The hardest call a CPO makes during the AI inflection is not whether to ship an agent product. It is whether to keep selling the SaaS product that pays the agent product's payroll. This is the decision framework I use: four diagnostic questions that route to three operating modes.
The four diagnostic questions
Ask them in order. Each answer routes the call.
First, can the legacy architecture support the successor's quality bar? Picture the agent-native product you would ship in twelve months if your current architecture were the only constraint. Can the legacy data model, latency profile, and security posture carry it? Do not accept wishful thinking from engineers here. Have a senior engineer write a one-page sketch of how the successor ships on the legacy stack. If the sketch needs three foundational rewrites, the architecture cannot support it, and it is forcing a sunset regardless of the customer math.
Second, is the legacy customer base the right ICP for the successor? If yes, sunset is operationally clean, because you just migrate them. If the base is the wrong ICP, sunset means losing those customers, and the real question becomes whether you can replace them faster than you lose them. Most companies discover the buyer for the agent product is a different person than the buyer for the SaaS product. That routes toward split.
Third, can the company afford the gross margin trough? Margin compresses from 78-82% for mature SaaS to 58-65% for early Service-as-Software during the ramp, bottoms around month 12, and recovers toward 70-75% by month 24 to 30. Work out how much cash the company burns at the bottom on its current trajectory. If you do not have the runway, the call cannot be sunset. It is split, where legacy revenue defends the cash position, or refresh.
Fourth, is the buyer the same person? Same buyer, and migration is feasible with one relationship carrying both products. Different buyer means split. The two products need different sales motions, different CS playbooks, different positioning, and running both under one product team breaks one or both.
Routing to the three modes
Take the four answers and route. Sunset when the architecture cannot extend, the customer base is the right ICP, and the trough is affordable. Refresh when the architecture can extend, the base is the right ICP, the buyer is the same, and AI is a feature improvement rather than a business model replacement. Split when the customer base is the wrong ICP, or the buyer is different, or the trough is not affordable without the legacy revenue.
Most CPOs I work with want the answer to be refresh because it is the least disruptive. In 2026 the architecture and pricing realities of agent products are pushing more decisions toward sunset than most CPOs are comfortable with. The honest test is the four questions, not the comfort level. The cleanest tell for refresh: can you publish a single roadmap that contains both products without lying about how they relate.
Run the four questions on a single page before Friday. Show the page to your CFO, just the CFO, and ask what would have to be true for the gross margin trough to be acceptable to the board. Their answer is your next month of work.