Am I the Bottleneck for My Own AI Agents?

How would I build an AI-native company? I explain how I design work around AI agents, software and people: define the outcome, find the bottleneck and decide who owns each decision. I'm Linus Ingemarsson, CEO and co-founder of Alice Labs. I help companies build AI and agentic systems, and I'm working through these same questions inside our own company. Work with me and Alice Labs: https://alicelabs.ai/en I've started setting up work so agents bring me things to review, approve or decide. I ask when work should reach me and what I need to contribute. I can be in the loop for a good reason and still become the person everything is waiting for. WHAT DOES AI-NATIVE MEAN IN PRACTICE? I look at the tasks inside a role: finding information, interpreting it, making a decision and taking action. A meeting might gather updates or resolve disagreements. I ask which tasks produce something the customer needs, then define the outcome before choosing how to do the work. For onboarding, a completed checklist only matters if the customer is ready to use the service. FIND THE CONSTRAINT IN THE BUSINESS I trace the customer journey from discovery and purchase to delivery, retention and feedback. The best starting point depends on what limits the outcome: opportunities, qualification, capacity, retention or work that costs too much. In an illustrative example, agents can prepare 20 onboarding cases a week, but a required approval step can handle only five. Preparing more cases creates a larger queue. It may reduce preparation cost, but completed customer outcomes are a separate measure. I explain how I use the theory of constraints: find the constraint, improve how it's used and check where it moves. DESIGN HUMAN REVIEW AND AGENT AUTONOMY For each review, I ask what decision I'm making, what could go wrong without it, what information I need and whether someone else could own it. A useful handoff shows the relevant facts, uncertainty and the decision required. Sending a long output for someone to inspect can leave them doing the work again. I choose a mechanism for each step. Fixed rules and calculations can stay in ordinary software. A defined sequence can use a workflow. An AI agent is useful when the next action needs to adapt. I decide what requires approval, what can run within limits and when the system must stop and ask. Someone still owns the outcome. GIVE PEOPLE A ROLE WORTH DEVELOPING I want more time for customers, trust, disagreements, leadership and responsibility. I connect AI upskilling to the actual work: describing good outcomes, spotting missing context and explaining why a result is good or poor. I involve the people who do the work in redesigning it, and make ownership of results, exceptions and rules explicit. MAKE THE RIGHT INFORMATION AVAILABLE I look at the current agreement, recent conversations, delivery constraints and company rules behind a decision. That knowledge may be spread across email, CRM records, documents and people's memories. It needs an owner, a way to stay current and a trusted source when records disagree. More context is not automatically better; the question is whether the agent has the right information. Access to a contract also does not mean permission to change it. OWN YOUR EXAMPLES, FEEDBACK AND EVALS I explain why a company needs examples of good work and a clear standard for judging results. Feedback should lead to an actual improvement: better guidance, a corrected source or an example available on a later run. Saving feedback does not automatically retrain the underlying model. An eval is a test case with a way to judge the result. I turn important requirements and recurring failures into tests, keep separate cases to check whether improvements generalize, and verify real outcomes. An agent saying an action succeeded is not proof that it did. WHERE I WOULD START I trace real cases through one important flow, find waiting and rework, then redesign a small part with the people who know it. I define handoffs, permissions, success criteria and ownership, test supervised work where appropriate, and compare quality, waiting time and total cost with today's process—including review, corrections, integration and maintenance. CHAPTERS 00:00 More preparation, the same bottleneck 00:26 Designing an AI-native company 01:48 What work actually is 03:50 Find the constraint in your business 06:18 When work should reach a person 07:59 Software, workflows or AI agents? 08:50 Human roles, ownership and upskilling 11:03 Context, information and permissions 13:14 Good work, feedback and AI evaluations 15:59 Where to start: one flow, real results What work is waiting for your approval, and which decisions really need you? Let me know in the comments. Subscribe for more practical videos about AI agents and business design. Learn more: https://alicelabs.ai/en #AINative #AIAgents #AliceLabs