AI or Die: Why Leadership, Not Technology, Decides Who Survives

By Anas Semesmieh · October 23, 2026

In August 2026 a ticket landed on my team. Known AI bots were being blocked by our firewall, and the pages on one of our websites were barely being indexed. We had been hardening the edge against bad bots. That is good, diligent security work. It was also shutting the door on the crawlers that decide whether an AI assistant can find, read and recommend our stock.

That ticket is the cleanest way I know to explain the title of a talk I gave at the AAICT AI Workshop on 23 October 2026. "AI or Die" is a headline. The truth is narrower and harder: you do not die from not using AI. You die from being invisible to it, and from being out-executed by whoever isn't.

1. The front door has moved

Search used to end with a click on your site. Increasingly it ends inside an answer. The figures below are from the sources I used in the talk.

Being cited is the new ranking. If your content cannot be reached by the machines that answer the question, your customers' questions get answered without you.

2. Your customers already use AI. Your industry doesn't.

In Australia, 17.4 million people use AI and 5.2 million use it daily (Telsyte, Jun 2026). Only 12% of Australian businesses used it in 2024 to 25 (ABS). The industries we serve are lower still: transport and warehousing 1%, agriculture 3%, construction 6%, wholesale 14%, mining 18%.

Inside organisations the gap is governance. Only 30% of Australian organisations have a generative AI policy, and only 24% have given staff any AI training, against 39% globally. Meanwhile 48% of Australian employees use AI against company policy (KPMG and the University of Melbourne). Banning it does not stop it. It only stops you seeing it.

3. Three years, not three months

The easy mistake is to read a good AI result as a recent event. Ours stands on groundwork that started in 2023, and it came in stages:

• 2023: ask how to measure, not whether to buy. The first conversations with the vendor were about how we would know a coding assistant was working. A cross-team AI community formed around that question.
• 2024: make experimenting public. A company hackathon let teams show what generative AI could do on real systems, including a prototype that one team described as "the best junior dev ever".
• 2025: from a pilot to a way of working. A small group of power users moved to the enterprise tier. Command-line agents and specification-driven development followed.
• 2026: make it somebody's job, then open it up. Dedicated AI enablement roles were appointed, and a program put agent-building tools in the hands of people outside engineering.
• Late 2026: agents leave the demo stage. AI-assisted code review, model access for every team, and the first business agents in production.

4. Three receipts

Receipt one: spec-driven delivery, measured as an experiment

We ran spec-driven development on a real product, deliberately, as a measured experiment and not a vibe. The specification became the artefact the agent worked from. The product shipped materially faster than the traditional approach would have allowed, and ahead of schedule.

I try to give the honest version of the research too. A 2023 controlled experiment found developers completed a task 55.8% faster with an AI pair programmer. A later study of 4,867 developers at Microsoft, Accenture and a Fortune 100 company found 26% more completed tasks, with the biggest gains for less experienced developers. And METR found experienced open-source developers were 19% slower with early-2025 tools. Their 2026 follow-up struggled to run, because developers no longer wanted to work without AI.

Receipt two: stop making it an engineering thing

We opened an AI program to staff from across the business, chosen for proven adoption and not for job title, and trained them on low-code agent tooling. Each team took a real business problem. Mine built an assistant that produced a customer brief with a suggested next step, replacing hours of manual preparation. We were one of the winning teams. None of the people on it were engineers, and none were on my platform team.

The people closest to the problem built the thing that solved it. They did not need us. They needed access, a safe platform, and permission.

Receipt three: agents stopped being demos

The first business agents went into production on our AI platform. They shipped with a golden test set and monitoring, like a business product and not a prototype. That is the difference between an agent you can keep and a demo you quietly turn off. More are on the way in other corners of the business, including HR and legal work.

5. What AI-aware leadership actually bought us

Microsoft's Work Trend Index 2026 found that 67% of AI's impact is explained by organisational factors and only 32% by individual ones. Google's DORA research says the same thing in one line: AI amplifies what is already there. Be specific about what that meant for us:

1. They made it somebody's job. Actual roles, with accountability. Not a working group, not a side-of-desk passion project.
2. They funded a real platform. Azure AI Foundry with model guardrails, private networking, managed identity, an AI gateway and a dedicated compute pool, all defined as code.
3. They gave access to everyone. Copilot Studio and AI Foundry went to every team, which is why non-engineers could build working agents within weeks.
4. They made experimenting public. Hackathons, demo days and winners. Leaders went first, in public, where it could have been embarrassing.
5. They gave a budget with a path, not a wall. A per-user AI budget with a documented exception process.
6. They were willing to pause. Copilot Studio was paused for a week to tighten data access before staff started building agents. Slowing down on purpose is a leadership act.

6. If you take five things

1. Make AI enablement somebody's actual job. It is the highest-leverage item here, and it is a leadership decision, not a technical one.
2. Give everyone access, tied to a real problem. A few non-engineers with a real problem beat a hundred seats with none.
3. Check whether AI can find you, tonight. Look at your bot rules. We were blocking the crawlers that decide whether an assistant recommends us. You probably are too.
4. Build the compounding layer, not just the tooling. Shared memory, shared skills, shared scripts. Individual productivity adds up. Organisational memory multiplies. I wrote up one version of that in Never Start From Zero.
5. Measure outcomes, not usage. "Eliminate 100 hours of work, waiting or effort every month" is a target. "Seats activated" is a vanity metric.

The honest part

We were lucky to have leadership that understood this early, and we made sure we used it. I cannot take credit for the conditions. I can take credit for what we built inside them, and that is the deal I would offer any engineering leader. If your leadership has already made these calls, you have no excuse: go build the compounding layer. If they haven't, that is the thing to go and fix. Not your tooling. Not your model choice.

"AI won't replace humans, but humans with AI will replace humans without AI." Karim Lakhani, Harvard Business School.

The headline is "AI or Die". The truth is narrower: enable, or be out-executed by whoever did.

Sources

SparkToro zero-click study (Jun 2026); Pew Research Center (Jul 2025); McKinsey, New front door to the internet (Oct 2025) and The State of AI 2026 (Aug 2026); Adobe AI Traffic Trends Report (Aug 2026); Similarweb AI referral traffic (Sep 2026); Seer Interactive AIO click-through update (Apr 2026); Cox Automotive AI in Auto Retail Tracker (Aug 2026); Telsyte Australian AI Study 2026; Roy Morgan AI tools usage (Jun 2026); ABS Characteristics of Australian Business 2024 to 25; KPMG and University of Melbourne, Trust in AI 2025; Microsoft Work Trend Index 2026; Google Cloud DORA 2025; Peng et al. (2023); Cui et al. (2025); METR (2025, with a 2026 study-design update); Karim Lakhani, Harvard Business School. Internal Pickles figures come from our own delivery and platform records.