The best AI candidate is not the task you hate most

The report you dread every month turns out to be hard for a reason. Which is its own small disappointment.

You thought it was drudgery. Same numbers, same slides, same questions from people who didn’t read the last version.

Then somebody asks where a figure came from.

Why a customer was excluded.

Whether the new definition still applies.

The misery wasn’t evidence that the work was empty.

It was evidence that the work had become where your organisation keeps its decisions.

You’re not failing because you want help with it.

You’re tired. There are deadlines, messages, people waiting for answers, and all the ordinary obligations that don’t get less real because a process map thinks they should.

When a model offers you a draft, taking it isn’t laziness or betrayal. It’s a sensible response to work that keeps asking for the same pieces of you.

This has no villain in it. It needs a better first move.

The task you hate most may be the worst place to begin, because it’s carrying the exceptions nobody named.

2015 — You wish the monthly report would make itself.

2020 — Somebody maps the process and finds seven exceptions.

2024 — The model produces a draft.

2026 — You still own the seventh exception. It just arrives faster.

A task is usually several

“Automate the monthly report” isn’t a decision.

It’s a wish wearing a lanyard.

Your useful unit is a bounded case. The trigger. The authoritative sources. The normal route. The decision that matters. The evidence behind it. The outcome. And the point where a case stops being ordinary.

Which is why GSA’s CODY application has three pathways — report, checklist, awardee memo — rather than one grand promise to automate procurement.

It gathers and prepares.

The procurement authority stays with a person.

That sounds fussy until your work touches a contract, a payment, an audit finding, or a customer whose situation doesn’t fit last month’s categories.

A rules engine suits a stable policy. An unattended bot bridges a repeated handoff between systems. Document processing extracts a defined class of document. An assistant prepares an answer while you stay accountable for sending it.

Those aren’t four levels of cleverness.

They’re four different boundaries.

The old question was whether the machine can do it.

The better question is which part of this case becomes too consequential to guess.

That’s the correction. You’re not choosing between human and machine.

You’re choosing where a case becomes too consequential to guess.

GSA’s contract-closeout example supported more than 12,000 closeout transactions a year and estimated around 25,000 staff hours saved.

That doesn’t prove your closeout work pays back. It shows you the shape of a good candidate.

Repeated document and reconciliation work under a stable rule set, while claims, missing evidence and disputed financial actions have somewhere else to go.

The number isn’t the argument. The frontier is.

Volume is an invitation

Volume has a persuasive sound. It makes your unpleasant job look like an obvious saving waiting to be collected.

But 11,531 requisitions processed across four months, at a reported fifteen minutes of manual labour each, describes a strong signal. Not a universal answer.

The short period matters.

The system changes. Credentials expire. An error message changes its wording. And the happy path starts sending its failures into your afternoon.

Which is why an invoice reminder can be a better candidate than invoice judgment.

GSA’s payables bot starts your reminders on day one, escalates to supervisors on day fifteen, marks an invoice past due on day thirty.

Explicit, repeated decisions on a clock. It doesn’t authorise a payment or decide what a disputed invoice means.

The system owns coordination. You still own the defective case.

And variability isn’t a nuisance at the edge of your work. It’s a compounding cost.

Each new input format can combine with a rule, a source system, an account type, an authority, or a customer condition of yours.

It adds tests, monitoring categories, training, reconciliation, and fresh chances for a later change to break something of yours that used to work.

The selection frameworks make this point less elegantly than they should. In one ranking of thirty-two criteria, number of exceptions and repetitiveness tied at seventeenth.

Don’t carry that into a meeting as though it settled anything. Read it as a refusal to treat exceptions as an implementation afterthought.

Your ranking gets honest only when it names volume, variants, data, rules, costs, benefit mechanism and the human route together.

Where the old work returns

The human remainder isn’t an embarrassment in your design. It is the design.

Once a bounded transformation is automated, what’s left concentrates into judgment, relationship management, quality control, change ownership, and the cases nobody can safely make ordinary.

McKinsey’s analysis found 4% of work activities required creativity at a median human-performance level, and 29% required sensing emotion at that level.

Those aren’t a comforting map of what machines will never touch. They’re a reminder that an occupation is a mixed bag, and the residual bag can get heavier.

A known exception needs an evidence bundle, a service target and a named owner. An unknown failure needs containment and investigation, and maybe a change to your boundary.

Without that distinction, “human in the loop” means you repair the model’s errors with no queue, no authority, no measure and no budget.

The difficult work didn’t disappear. It became uncountable.

The audit precedent is blunt about the good version.

An Army bot identified transactions for 96% of audit samples in a day, where the manual process took five to ten.

It supported deeper review. It never replaced the auditor’s conclusion.

That’s partial automation at its best. Retrieve and arrange the evidence faster, and refuse to turn the meaning of the evidence into a silent output of the retrieval system.

You might read all this as an argument for keeping the boring work human. It isn’t.

Some of your boring work should go. The report can assemble faster. The account history can be retrieved. The reminders can go out without you carrying a date in your head.

But a process turns cruel when it calls the meaningful remainder a failure of automation, rather than the condition that made your job matter in the first place.

A draft is not a service

The model’s draft is the easy moment. It arrives cleanly, speaks in the approved tone, makes the work look solved.

Enterprise adoption starts where your feeling ends.

Your pilot needs a defined queue, approved sources, a stated human-only boundary, an event that ends it, and a baseline. Request volume, handling minutes, recontact, quality, escalation, cost per completed case.

Without the baseline, time saved is an assertion with a good interface.

Then the records have to become something more demanding than a shared drive.

A data product has an owner, defined meaning, access rules, a way for you to correct a defect, and a current source of truth.

Your system has to show which source supported a draft, and revoke a document or entitlement when it shouldn’t be there any more.

The pleasant chat box isn’t exempt from permissions because it’s pleasant.

Before release, the configured service meets held-out cases and adversarial ones. Another customer’s data. A prompt buried in a source document. A stale policy. A case that must go to a person.

An average accuracy number hides one unacceptable error inside the cases that carry your consequence.

Faster is not cheaper

Your first release should be narrow enough to reveal real work. A small trained group, or one queue.

Measure your eligible cases, drafts shown, accepted-without-material-edit rate, material edits, overrides, time-to-send, sampled accuracy, and what happened downstream.

A high acceptance rate may be trust. It may be passive rubber-stamping.

A low rate may be rejection. It may be one repairable source.

Usage, workflow performance and financial return are three different things.

Support, source refresh, model-change handling and a manual fallback belong at the beginning of a real service. Not after the launch announcement.

The task that hates you hasn’t been removed. It has acquired a run rate.

And your saved minute doesn’t become a saving by being impressive.

It becomes cash only if your employer avoids hiring, cuts overtime or contractor spend, closes a site, or removes a contracted service.

With staffing fixed, it buys throughput, resilience, speed, or time for better work instead.

Those can be good outcomes. They’re not interchangeable outcomes.

So the last gate was never whether people used it.

Your review reconciles case mix, seasonal demand, training time, new rework, supplier consumption, and the capacity nobody could actually redeploy.

Then it gives you three honest endings. Fund an extension. Hold the service where it is. Or retire it and keep the lesson.

NIST notes that simple payback excludes every benefit and cost after the payback period. Which is why a quick recovery story is biased against a slower, stronger return.

The opposite error is easier to spot. A rescue mission for a pilot, because nobody wants to admit the seventh exception was the work.

You’re not wrong to want the report to make itself.

You’re not sentimental for noticing the hardest case is still yours.

The work people most want to hand away is often variable, politically loaded, or full of exceptions. That’s precisely where an organisation hid its decisions.

Let the ordinary case go quickly. Let the draft arrive.

But don’t let anybody call your remaining judgment a defect just because it finally became visible.