The Work a Machine Should Do

Last week I ended with a promise. I said I would move this conversation out of the strategic layer and into the operational one, down to where the repetitive, high-labor work actually sits inside a school district.

We have already established that some of the safest opportunities for AI sit outside the classroom. They live in the operational systems underneath it: finance, HR, procurement, transportation, facilities, and the other functions that keep a district moving.

“Start in the back office” is a location. It tells a leader where to walk and nothing about what to do once they arrive, and some of the work that looks easiest to hand to a machine contains exactly the judgment we should be most reluctant to give away.

So the question shifts. What work should we actually give the machine?

Walk through almost any central office and you will find people doing things machines are very good at.

Someone receives information in an email, keys it into a system, then keys part of it into a spreadsheet. Someone else downloads a report, rearranges the columns, cleans up the formatting, and sends it to a colleague who pulls pieces of it into a second report.

Employees check whether forms came back and send reminders when they did not. They compare two systems that hold the same data, assemble recurring documents, update status lists, and answer the same basic question forty times a month.

Those are real jobs held by capable people. We have also built a lot of those jobs around work that stopped earning its keep years ago, and that distinction is where the operational AI conversation should start.

The first question most districts ask is what AI can do for them. A better one: where are our people spending time doing work that a person should not have to do?

Start with the technology and everything begins to look like an AI opportunity. Start with the work and you will sometimes find that AI is beside the point, because a workflow rule solves it, or a system integration, or a better intake form, or removing four approvals from a request that never needed seven.

Sometimes the smartest automation strategy is to stop doing the work.

This is an old operational lesson wearing new clothes. Automating a poorly designed process makes the bad process run faster, and the result is a district that becomes efficiently inefficient.

I have written before that AI is entering school systems through points of pressure, which is predictable. People reach for tools where the workload is heaviest, the friction is greatest, and the need for capacity is most immediate.

Pressure is a useful signal. Treating it as a diagnosis is where districts lose money.

Consider an HR specialist who spends six hours a week reconciling information across four spreadsheets. We have found the pressure and we have not yet found the answer.

Maybe AI should perform the reconciliation. Maybe workflow automation should move the information so that nobody touches it. Maybe the two systems should be integrated until the reconciliation disappears. Maybe the report at the end of all that work stopped being read three superintendents ago.

Pressure tells you where to look. What to do when you get there is a separate decision, and it is the one that determines whether the money was well spent.

This is why I would be careful about walking a district asking employees what repetitive tasks AI could take off their plate. The question sounds reasonable, and it quietly assumes the task deserves to survive.

Ask them to show you where the work is hard, then find out why it is hard.

In operational improvement work, the most important activity often never appears on an org chart or in a procedure manual. It lives in spreadsheets, email folders, personal checklists, shared drives, and side databases. It lives in the memory of the employee who knows that after Step 7 you have to call Susan, because otherwise Step 8 never happens.

I think of this as the hidden factory, the unofficial work people invented to make the official system function.

Most of it sits in the handoffs: HR to payroll, a school to the central office, one finance system to another, the follow-up after an application is submitted, the routine communication between departments, and the report assembled from information already stored in four other places.

Most of that labor surrounds the core work of a department without being part of it. It is friction, and removing friction takes nothing away from the employee. It usually reveals them.

When the hours spent moving, cleaning, chasing, formatting, checking, and reentering information come off the calendar, what remains is the work we hired the person to do: analysis, coordination, and judgment.

Here is where the distinction gets harder. Some work looks repetitive because the same general situation arrives every day, while the activity underneath it is judgment.

Take an employee handling parent concerns. The calls come in daily and the process looks identical every time. Receive the concern, review the information, respond.

A machine can help here, and meaningfully. It can organize the information, summarize prior communication, draft a response, identify the applicable policy, and route the issue to the right person.

Owning the work is a different matter. The value of that employee sits in understanding what the parent actually means, recognizing the history behind the issue, sensing whether the situation is escalating, knowing who needs to be pulled in, and choosing how the response lands.

The machine sees the transaction. The experienced employee sees the system around the transaction.

That is the same boundary I have been describing at the strategic level. AI can generate an answer, understanding the constraint is a second capability, and owning the decision is a third.

The principle holds for a superintendent weighing district strategy and for an HR employee working an unusual personnel situation. As work moves toward ambiguity, consequence, context, and relationships, human judgment carries more of the load.

When you find high-labor, repetitive work, you have five options.

1. Eliminate it.

Start here, and start with purpose. Ask whether anyone reads the report, whether the approval is actually required, and why the information gets entered twice. Ask whether the control still mitigates a real risk, or whether it survives because it has always been there.

There is little value in paying to automate work that should disappear.

2. Simplify it.

Some processes serve a legitimate purpose and have picked up complexity along the way. Five approvals became seven, one form became three, and a simple request became a twelve-step workflow. Strip the complexity that creates no value before layering intelligence on top of it.

3. Automate it.

This is where machines earn their keep: moving information, matching records, checking explicit rules, monitoring thresholds, sending routine notifications, populating documents, compiling recurring reports, routing transactions, and flagging missing fields.

When the input is predictable, the rules are known, exceptions are rare, and the outcome can be defined clearly, a person may never need to touch the transaction at all.

Notice something. Much of that list requires no AI whatsoever, and ordinary workflow automation still solves plenty of these problems more cheaply, more predictably, and with a far easier audit trail. AI gets no bonus points for being newer.

4. Augment it.

Other work benefits enormously from a machine and should stay human-owned: summarizing information, drafting routine communication, comparing documents, surfacing anomalies, organizing large volumes of data, and preparing a first analysis for someone else to challenge.

Here the machine accelerates the thinking and the thinker stays in the chair.

This is probably where most of the near-term value sits for districts, because it releases capacity without asking leaders to surrender judgment.

5. Protect the human judgment.

Some work should move slowly, if at all. Interpreting ambiguous situations, resolving exceptions, making tradeoffs, negotiating, coaching, determining consequences, and deciding where context matters as much as the data all belong to people, along with any call someone eventually has to answer for.

AI is now fluent enough that a well-written answer reads like a sound decision. Producing an answer and owning a decision are separate capabilities.

Share

Before deciding what to do with repetitive work, run it through five questions.

Why does this step exist? If nobody in the building can explain what it protects or produces, the conversation starts with elimination rather than automation.

Is the input predictable? Roughly the same type of information should arrive in roughly the same form every time.

Are the rules explicit? Ask whether you could teach a new employee to process the transaction without leaning on experience or interpretation.

How many exceptions occur? One in a hundred, and automation may be ideal. Four in ten, and the process is not standardized yet, so automating it will move the interpretation into an exception queue and call that progress.

What does the employee know that the system does not? History, relationships, local context, policy nuance, and a promise made at a board meeting six months ago all shape the answer. If that knowledge routinely changes the outcome, the work is not ready to hand over.

Then one more, and it matters most of all.

What happens when the machine is wrong? A machine formatting a monthly report incorrectly is an inconvenience. A machine shaping a personnel, disciplinary, financial, or student decision incorrectly is a different category of problem, and oversight should scale with the consequence of the error.

If I were starting this work inside a district, I would leave the product catalog alone and go walk processes.

Pick HR, finance, transportation, purchasing, student services, special education administration, facilities, and technology support. Ask employees to show you what happens from the moment work enters the department until it leaves, including the steps that never made it into the manual.

Look for the same information typed twice, the reports that take hours to assemble, the checks performed by hand, and the questions that require opening two systems before anyone can answer them. Look for the spreadsheets that exist because the primary system does not produce what a department needs. Look for the employee who has quietly become the integration between two systems that do not talk.

Those are operational signals, and each one ends in a decision: eliminate it, simplify it, automate it, augment it, or protect it as human work. That is a better starting point than a platform selection.

There is a bill attached to this. Process walking takes time from the same people who are already stretched, and it usually surfaces work that somebody built, defends, and takes pride in. Budget four to six weeks per function, and expect the first two conversations in every department to be defensive. That reaction is normal when people feel audited, and it eases once employees see the exercise removing work instead of counting it.

District work is not getting simpler. Compliance demands hold, parents expect communication, and labor constraints now reach into departments across the organization. Financial pressure raises the bar on analysis, and new technology creates its own responsibilities even as it retires old ones.

Counting the positions AI can replace is the narrow version of this question. The larger one: how much human capacity can we release from low-value processing and redirect toward judgment, relationships, problem solving, and leadership?

Imagine pulling five hours of repetitive administrative work out of a principal’s week, or three hours of data manipulation away from an HR specialist, or removing the routine transaction work that eats part of a finance employee’s day.

That creates capacity. What the district does with that capacity is a leadership decision, and it should be made deliberately rather than absorbed by whatever fills the gap first.

I will admit I do not have a clean answer for what a district should do with released capacity in a year when the budget is shrinking. The pull toward converting it straight into savings is real, and so is the cost of doing that before anyone has redesigned the role. That tension deserves to be settled in the open.

I have argued that districts should build AI capability from the desk upward. People learn the tools inside real work, departments capture recurring work as shared capability, and enterprise platforms come later.

There is another reason that sequence matters. The people closest to the work know where the machine belongs.

They know which five-minute task happens fifty times a week. They know which spreadsheet takes half a day to rebuild. They know which simple transaction actually contains six judgment calls. They know which workaround exists because the official process does not work. And they know which step everyone could stop doing tomorrow without anyone noticing.

That knowledge is part of the capability districts need to build.

Within a few years, most districts will have access to similar AI tools, so the differentiator will be operating discipline. Districts that understand their processes will know where machines can safely take over. Districts that understand their people will know where machines should only assist. Districts that understand their systems will spot the work that should be eliminated before anyone tries to automate it, and districts that understand judgment will know where the machine stops.

The advantage will belong to the district that understands its work well enough to know what the machine should do, what the human should do, and what nobody should be doing at all.

We are going to hear a great deal over the next few years about what AI can do, and I am increasingly convinced that is the less interesting question. The harder one is what we should allow it to do.

That answer will not arrive in a vendor demonstration or a list of use cases. It comes from understanding the actual work inside the organization: its purpose, its friction, its exceptions, its risks, and the judgment people apply to keep it moving.

So before asking what AI could automate in your district, follow the labor. Find the work, understand why it exists, and then decide who, or what, should be doing it.

The goal is to need less human effort for the work that does not require being human.