Your AI Copilot Is a Margin Trap. Build for Replacement, Not Assistance.

TL;DR

• Bolting AI onto legacy platforms destroys unit economics: inference costs on top of existing infrastructure without rethinking the workflow is a margin trap
• The shift is from productivity (do work faster) to replacement (do the work for you), and most "AI-enhanced" products are optimising workflows that shouldn't exist
• If the AI is optional, it's a gimmick; if the product breaks without it, you're building something real

---

Established software companies are racing to slap "Generative AI" stickers on platforms architected a decade ago. Every product demo has a chatbot now. Every roadmap has an "AI Assistant" workstream. Every earnings call mentions generative AI at least fifteen times.

They call it modernisation. I call it a margin trap.

The new compute paradigm

AI is not just another feature you add to an existing product. It's a new compute paradigm. The difference matters.

Adding search to a product was a feature. Adding collaboration was a feature. Adding AI to a product built on pre-AI assumptions is like adding a jet engine to a bicycle. The bicycle wasn't designed for it. The frame can't handle the forces. And the fuel costs make the whole thing economically absurd.

When you bolt inference costs onto a platform that was never designed around them, you don't get a smarter product. You get the same product, slightly augmented, with dramatically worse unit economics. The AI economics framework in the handbook covers how to model inference costs, margin thresholds, and pricing before you commit. Every AI call costs money. Every token has a price. If the underlying workflow hasn't changed (if users are still navigating the same screens, filling out the same forms, following the same multi-step processes) then you've added cost without removing cost. That's a margin trap.

The companies that will win aren't the ones adding AI to existing workflows. They're the ones using AI to eliminate workflows entirely.

From productivity to replacement

The old SaaS deal was straightforward. Pay us money, and our tool will help you do your work faster. CRM helped you manage relationships faster. Project management tools helped you coordinate faster. Analytics platforms helped you understand data faster. The value proposition was productivity: same work, less time.

The promise of agentic AI is different in kind. It's not "do the work faster." It's "do the work for you."

That single shift breaks the legacy model in ways that most product leaders haven't fully internalised.

Consider a traditional expense management platform. The legacy model: user logs in, uploads receipts, categorises expenses, fills out a report, submits for approval, manager logs in, reviews, approves. The "AI-enhanced" version: same workflow, but a copilot auto-categorises some expenses and pre-fills some fields. Faster, but structurally identical.

The agentic model: an agent monitors your email and bank transactions, identifies business expenses, categorises them against company policy, assembles the report, routes it for approval, and handles the reimbursement. The user never logs in. The form never exists. The workflow never happens.

Same outcome. Entirely different product. Entirely different architecture. Entirely different cost structure.

The first version adds AI cost on top of the existing platform cost. The second version replaces the platform with an agent. One is a margin trap. The other is a new business.

The interface inversion

Most platforms are designed to keep users inside the app. Session duration, daily active users, feature engagement: these are the metrics that SaaS businesses optimise for. The entire UX is built to make the user stay longer, click more, and engage deeper.

An AI-native product succeeds when the user doesn't have to log in at all. The goal is the outcome, not the interface. The best expense report is the one that files itself. The best CRM update is the one that happens without a human touching a record. The best project status is the one that assembles itself from commit history and ticket movement.

This is an inversion of everything the SaaS industry has optimised for over the past fifteen years. And it's why incumbents struggle with the transition. Their entire business (pricing, metrics, UX, infrastructure) is built around the assumption that users are inside the product doing things. Remove that assumption and the business model needs to be rebuilt from the ground up.

The companies that can't make this transition will keep building copilots. A copilot helps you fill out a form faster. An agent executes the transaction via API so the form never exists. One is an incremental improvement. The other is a category reset.

The assistant roadmap problem

If your roadmap is full of "Assistants" and "Copilots," step back and ask a harder question: should this workflow exist at all?

An AI assistant that helps a user navigate a complex configuration screen is solving the wrong problem. The right question isn't "how do we make this screen easier?" It's "why does the user need to configure this manually in the first place?"

An AI copilot that helps a user write a better query is solving the wrong problem. Instead of asking "how do we help users write SQL?" ask why the user is writing SQL at all when an agent could go from business question to answer directly.

This is uncomfortable because it means cannibalising features you've spent years building. It means admitting that some of your most complex, most differentiated workflows are actually liabilities in an AI-native world. The complexity you're proud of, the thing that creates switching costs and moats, is exactly the thing an agent can route around.

The trap is adding intelligence to the existing complexity. The opportunity is using intelligence to eliminate the complexity.

The test

A simple litmus test will tell you whether your AI strategy is real or cosmetic.

Remove the AI from your product. Does the product still work? If yes, the AI is a feature. It's nice to have. It might even improve metrics. But it's not a strategy. It's a sticker.

Now imagine the inverse. Build the product where the AI is the product. Where removing the AI means the product doesn't function. Where the value proposition is impossible without inference, without agents, without generative capability.

That's the difference between modernisation and transformation. One adds a layer. The other rebuilds the foundation.

The winners in the next era of software won't be the ones with the best chatbots running on the latest models. They'll be the ones with the courage to burn down legacy workflows and rebuild from agentic first principles. And the first thing that needs rebuilding is the pricing model. To look at a decade of product investment and say: "That solved the old problem. The new problem is different."

If the AI is optional, it's a gimmick. If the product breaks without it, you're building the future.

---

Frequently Asked Questions

Isn't this too risky for established companies with paying customers?

The risk of doing nothing is higher. Established companies don't need to burn down their existing product overnight. But they should be building the agentic replacement in parallel, not just layering copilots onto the legacy platform. The companies that only invest in "AI-enhanced existing workflows" will be outflanked by startups that build AI-native from day one, unburdened by legacy architecture and legacy thinking.

How do you price a product where the user never logs in?

Outcome-based pricing. Instead of charging per seat or per month for platform access, you charge for the value delivered: expenses processed, reports generated, transactions completed. This aligns the business model with the AI-native value proposition: the customer pays for outcomes, not for time spent inside your product. It's a hard transition for companies built on seat-based pricing, but it's where the market is heading.

What's the difference between a copilot and an agent in practice?

A copilot augments a human doing a task. It suggests, auto-completes, and assists. The human is still in the loop for every step. An agent executes a task end-to-end with human oversight at defined checkpoints. The distinction isn't about autonomy for its own sake. It's about whether the architecture assumes a human is present for every operation or whether it assumes the human defines the goal and reviews the outcome.