brand copy, the traditional approach is to have AI study a pile of what you consider qualified copy — but the final output ends up commercial-sounding, like soulless copywriting.

Sherif's suggestion: users should feed AI the full record of how copy gets revised again and again — which words get cut, meeting notes about the copy, why the brand lead ultimately insisted on replacing these words and keeping those ones... inject all of it into the AI.

In other words, let AI see not a static result, but the team's working path.

It will ultimately learn a direction, not just fixed patterns.

Sherif's original words in the interview: "Taste is not a workflow. Taste is how your team keeps editing the workflow."

This is like an intern joining a team — have them read documents all day and they won't align with the team's style. But have them sit in on a few meetings, watch colleagues argue, discuss, reach consensus, and they'll quickly understand: this is what the team's taste is.

Method 3: Bring AI in Early

Sherif makes a bold point in the video: as model technology improves, AI will inevitably shift people from executors to process architects.

In other words, the "delivering results" that the industry keeps talking about right now isn't where AI's value lies — its greatest value should be driving decisions.

Based on serving nearly 300,000 customers across 200 countries, Sherif believes that having AI write a Jira ticket for you or generate a piece of content doesn't actually save people time. Because what hinders enterprise efficiency isn't one person struggling with documents or a specific task. It's:

The friction that arises when multiple departments and roles try to coordinate and push processes forward.

That's the real time sink.

He believes the greatest value of AI going forward will be making decisions.

The ideal state: an Agent should automatically identify who a task should be assigned to (based on history), explain which dependent task it traces back to, auto-generate deadlines (based on project rhythm), auto-link relevant knowledge documents, then send Slack notifications to the right people — flagging who needs to review, who needs to approve.

Achieving all of this absolutely requires an Agent with taste, one that fully understands team preferences, how the organization operates, its rules and boundaries, then makes professional judgments and pushes processes forward.

Based on this, Sherif's advice is: users should start using AI from the very source. Break tasks down for AI and tackle them bit by bit right now. For example, to write a PRD: have AI organize user interviews first, then synthesize user needs, then list edge cases, then write solutions, and finally fill in structure. Operations, design, brand — every department's work should follow this approach.

Ultimately, AI "taste" cultivation begins the moment a user initiates a task and continues throughout: what does this team value, how do they frame questions?

"AI shouldn't reverse-engineer an organization's workflow by looking at results. It should be directly injected with operational source data: knowledge graphs, event streams, conversation records," Sherif says. "The more granular the task breakdown, the more texture the output will have."

---

Sherif's Outlook

As models get cheaper and knowledge graphs and workflows become replicable, AI taste will become an irreplaceable moat. Whether it truly understands users' ways of working — even their ways of living — depends entirely on how you use it.

Image credits | Unsplash, Nano Banana

Share with Us

If you're a builder figuring out AI products on the front lines, we'd love to hear about your practice.

Contact info in the card below: "WeChat"