Three of the strongest sessions at ** SaaStr AI 2026** came from three very different seats. Sharif Mansour runs AI and the product management craft across Atlassian’s 20+ apps and 450 product managers. Eleanor Dorfman leads the commercial and industries sales team at Anthropic. Rory O’Driscoll has been a software investor at Scale Venture Partners for 30+ years.
Different companies. Different jobs. Different stages. Almost the exact same lesson.
Every popular claim about AI right now comes packaged as a binary. Chat is the universal interface, so kill your UI. No, build dedicated experiences and skip chat. Reimagine your product from scratch. No, bolt AI onto what you have. Hire 10x AI builders. No, you can’t afford juniors anymore. Software is dead. No, software is fine.
The binary is the trap. The teams winning right now are running both sides of every one of these at once. And underneath the both/and framing, the same concrete primitive surfaced in talk after talk. Skills. The harness. The thin layer of software that turns a raw model into something a business can depend on.
Here is what each session said.
Session 1: Sharif Mansour, Head of AI and Product Management Craft at Atlassian, on the Three Contradictions of Shipping AI Into Real Products #
Sharif Mansour opened with a line that set up the entire conference: for every claim he has read about how to implement AI, there is an equal and opposite claim that is also true. So instead of picking sides, Atlassian went looking for where the answer sits in between. The context matters. This is not a startup with one app and a clean slate. This is 20+ apps, most of them years old, six of them AI-native, with more than 5 million users on the AI features alone.
Contradiction one: chat as the universal interface vs. dedicated UI.
Two years ago, adding chat to the products was controversial inside Atlassian. One camp said chat is a terrible experience for most tasks, so just build the features people want and skip it. The other camp said nobody knows what people will do with these models, so ship chat and watch. Atlassian had to build a chat backend to power AI across the portfolio anyway, so they shipped it (it is called Rover) and learned from it.
The analogy Sharif used: the command line never died. DOS was the universal interface to the operating system, and over the years specific use cases got pulled out into dedicated apps for spreadsheets, documents, and games. The terminal is still here. Both layers coexist. Chat is that universal interface for AI, the infinite use case, and you pull specific workflows out of it into dedicated UI.
In Confluence Whiteboards, they watched users type things chat could not yet do. “Group all my sticky notes into common themes.” So they built it as a feature: select the cards, group them. Then they watched users try to push those notes into a Jira backlog through elaborate prompts that failed, and they built that as a workflow. Three patterns came out of this:
- Automate the prompt. A repeated prompt becomes a button.
- Prompt to workflow. A prompt that crosses systems becomes a multi-step capability.
- Conversation to UI and back. Some flows start in chat (“pull all my Salesforce and Google Drive feedback onto a whiteboard”), move into a dedicated experience, then return to chat.
On chat usage, they assumed it would fade once people standardized on outside tools. It did not. Millions use Rover chat every day, even though those same users also run Gemini and Claude Code and everything else. The lesson Sharif drew is workflow proximity. People reach for the AI closest to where they already work.
Contradiction two: reimagine from scratch vs. bolt AI onto existing workflows.
“Bolt-on” already sounds like an insult. Nobody wants to admit they did it. Atlassian did it on purpose. Two and a half years ago they were one of the first B2B vendors to put agents into the platform, and they added an agent step into an existing Jira automation workflow. Arguably the best thing they did, because they had no idea where the market was going and bolting on let them learn fast.
What they learned: customers immediately went past single agents into branching, conditions, and multi-agent chains. A ticket fires, an agent picks it up, a marketing agent talks to Canva to generate assets, a social agent posts them. Customers were automating workflows they already had. That pushed Atlassian to a principle that now governs the whole portfolio: every problem you solve for humans, you solve for agents. Humans need tools, context, goals, accountability, and awareness of what teammates are doing. So do agents. The primitives are nearly identical. Anything you can do with a human, you should be able to do with an agent, and if a design breaks that rule, the team has to justify why.
For new products, reimagine. For existing products with users and workflows, bolt on and evolve. Both. Contradiction three: hire the 10x AI builder vs. build the 10x team.
Atlassian ran about 10 projects staffed with AI builders, people whose main job is shipping code with AI tooling, drawn from PM, design, and engineering. Early results felt incredible. Then after a few weeks they slowed down badly. Everyone was rowing and no one was steering. The PMs and designers organically drifted back to their old jobs, giving the team customer context, vision, and decisions. The takeaway is not that AI builders are fake. It is that “everyone becomes an AI builder” is the misconception. As engineers churn through far more work, the PM and design ratio that used to feel like 1:10 or 1:20 now feels like 1:30 or 1:40, which makes the people steering more important, not less. If your steerer is heads-down live coding, nobody is directing the ship.
The hiring twist worth stealing: Atlassian flipped its talent pyramid. They now hire more juniors and more seniors, fewer in the middle. Their internal research found juniors are 19 to 30% more likely to use AI and almost twice as likely to experiment with new tools, while seniors are far better at spotting slop and refining output because they are used to coaching juniors. So they run an AI Builders Week every couple of months, four synchronized days where juniors teach new tools and seniors teach quality control. The hardest thing for a 10, 20, or 30-year veteran is unlearning how they work. The juniors never learned the old way, so they treat vibe coding as simply how you code.
Running underneath all three contradictions: skills. Every feature Atlassian built for humans, they exposed as a tool for agents, on and off their platform, through MCP. “Group my sticky notes into themes” became a human feature and an agent tool. Build it once for people, expose it everywhere.
Top 3 learnings from Atlassian:
- Ship chat to find your roadmap. The universal interface tells you which workflows to pull out into dedicated UI, through three patterns: automate the prompt, prompt to workflow, and conversation-to-UI-and-back handoffs.
- Make “anything a human can do, an agent can do” a design principle. Every problem you solve for humans (tools, context, goals, accountability) you have to solve for agents, and the primitives are nearly identical.
- Build the 10x team, not the 10x builder. Flip the hiring pyramid toward more juniors and more seniors, and expose every human feature as an agent skill so you build it once.
Top 3 mistakes Atlassian made:
- Almost did not ship chat at all. Two years of internal debate over whether chat even belonged in the products nearly cost them the discovery engine that ended up driving their roadmap.
- Assumed chat usage would fade. They expected to learn from chat and then take it away. Instead millions use it daily, and treating it as temporary would have been the wrong bet.
- Staffed “everyone as an AI builder.” About 10 projects took the immediate speed gain and then stalled because everyone was rowing and no one was steering. They had to walk it back to a team model.
Session 2: Eleanor Dorfman, Head of Industries at Anthropic, on Building an AI-Native Sales Org From Scratch (Without a New Stack) #
Eleanor Dorfman’s session was the operator’s version of the same idea. The trigger was the Opus 4.6 launch in December 2025, which turned demand vertical overnight. Anthropic came back from winter break to a market they had not staffed or planned for, and even if they had wanted to 3x, 4x, or 5x the sales team, you cannot absorb that many people that fast without torching your bar and your customer experience. So the question in January was: how do you build an AI-native sales org from scratch?
The thesis is the most important part, and it is the opposite of what most people assume AI-native means. It was not a teardown. Anthropic had already spent three years building a strong stack: Clay for enrichment, LeanData for routing, Salesforce as system of record, Gong for call coaching, Ironclad for contracts, Slack for coordination, Intercom’s Finn for support. The move was not to replace any of it. It was to thread Claude through and around it so six tools became one coherent customer journey instead of six tools with AI bolted on. As Eleanor put it, it is not a new stack. It is being intentional about the existing one.
Four investments came out of four constraints they could not move:
Self-service enterprise. Eleanor had spent 15 years believing enterprise plans get gated by a human, because customers want to talk to an AE. They threw that orthodoxy out. Claude and Clay qualify and enrich every lead, then split into two funnels. The self-serve path runs through Finn, which guides the buyer to an enterprise plan with terms, invoicing, provisioning, and onboarding, fully self-served. The sales path routes qualified leads to BDRs and AEs. The result: 54% of new enterprise logos in 2026 came through the self-serve funnel. They do not treat self-service as a downgrade. They treat it as friction removed for buyers who want to move.
Thread Claude through the six-tool stack. Claude is the connective tissue that makes the tools talk. When a lead comes in, Claude and Clay do the account research, prioritization, and record updates, and pull history from Slack, Google Docs, and prior Gong calls so the AE walks in with full context. When an AE is ready to send a proposal, Claude generates it inside policy, aligned to the customer and the negotiation history, and uploads it into Ironclad. Forecasts are largely run by Claude and then inspected by managers, which turns forecast calls into discussion instead of gotcha sessions.
Make Slack the front door for every support function. The old reality was a sea of DMs, where AEs needed institutional knowledge or a desk near deal desk to get a quote approved, and East Coast and Europe reps stayed up late chasing approvals. The fix: reps stopped going to the systems and the systems came to them. Slack in, ticket out. Claude triages, resolves on precedent and policy where it can, and escalates with full context (customer contacts, history from email, Salesforce, and Gong) to a human when it cannot, then tells the AE so they can set expectations.
Encode the best reps as skills. A three-person go-to-market productivity team, plus reps sharing their own builds, turned the patterns of top reps into five skills every rep uses daily:
- Morning brief. Built on every connector they run on (Gmail, Gong, Slack, Google Docs, calendar, Salesforce, Intercom). One prompt, and it prioritizes the day: what is on the calendar, what emails and Slacks are unanswered, what initiatives are due.
- Call prep. A one-pager before every call: who is on it, what they care about, historical context, a good outcome, discovery questions, positioning, competitive landscape. Reps still have to read it.
- Customer follow-up. Extracts action items from email, Gong, Salesforce, and Slack, drafts responses into the email provider, and reminds you the next morning if you did not hit send. Humans stay in the loop.
- Competitive intel. Instead of a battle card product marketing refreshes quarterly, Claude generates a dynamic, interactive, customer-tailored battle card on demand.
- Create an asset. Any deal, any stakeholder, any stage: a custom one-pager, landing page, ROI calculator, or prototype, on brand, generated on the fly. Some reps drop discovery call transcripts into Claude Code to build prototypes mid-cycle.
The number that explains all of it: AEs had been spending 70% of their time on internal processes, not on customers, and the volume of support tickets from customers who literally could not buy was climbing fast. The point of threading Claude everywhere was to give that 70% back. The context-repetition problem, re-explaining who you are and how you write across every chat, gets solved with org-level skills that hold brand, tone, and context consistent across marketing, sales, and customer success. Raise the floor for everyone, remove the ceiling for the best.
Top 3 learnings from Anthropic:
- It is not a new stack. The fastest path to AI-native was threading Claude through three years of existing investment in Salesforce, Slack, Gong, Clay, and the rest, not ripping and replacing.
- Self-service is part of the enterprise journey, not a downgrade. 54% of new enterprise logos came through the self-serve funnel once they stopped gating every enterprise plan behind a human.
- Encode your best reps as skills. Five daily skills plus org-level skills raised the floor for the whole team while giving AEs back the 70% of time they were losing to internal work.
Top 3 mistakes Anthropic made:
- Did not plan for the demand. Extensive annual planning was wrong within weeks of the Opus 4.6 launch, and they came back from break with no staffing or processes ready for vertical demand.
- Held the human-gated orthodoxy too long. The belief that enterprise plans must be touched by a human persisted until the buying experience broke, with customers sending angry messages because they could not buy.
- Forecasting is still unreliable. Every forecast call still opens with a 10-minute debate about how to forecast, and accuracy is openly a work in progress, not a solved problem.
Session 3: Rory O’Driscoll, Partner at Scale Venture Partners, on Whether Software Is Actually Dead #
Rory O’Driscoll took the same instinct and pointed it at the question every software investor and founder is getting: is software dead? After 30 years making money in software, he noted, that one feels personal.
He started with what nobody disputes. AI eats the work, not just the software, which is a far bigger market, and it is moving into spaces like healthcare and legal that were terrible software markets before. He also pointed out how knowable the trajectory was: ImageNet in 2012, transformers in 2017, the scaling laws paper in 2020 that effectively said money in equals results out, ChatGPT in 2022, and the situational awareness paper in 2024 that all but named the trade. And still everyone struggled to act on it.
Then the math. Hyperscaler AI CapEx this year is roughly $688 billion. Revenue coming out the other side is roughly $110 billion, about $89 billion of it from the two leading foundation model companies. That is half a trillion dollars more going in than coming out. AI is in invest mode, and on his estimates the revenue from the two big model companies does not surpass cumulative CapEx until 2031 or 2032. Five or six more years of invest mode. His warning: when you are spending more than you take in, you can hit a hiccup at any point, even if the long-term story is great. To make these numbers work, foundation models have to capture an appreciable slug of total knowledge-worker wages, on the order of 15 to 17% overall and north of 25% for software developers. Possible, maybe likely, but not without bumps.
On where to invest, Rory borrowed Jensen Huang’s stack: energy, chips, infra, models, apps. The bottom three are about making AI, and roughly 80% of that spend is not venture-shaped (chips are the giant block, mostly Nvidia). The models are the apex predator that has captured the value and probably keeps doing so. The right side, apps, is everyone else: how software uses AI to deliver value to customers. The trillion-dollar question is how that revenue splits between the model companies and the software companies built on top.
His answer to “is software dead” is really three answers, because the question hides three different questions:
Are post-2022 AI companies just going to get rolled by the foundation models? Not all of them, if they have real defensibility. He walked through the categories that are clearly safe: software plus sensors (OpenAI is not shipping touch sensors), marketplaces with network effects (Scale invested in Paraform in recruiting), proprietary non-public data, and full-stack businesses that become the service rather than sell to it (Range doing wealth management with LLMs). The trickier middle is the data flywheel that builds from usage and the forward-deployed-engineer model. The frame he likes better than “harness”: this is the new LAMP stack. Just because every B2B app gets built on the same framework (access a model, reason, distill into something useful for a business user) does not mean they are the same app, any more than every LAMP-stack app was the same.Is pre-2022 software dead? About 10% of Scale’s older portfolio was dead on arrival when ChatGPT shipped, because they had solved an AI problem that suddenly cost a few dollars per million tokens. The remaining 90% splits into rough thirds: insulated (often using predictive AI, orthogonal to GenAI), additive (a company like Drone Deploy where AI on top increases the product’s value), and genuinely threatened plain-vanilla B2B that automates a workflow and now has to move or die.Do you need your own model? Mostly no, but not always. Voice, video, and image had model options that were not pure LLM plays. Eleven Labs built a real business on its own speech model. There is room for specialist models in coding, chemistry, and physical science.
On multi-model, Rory said yes, we live in a multi-model world, and the model companies want that least of all. Unlike cloud, where you got locked in, an API call lets you rotate models, which is exactly why every model company is racing to build harnesses that eventually lock you back in. The live example was Replit: Sonnet builds the app, and when it hits a hard problem it brings in a second model as the “architect” to check the work. One engine checking another, in real time.
The closing argument is the one founders should hold onto. The SaaS multiple collapse is real, and it happened because the SaaS growth rate collapsed, from an average around 30% to sub-10%. Wall Street is not punishing these companies for nothing. So you need category conviction. If your category blends into the undifferentiated low-growth mass, there is no point. But just as Microsoft owned the PC era while hundreds of client-server companies won, and AWS owned cloud compute while a generation of B2B companies won on top, Rory expects two or three foundation models to dominate with oligopoly profits and a whole field of software opportunities on top. Software is not dead. It got harder.
His heuristic on compute intensity: the standard deviation across companies has exploded. Model companies spend 70 to 80% on compute. Coding companies 50 to 60%. Most app companies allocate roughly 10% of cost of goods to LLMs and build their value with the other 90%. Traditional B2B adding AI features lands around 8 to 20% and needs to get toward 5 to 8%. The exception: if the AI is so good the product sells itself, your AI spend is effectively your sales and marketing, and you can run it much higher. What you cannot do is carry 700 people in sales and marketing and a 50% compute bill. You get one or the other.
Top 3 learnings from Rory:
- AI is in invest mode for five or six more years. $688 billion in CapEx against roughly $110 billion in revenue means a hiccup is possible at any point, even if the long-term story holds.
- Software is not dead, but defensibility no longer comes from the model. It comes from proprietary data, network effects, sensors, full-stack ownership, and category position.
- We live in a multi-model world, and the LAMP-stack-style layer on top of the model is where differentiation lives. Same framework, very different apps.
Top 3 mistakes Rory was candid about:
- About 10% of the pre-2022 portfolio was wiped out overnight by ChatGPT. They had backed AI problems that suddenly cost a few dollars per million tokens to solve.
- They may be grading the threatened third too kindly. Rory admitted they were probably squinting optimistically at how much plain-vanilla B2B survives.
- The trend was knowable for over a decade and they still struggled to act on it. Clear signals from 2012 to 2024 did not translate into decisive early conviction.
Stop Picking Sides, Start Building Skills #
Three sessions, three sets of conclusions that rhyme.
The binary is the trap. Chat or UI, bolt-on or rebuild, junior or senior, software dead or alive, single model or multi-model. In every case the team that is winning runs both sides at once and treats the “or” as a false choice.
The advantage is the stack you already have. Atlassian’s value is 20 apps with real workflows and 5 million users. Anthropic’s value was three years of investment in Salesforce, Slack, Gong, and the rest. Scale’s surviving portfolio wins on proprietary data, network effects, and category position. Nobody won by starting over. They won by threading AI through what already worked.
The primitive everyone landed on is the skill. Atlassian exposes every human feature as an agent tool. Anthropic encodes its best reps as five daily skills and governs the org with org-level skills. Rory describes the same thing one layer down as the harness, the new LAMP stack, the software around the model that turns a raw model into a dependable system. Whatever you call it, the work of the next few years is building that layer, because the raw model is increasingly a commodity and the skill layer is where the differentiation and the data live.
And the operating model is lean. Atlassian’s PM ratio is stretching to 1:30 and 1:40. Anthropic ran this with a three-person productivity team and gave AEs back 70% of their time. Fewer people steering more leverage, with AI doing the rowing.
The AI model is the commodity. The data model, the workflow, and the skill layer you build on top are the moat. That was the thesis going into SaaStr AI Annual 2026. A product leader, a sales leader, and an investor each spent an hour proving it.