Junaid Haq: stop bolting AI onto broken operating models

Problems stay the same, mindsets have to change

Muzamil opens by naming the shift he keeps hearing about: the AI conversation is cooling from “it will take every job” toward something more grounded, and Noodle Seed’s own work is about connecting agents to business cases that survive the hype. He introduces Junaid, currently running business transformation inside Amazon’s global procurement and sustainability organization, previously close to a decade at AWS selling professional services into large European enterprises.

Junaid’s first move is to reset the frame. “Problems have almost or and always remained the same. It’s our approach and mindset uh on how we tackle them and solve them.” He traces a line from Lean Six Sigma and TPM through RPA to today’s AI wave, and lands on the diagnosis that runs through the rest of the conversation: companies are trying to solve new problems without changing the operating model underneath. By operating model he means the concrete stuff. Who does what. Governance layers. Tracking. How people are upskilled to work alongside agent systems.

He then references the recent MIT finding that 95% of AI use cases are not yielding P&L-level benefits. His read is that the number is probably right, and the reason is structural. “Unless you redesign your organization and your capabilities uh to work alongside these agent and AI systems the benefits will not be uh trackable, sustainable and implementable across the organization.”

Zero-based design, not RPA in a new coat

Muzamil pushes on whether redesign is even practical, citing companies like Meta and Salesforce that pushed hard on AI-first restructuring and are now rolling parts of it back. His worry is real: how do you move fast enough not to fall behind, without destroying working systems out of FOMO.

Junaid contrasts the old automation mindset with what he sees working now. In the RPA era, you would take a fifteen-step process and automate step five and step eight. Linear. Additive. The new question is different. Given that AI exists today, how would you redesign this process. He names the technique: zero-based design thinking. Start from the outcome the business is trying to drive. From that outcome, determine the decision rights. Then work out the minimum number of steps and where an AI-native workflow belongs.

The key word is “native.” Junaid distinguishes between digitizing an ecosystem, which is what the last twenty years were about, and building an AI-native ecosystem, which is what the next phase requires. The first treats AI as a tool inside a legacy process. The second treats the process itself as something to be rewritten.

He is also honest about the mess. Teams are in an experimentation, “storming, norming” phase. New tools land before people have finished figuring out the previous ones. Data is a persistent problem that cannot be fully solved before other work begins. His answer is agility and cross-functional teams, not a masterplan.

Compliance at Amazon scale is where the theory gets tested

Muzamil asks Junaid to clarify his current remit, since his title has moved from AWS to the business side of Amazon. Junaid explains he sits in the global procurement and sustainability organization, running a program that ensures compliance across ESG regulations, the EU corporate social responsibility due diligence act, and human rights acts, across a supplier base of more than 100,000 vendors providing indirect procurement services to Amazon’s global retail operations.

This matters because it grounds everything he says next. He is not theorizing from a keynote stage. He is running compliance and traceability at a scale where the operating-model question is not abstract.

A minimum viable operating model when the tools change every quarter

Muzamil asks about strategic thesis. Where does leadership think the world is going in two to five years, and how do you plan when tooling changes every quarter. Junaid splits the horizons. On the business side, the P&L still needs a two-to-three-year view for stability, resourcing, and strategic pillars. On the tech side, planning that far ahead is not credible.

The bridge is what he calls a “minimum viable operating model,” anchored on capability and people. Common themes are surfacing that let organizations plan without pretending to know which specific tool will win: how to design AI-native workflows, how to move from a process-oriented organization to a product-led one, how to build a generalist mindset in business end users that carries context into the workflow.

Two capabilities keep recurring. Data literacy, because “unless he understands that whatever tool comes then it will you know not give any fruition or any benefit to the end user.” And design thinking, working backwards from the outcome, which used to sit with product managers and now has to sit with business end users too.

Experience is the missing ingredient in most failed deployments

Muzamil offers his own read from a technical background. Since Claude’s Opus models became mainstream, the value of raw technical skill has collapsed because 85 to 90% of what a developer used to do is handled by the model. What remains valuable is the ability to break a problem into its data-flow nodes, understand what each step does, and stitch a system together. That is a domain problem, not a coding problem. He extends it into a hiring worry: if juniors cannot get in and domain expertise is what actually unlocks AI, where do the next expert operators come from.

Junaid agrees that experience is the missing ingredient. “Without experience being brought to these AI developments the outputs are not very optimal. Uh in fact they’re you know below suboptimal and that’s where the tools are failing miserably.” He gives a concrete example: business requirement documents used to take three to four months. With prototyping and AI, that cycle now compresses to days or weeks, but only if someone brings the right background context and captures learnings across the team. Without that, the output is worthless.

He is candid about the labor implications. The senior layer will likely grow. Middle-management orchestration burden will shrink as multi-agent architectures absorb the admin load. But if you do not hire enough juniors, the pipeline breaks. He does not pretend to have the answer. He says it will take years to mature, and it requires new questions: how do you feed strategy to an AI, how do you keep it responsible, how do you validate output, how do you monitor bias.

From task executors to orchestrators and contextual guardians

Muzamil asks what new roles will emerge. Junaid answers from his own function. Today his teams do vendor segmentation, audit coordination and scheduling, SLA tracking, chasing suppliers for evidence, remediating findings, and reporting. Most of the time is manual follow-up.

When that manual load is released, the same people become orchestrators. They have time to fine-tune models, run A/B tests, interview suppliers, feed insights back, and continuously tweak processes every two months instead of writing an SOP and revisiting it a year later. Task executors become trusted advisers. Risk orchestrators. Audit orchestrators.

On the data side he sees another shift. Today’s data owners spend most of their time building pipelines and BI reports. In the future, he says, you get “data owners becoming contextual business guardians.” Their job becomes controlling data quality, contextual efficacy, and consolidating disparate datasets into a single pane of glass aligned to a modern data strategy.

Democratization, the flywheel, and outcome-based contracts

Muzamil then asks Junaid to put his old AWS hat on. His observation: the top 0.001% of businesses used to run on a completely different level from everyone else. AI is democratizing that stack. A fifty-person company can now do what a five-hundred-thousand-person company did five years ago. He asks how enterprise-focused hyperscalers respond, and whether the real user-experience gap is what smaller, plug-and-play services are exploiting.

Junaid acknowledges the pressure directly. Per-unit economics have come down every year at AWS. Amazon’s flywheel, more variety, better pricing, more consumption, applies to cloud services as it does to retail. He points to Amazon Quick Suite as an example of packaging capability through Prime subscription so end users get access without heavy technical lift.

He also confirms the internal conversation has changed. In his sales days the recurring client question was how to make it easier. Professional services teams are pivoting to compress migrations from months to weeks to days, and to bring agents into their own delivery, scoping, deployment, security testing, and dev-ops.

The commercial model is changing with it. Contracts are moving from time-and-material and fixed-scope toward outcome-based structures with skin in the game. “You either do contingency, you put skin in the game and you say, ‘Okay, I’m going to save you $100 million or $50 million and you invest $10 million over the next 3 years.’” That structure, he argues, works for both SMBs that need fast value realization and enterprises that need durable transformation.

Trust as the new line of business

Muzamil closes by asking whether AI transformation will drag on the way digital transformation did, or whether this is a different beast. Junaid’s answer is net-positive but not utopian. Tactical repeatable work will be disrupted. New roles will emerge. Business models will be rewritten. The benefits only compound at organizational level if people are brought along. Otherwise the value stays trapped in a handful of pilots or individuals.

His picture of the winner is specific. “We are probably moving to a world of generalist where solving problems is not difficult. Having common sense and common knowledge and that experience uh is more important.” Curiosity, adaptability, and the ability to ask the right questions matter more than deep specialization in a single tool.

Before wrapping, Junaid offers one concrete example of a business that does not yet exist but probably should. In a conversation in California the previous month, he encountered the idea of LLM decision traceability as a service. If a C-suite or finance function is making decisions with AI in the loop, someone needs to maintain the log of those decisions, allow them to be revisited, and evaluate the branching consequences. He frames it as a new consulting or advisory line of business built around one thing: maintaining trust in the AI ecosystem.

That framing captures the argument Junaid has been making the entire episode. Trust, workflows, roles, and operating model are the actual substrate of AI value. The tools are downstream of all of them.