Value AI Solutions Book a discovery call
← Journal

How to tell which work is actually worth automating

A dense field of dark geometric blocks in shadow, with five lit from within in warm amber

Most automation projects start in the wrong place: someone books a demo before anyone has named the task worth fixing. The tool gets chosen, then the work gets bent to fit it.

The question is never whether something can be automated. Almost anything can. The question is whether it should be — and that is the step most projects skip.

Automation projects fail for reasons that aren't technical

The failure rate is not a secret. An Ernst & Young study, quoted by Forbes, found that 30% to 50% of robotic process automation initiatives globally do not succeed. In manufacturing, a Vention 2025 survey found nearly one-third of automation projects do not perform as expected — with the top reasons being picking the wrong technology (50%), lack of internal expertise (39%), and cost overruns (32%).

Read that list again. Not one of those is a software defect. Wrong task, wrong tool, wrong team — the build was never the hard part.

Newer technology has not fixed this. Fiddler AI reports that AI agents fail between 70% and 95% of the time in production settings, depending on task complexity and how you measure success. On the WebArena benchmark, the best GPT-4-based agent completed only 14.41% of tasks end to end, against 78.24% for humans. Reliability degrades under repetition too: the same source cites an agent success rate dropping from 60% on a single run to 25% across 8 consecutive runs.

An agent that works in a demo and fails one run in three is not a system. It is a liability with a nice interface.

The wasted hours are real — but they're scattered

There is no shortage of evidence that manual work eats the week. Clockify reports employees spend 62% of their time on repetitive tasks and 4 hours 38 minutes per week on duplicate work. Asana's Anatomy of Work Index puts 60% of time at work on "work about work" rather than skilled work. Smartsheet found over 40% of workers spend at least a quarter of their work week on manual, repetitive tasks.

Data entry earns its own line. A typical office worker spends 10% of their time on manual data entry into business applications, and 1.5 hours a week copy-pasting between applications.

Here is the trap in those numbers. They are aggregate, and aggregate hours do not automate. Sixty-two percent of the week spread across two hundred small, varied, judgment-laden interruptions is not an automation opportunity — it is just what a job looks like. The hours worth recovering are the ones that clump: the same sequence, on the same trigger, many times a month.

Finding those is the actual work.

Four questions that disqualify most of the list

Every operator has a list of things that annoy them. That list is not a roadmap. These four questions are the filter we run before proposing a build, and they usually kill most of the list on the first pass.

Is it repeated, or just recurring?

A task run 30 times a month in the same sequence is a candidate. A task run once a quarter, with different logic each time, is not — no matter how much it is resented. Frequency without consistency is not automatable; it is just frequent.

Can the decision rule be stated in one sentence?

"If the order total is over the threshold, route to manager approval" is a rule. "Use judgment based on the client relationship" is not — it is expertise that has not been written down yet. If nobody in the room can articulate the rule, the automation will encode a guess.

Does the data already live in a system?

Work that requires opening five tabs to assemble one number is expensive to automate and usually points to a data problem underneath. Fix where the data lives first. Automating retrieval from a mess produces a faster mess.

What breaks when it goes wrong?

A misrouted internal notification is an annoyance. A miscalculated customer refund is not. High-consequence steps still get automated — they just keep a human checkpoint inside the flow. Decide which one you are dealing with before, not after.

Run a real example through it. Reconciling shipment confirmations across three carrier portals against an internal tracking sheet, done roughly 25 times a week at about 6 minutes each, following the same three rules every time, with a human review step before anything is finalised. That clears all four: repeated, rule-based, fully digital, low-risk. (Illustrative figures — your volumes are the ones that matter.)

Most items on the list will not clear it. That is the point. The filter's job is to tell you what to stop considering, so the effort goes to the two or three things that survive.

Measure the before, or you have no after

If you cannot say how long the task takes today, you cannot say what the automation saved. "It feels faster" is not a result, and it is not something you can take to a board.

Before building anything, capture the baseline: how many times the task runs per week, how long one pass genuinely takes end to end, how often it has to be redone, and what a mistake costs when it happens. Take it from observation, not from asking people to estimate — self-reported task times are consistently wrong, usually in the flattering direction.

Then measure the same four things after. The delta is the entire argument for the project. Everything else is vendor arithmetic.

We are deliberately not publishing a payback table here. The honest answer is that ROI depends on volume, loaded cost, and error rate in your business, and any generic percentage you have read was measured somewhere that is not your company.

Start with the work, not the workflow tool

The highest-leverage decision in an automation project is made before a single scenario is built. Pick the wrong task and the best implementation in the world joins the failure statistics. Pick the right one and the build becomes the easy part — usually wiring together tools you already pay for, rather than adding new ones on top.

Find the work worth automating first. Then build it.

Sources