Which process to automate first — and which to leave alone

The question arrives in almost the same words every time. Can AI do this? Someone describes a process — quotations, supplier invoices, the WhatsApp enquiries that pile up overnight — and waits to be told yes or no.

The answer is nearly always yes. That is what makes it a bad question. Almost anything in a small business can be automated to some degree in 2026. The useful question is narrower and harder: which process first, and how would you know if you picked wrong?

This is the framework we use before writing a line of code, including the part most people skip — deciding what to leave alone.

The most annoying process is rarely the right first one

Start somewhere boring. The process that irritates you most is usually the one with the messiest rules, which is exactly why it irritates you.

Irritation is a signal about variability, not about volume. The task that makes someone sigh is often the one where every case is slightly different, the exceptions are the interesting part, and the person doing it is quietly making judgement calls they have never written down. That is the hardest thing to automate and the worst place to spend a first budget.

The right first candidate is usually something nobody complains about, because it is dull, repetitive and nobody thinks about it at all.

Four tests a first candidate has to pass

Run any process through these four before considering it. A candidate that fails one is not disqualified forever — it is disqualified from being first.

Test	Green light	Wait
Volume	Happens daily, or many times a week	A few times a month
Rule stability	The rules have held for a year or more	Changed twice this year
Cost of being wrong	A mistake is caught and reversed internally	A mistake reaches a customer or a ministry
Input quality	Arrives as text, files or form entries	Lives in voice notes, paper, or someone’s head

Volume: do the arithmetic before the demo

Automation is worth building when the hours it removes cost more than the system does. That is a calculation, not an opinion, and you can do it in ten minutes.

Take the loaded cost of an hour of the time being spent. In Kuwait, an administrator on KD 450 a month works roughly 176 hours, which is about KD 2.6 an hour before anything else. Add visa costs, indemnity accrual and the share of overheads that person carries and you land nearer KD 3.3. Use your own figures — the point is the method, not our number.

Now multiply by the hours the process actually consumes each month. Not the hours it feels like. Ask the person to count for one week.

Rule stability: are you automating a moving target?

If the rule changed twice this year, wait. You will spend the build reacting to changes and the result will encode a policy that is already out of date.

This is the test people resist most, because a changing rule feels like an argument for automation — surely software adapts faster than a person? It does not. A person absorbs a rule change in a corridor conversation. Software absorbs it in a change request.

Automate the stable core of a process and leave the volatile edge to a human. A quotation system that calculates the standard price list automatically and routes anything unusual to a person is worth building. One that tries to encode every negotiation is not.

Cost of being wrong: reversible or not?

Sort errors by who sees them. An error caught internally costs minutes. An error that reaches a customer costs a relationship, and one that reaches a ministry costs considerably more.

This is the axis that decides how much human review the design needs, and it is where most of the real engineering sits. The NIST AI Risk Management Framework is the reference we point clients to when they want the formal version: the useful idea is that risk is a function of consequence, not of how impressive the technology is.

In practice this means a first automation should sit where a wrong answer is embarrassing rather than expensive.

Input quality: where does the data actually live?

If the input is a WhatsApp voice note, you have two projects, not one. The first is turning speech into structured text. The second is the thing you actually wanted.

Processes that begin with typed text, structured files or form entries are ready now. Processes that begin with paper, photographs of paper, or spoken Arabic are entirely possible — we build them — but they carry a preprocessing stage, and that stage is where the accuracy budget goes. We wrote about exactly where that breaks in reading Arabic invoices with AI.

A worked example, including the part where we say no

A Kuwait trading company prepares quotations from supplier price lists. Sixty quotes a month, about twenty-five minutes each. That is 25 hours a month.

At KD 3.3 an hour, that time costs roughly KD 82 a month, or just under KD 1,000 a year. A quotation system that reads the price lists, applies margin rules and drafts the document is a real build: somewhere between KD 1,200 and KD 2,500 depending on how many suppliers and formats are involved, plus KD 30 to 60 a month to run.

The arithmetic says no. Payback lands somewhere past the two-year mark, and that is before anyone maintains it. We would tell this company to leave it alone and spend the money elsewhere.

Change one number and the answer flips. The same company at four hundred quotes a month is spending 167 hours, around KD 550 a month, KD 6,600 a year. The same build now pays for itself inside a quarter. Nothing about the technology changed — only the volume.

This is why we ask for the counted hours before quoting anything. If the number does not support the build, the honest answer is the cheap one.

What to leave alone

Some processes should stay human, and not only for the first project.

• Work where the exception is the job. If eighty per cent of cases are unusual, you do not have a process, you have a series of decisions.
• Judgement calls with no written rule. If two experienced people would answer differently, software will simply pick one of them and hide that it did.
• Relationship work. Chasing a late payment from a client of fifteen years is not a workflow.
• Low volume, high variance. Anything that happens five times a month and differently each time will cost more to maintain than to do.
• Anything nobody can describe. See below — this one is the most common by a distance.

If nobody can write it down, don’t automate it

Ask the person who does the task to write the steps on one page. If they cannot, the problem is not a missing system.

This exercise fails more often than anyone expects, and the failure is valuable. It surfaces the undocumented decisions the whole process actually rests on: the supplier who always gets a different margin, the customer whose orders skip approval, the exception everyone knows and nobody recorded.

Once it is written down, two things become true. You can automate it — and quite often, you no longer need to. Half the value of automation projects is the documentation they force, which is why we start with the page and not the platform.

Four mistakes we see repeatedly

• Buying the tool before defining the process. The platform is the last decision, not the first.
• Automating the exception instead of the rule. Handle the standard case perfectly and route the rest to a person.
• Reaching for a model where a rule would do. Google’s own engineering guidance opens with it: don’t be afraid to launch a product without machine learning. A lookup table that is right every time beats a model that is right most of the time.
• Measuring nothing before starting. Without the hours counted beforehand, you cannot tell afterwards whether it worked.

How we scope it

One process, thirty days, on your real data. That is the whole shape of our 30-day AI pilot, and the deliverable is a working system plus a number telling you whether to scale it or stop.

Both of those are acceptable outcomes. A pilot that ends with “the volume does not justify this” has saved you a year of paying for something you did not need — and it cost thirty days to find out instead of finding out afterwards.

If you want the longer view of what an agent does once it is running, that is AI agents and automation. If you already know which process you would start with, tell us about it on WhatsApp and we will do the arithmetic with you before anyone talks about building.

Frequently asked questions

What is the minimum volume that justifies automating something?

There is no universal threshold, because it depends on what the hour costs you and what the build costs. The test is whether annual hours saved, priced at the loaded hourly rate, exceed the build cost plus a year of running it. As a rough orientation, processes consuming under ten hours a month rarely clear that bar in Kuwait; processes over fifty hours a month usually do.

Should I fix the process first, or automate it?

Fix it first, always. Automating a broken process produces broken output faster and makes the breakage harder to see. Writing the process down on one page usually reveals the fix, and sometimes the fix is enough on its own.

Can you automate something that arrives as WhatsApp voice notes?

Yes, and it is common in the Gulf. Treat it as two stages: speech to structured text, then the actual workflow. Kuwaiti dialect handles well with current models, but plan for a review step at the transcription stage rather than assuming it is solved.

What if our rules change every few months?

Then automate the part that does not change and leave the rest to a person. A system that handles the stable eighty per cent and routes the remainder is more valuable, and far cheaper to maintain, than one that tries to encode every variation and needs rebuilding each quarter.

How do I know whether the pilot actually worked?

By comparing against the number you recorded before it started. Count the hours the process consumes for one week before anything is built, then count them again after. If nobody counted beforehand, the pilot will produce opinions rather than evidence, and opinions are what you were trying to get away from.