The 10% problem: what AI can't automate in SaaS

AI is compressing SaaS execution. It can write code, summarize calls, draft documentation, generate test cases, build analyses, triage support themes, and turn a rough instruction into a passable artifact. For many teams, it is reasonable to assume that roughly 90 percent of the production surface will become automated or heavily assisted.

That does not mean 90 percent of the company becomes easy.

The hard part moves. When production gets cheaper, the bottleneck becomes the remaining 10 percent: deciding what should exist, what should wait, what should be trusted, and which outcome deserves attention now.

What can AI not automate in SaaS?

AI cannot own the company's commitment. It can recommend a feature, but it cannot be accountable for the outcome a VP promised to the board. It can compare customer notes, but it cannot decide which customer segment should define the next six months. It can draft a migration plan, but it cannot carry the trust cost if the wrong one-way decision reaches production.

That is the 10 percent problem.

The 10 percent is not only "strategy" in a big annual-plan sense. It is the dozens of operating choices that turn strategy into reality:

• Which customer signal should override the roadmap?
• Which agent proposal is aligned with the current outcome?
• Which decision is reversible enough to move quickly?
• Which review needs a human owner before execution continues?
• Which tradeoff has drifted away from the principle leadership set?

These are judgement calls. They are also the calls that determine whether AI acceleration creates leverage or noise.

Why the scattered stack hides the 10 percent

Most SaaS teams already have the raw material. Jira has the ticket. Linear has the status. Slack has the argument. GitHub has the implementation. Productboard has the customer signal. Docs have the stated plan. Agent sessions have the reasoning that produced a change.

The problem is that the decision is spread across all of them.

A VP can see delivery velocity and still miss that execution has shifted away from strategy. A product leader can read customer notes and still miss that an agent shipped the old assumption. An engineering manager can review the PR and still miss the business reason that should have constrained the solution.

AI makes this harder because it increases the number of plausible artifacts. More work gets done. More options appear. More comments, diffs, summaries, and plans move through the system. Without a shared decision layer, the company gets faster at producing work while staying slow at understanding whether the work matters.

What intelligent decision-making in SaaS requires

Intelligent decision-making in SaaS starts by treating decisions as first-class operating objects. A decision should have an owner, a reason, a source trail, an outcome link, a reversibility level, and a next review point. It should be visible to humans and available as context for agents.

That visibility changes the way teams work.

When a new agent session starts, it should know the outcome the team is committed to and the tradeoffs already made. When a human reviewer opens a proposal, they should see whether the work follows the current strategy or revives a deferred path. When leadership reviews progress, they should see not only what shipped, but which decisions moved the number and which created drag.

This is where a judgement layer built on decision intelligence becomes different from another dashboard. Dashboards show metrics. Decision intelligence connects metrics to the choices that created them.

The enterprise risk is not slower work. It is faster drift.

Before AI, drift often looked like missed deadlines or swollen roadmaps. With AI, drift can look like progress. Tickets close. PRs merge. Summaries sound coherent. The machine keeps producing evidence that something is happening.

But if the work is pointed at the wrong outcome, speed becomes a liability.

The enterprise buyer does not need a motivational layer around AI. They need a judgement layer that lets leaders ask: what are we committed to, how did we get here, what should happen next, and who should we trust to propose it?

That is the 10 percent problem. AI can help produce the work. The company still has to govern the judgement.

For the naming and question framework behind Ask The W, read Why it's called Ask The W.