Stop Automating Decisions. Start Automating What Comes Before Them.

An operations manager at a mid-sized company told me it takes her about two minutes to decide whether an order can ship. She was not exaggerating. She looks at a summary, she checks one thing against another, she says yes or no. Two minutes, forty times a week.

Then I asked what has to be true before she can look at that summary. She started listing. Pull the account history. Check what shipped against what was billed. Cross reference the support tickets. Reconcile two spreadsheets that do not agree. Write the summary. The list took her longer to say out loud than the decision takes to make. When we timed it, the work in front of that two minute decision ran close to forty minutes, every single time, and nobody in the building had ever measured it.

That gap is where AI actually belongs, and almost nobody is pointing it there.

Individual Productivity Is Up. Nothing Else Is.

The evidence on this has stopped being ambiguous. McKinsey's 2026 State of AI survey, published in August, found that 80% of people who use AI in their roles say it has improved their individual productivity, while only 37% of organizations attribute any EBIT impact at all to AI, a share that has not moved in a year. The group seeing material financial impact, meaning five percent of EBIT or more, sits at 6%, unchanged.

Read those three numbers together and the picture is unmistakable. Individuals are faster. Companies are not different. MIT's GenAI Divide report said the same thing more bluntly, finding that widely adopted consumer tools "primarily enhance individual productivity, not P&L performance."

The most useful framing of why comes from Microsoft's 2026 Work Trend Index, which found that organizational factors account for more than twice the reported AI impact of individual factors, 67% against 32%, and that only 19% of AI users sit in the group with both high individual capability and high organizational readiness. The bottleneck is not the person and it is not the model. It is the shape of the work around both of them.

We Have Been Aiming at the Wrong Half of the Process

Here is what I think is actually happening. Almost every AI deployment I see is aimed at one of two places: the task, or the decision itself.

Aim at the task and you get what the numbers above describe. A writer drafts faster, a rep summarizes a call faster, an analyst builds a chart faster. Real, small, and capped at the cost of the task. Aim at the decision and you hit a wall of a different kind, because the person accountable for the outcome will not sign their name to an output they cannot explain, and they are right not to.

I have argued the second half of that before, in From Prediction to Prescription, and I want to complicate my own argument. The interesting frontier is not AI moving further into the decision. It is AI taking everything that stands in front of the decision and handing the person a call they can make in two minutes with the evidence already assembled.

Call it the assembly. It is the pulling, checking, cross referencing, reconciling and summarizing that has to happen before a judgment is possible. It is almost always the expensive half, and it is almost never measured, because no one bills for it and it does not belong to any one person.

The sales data makes the scale of it visible. Salesforce's 2026 State of Sales report, fielded across 4,050 sales professionals in 22 countries, found that reps spend 40% of the week selling and 60% not selling, with the largest non-selling blocks being manual data entry at 22%, prospecting at 18%, and building quotes at 16%. That 60% is not waste in the sense of laziness. Most of it is assembly. It is the work that has to happen before somebody can decide which account to call, what to say, and what to offer.

The Line Is the Design Decision

So the practical question stops being "where can we use AI" and becomes something much sharper: in this process, where exactly does judgment enter?

Every recurring decision in a business has a point where the work changes character. Above that point it is gathering, formatting and reconciling, and any two competent people doing it should produce the same result. Below that point somebody weighs things that do not have a single right answer, and their name goes on the outcome.

Find that line and you have both halves of your AI strategy at once. Everything above it is a candidate for automation, and the case for automating it is measurable in hours you can count. Everything at or below it stays with a person, and now that person gets the decision with the preparation already done rather than a blank page and four browser tabs.

This is a different design instinct than the one most teams start with. It does not ask what AI can do. It asks what the person actually needs in front of them at the moment they decide, and then works backward through everything that currently has to happen to produce it.

It also explains why the context argument I made in Context Is the Whole Game matters operationally rather than philosophically. The assembly work is context work. Pulling the account history and reconciling the two spreadsheets is a person doing by hand what a connected system should be doing continuously. When the context layer is real, the assembly mostly disappears. When it is not, you have people performing retrieval for a living.

How to Find Your Own Line This Week

You do not need a platform to do this, and you should not start with one.

Pick one decision your team makes most weeks that has money attached and a name on it. Not a task, a decision. Which accounts get worked. Whether this order releases. Whether we bid this job. Then ask the person who makes it what has to be true before they can decide, and write the steps down in order.

Now time them. Roughly is fine. Then multiply by how often the decision happens and by forty eight working weeks, and you have an annual hours figure for the preparation behind one decision. Almost everyone I have done this with is surprised by the number, and the surprise is the useful part. You cannot aim at something you have never measured.

Finally, draw a line across the list at the point where judgment enters. That line is your scope. Everything above it is the project. Everything below it is the reason the project is safe to do.

The Takeaway

The reason 80% of AI users feel more productive while only 37% of companies see any financial impact is not that AI does not work. It is that we keep pointing it at tasks, which are small, or at decisions, which people will not surrender.

The value sits in between, in the assembly nobody has ever timed. Automate up to the line where judgment starts and stop there, and you get the part machines are genuinely good at without asking anyone to sign their name to an answer they cannot explain.

We do not need AI that makes the decision. We need AI that makes the decision cheap to make.

Tracy Thayne* is the founder of Expona, an AI-powered operational intelligence platform for B2B marketing. Read the Expona founder story or subscribe to the blog (below) for weekly insights on context, AI, and the operating model of the next decade.*