Which Business Processes Should You Automate First? A Practical Framework

Ask ten vendors what your business should automate and you'll get ten answers, each one matching whatever that vendor happens to sell.

This article gives you the answer we'd give if we were sitting on your side of the table: a way to rank your own processes, decide what genuinely deserves automation, and — just as important — what to leave alone.

The short version: automate the processes that score highest on three tests — how much friction they cause, how frequently they run, and how much they follow rules rather than judgment.

We call this the Friction–Frequency–Rules test, and the rest of this article shows you how to apply it.

Start with problems, not technology

Most automation projects begin the wrong way around.

Someone sees a demo, buys a tool, and then goes looking for a place to use it. Six months later the subscription is still active and nobody remembers why.

The businesses that get real value from AI start with a different question:

Where does our time actually go?

Not in theory — in practice, this week.

Before you evaluate a single tool, spend one week having each team note the tasks that eat their hours.

You are looking for the unglamorous stuff:

• Copying data between systems
• Answering the same customer question for the fortieth time
• Chasing a lead who filled in a form three days ago
• Formatting the same weekly report
• Moving information from emails into spreadsheets
• Sending repetitive reminders

That list — not a vendor's feature sheet — is your raw material.

The Friction–Frequency–Rules test

Score every candidate process from 1 to 5 on each of three questions.

1. Friction: how much does this process cost you?

Count the hours, but also count the errors, the delays, and the things that never happen because nobody has time.

A task that takes ten minutes but blocks a sale for two days has high friction.

A tedious task nobody is waiting on has less than you think.

2. Frequency: how often does it run?

Automation is a fixed investment repaid per execution.

A process that runs fifty times a day repays it quickly.

A process that runs quarterly almost never does — however annoying it is.

This single test kills most bad automation ideas.

3. Rules: could you write down how it's done?

If an experienced employee could describe the steps and decision points on one page, the process follows rules, and today's AI handles it well.

If their honest description is:

"It depends — I just know."

You're looking at judgment.

Modern AI has pushed the boundary here. It can now read an email and classify what the customer wants, draft a reply in the right tone, or extract details from an invoice — tasks that were "judgment" five years ago.

But genuine judgment — pricing an unusual deal, handling an angry key account, deciding whom to hire — should stay human, with AI at most preparing the groundwork.

Score the process

Multiply the three scores.

Anything above roughly 60 out of 125 deserves a serious look.

The pattern of the score matters too.

High friction with low frequency means fix the process, don't automate it.

High friction, high frequency and high rules means you've probably found your first project.

What this looks like in practice

Consider an illustrative example: a 15-person trading company scores three candidates.

ProcessFrictionFrequencyRulesScoreVerdictAnswering price/availability enquiries45480Automate firstFollowing up quotes that went quiet54480Automate firstNegotiating annual supplier contracts4114Keep human

The two winners share a profile you'll see again and again:

Customer-facing. High-volume. Rule-shaped. Currently done badly because people are busy.

In our experience of mapping processes with businesses, the highest scores cluster in the same four areas almost every time:

• Lead follow-up
• First-line customer enquiries
• Appointment scheduling and reminders
• Moving data between systems that don't talk to each other

Five things you should not automate first

A prioritisation framework is only useful if it also says no.

Leave these alone in round one.

1. Broken processes

Automation makes a process faster, including a bad one.

If your quoting process produces wrong prices, automating it produces wrong prices at scale.

Fix first. Automate second.

2. Rare, high-stakes decisions

Low frequency, high judgment is the worst possible automation candidate, whatever a demo suggests.

3. Processes you can't measure

If you don't know the current response time, error rate, or hours spent, you will never know whether automation worked.

Baseline first.

4. Anything where the relationship is the product

If clients pay you partly for the feeling of being personally looked after, automate the logistics around the relationship — reminders, scheduling, records — not the relationship itself.

5. Everything at once

One process, done properly, measured honestly, then the next.

Businesses that launch five automations simultaneously usually end up with five half-working ones and a sceptical team.

Sequencing: the first ninety days

Once the FFR test has produced a shortlist, sequence it with three tie-breakers.

Visibility of the win

Your first automation is also an internal sales job.

Pick something whose result the whole team can see.

"Every lead now gets a reply within two minutes" beats "our data hygiene improved."

Data readiness

An automation is only as good as what it can read.

If the process's information lives in one system — or even one inbox — you can start now.

If it's scattered across spreadsheets, WhatsApp threads and someone's memory, budget time for consolidation — or pick a different first project.

Reversibility

Prefer automations you can switch off without drama.

Customer-facing AI with a clean human hand-off is reversible.

Ripping out your invoicing system is not.

Then run it as a proper test.

Define the metric before you start, run the automation alongside the old way for a few weeks if you can, and review with real numbers.

If it worked, take the next item on the list.

This rhythm — one process per month or two, each one measured — compounds quietly.

Within a year it amounts to a different company.

Putting money on friction

One refinement separates teams that use the FFR test well from those that treat it as a parlour game:

Pricing the friction.

A score of 4 is an opinion.

"This costs us roughly $2,100 a month" is a decision-ready fact.

The arithmetic doesn't need to be precise — it needs to be conservative and written down.

Three prices to attach, one of which usually dominates:

Labour price

Hours per week × loaded hourly cost of whoever does it.

This is the price everyone calculates and, for customer-facing processes, usually the smallest.

Leak price

Revenue lost through the process's failures — the unanswered enquiries, the quotes never chased, the no-shows.

Estimate it from your own numbers:

Enquiries per month × share answered late × close rate × average deal

Then halve it, because optimism creeps into every one of those inputs.

Delay price

What waiting does downstream.

A two-day quote turnaround doesn't just cost the deal-in-hand; it costs the reputation for responsiveness that wins the next tender.

Hardest to quantify, so note it in words rather than forcing a number.

When you rank your shortlist by conservative monthly cost, the order often changes — and arguments end.

The back-office process everyone complains about loudly may cost a fifth of the quiet leak in enquiry response nobody had measured.

This priced list also becomes your negotiating anchor when providers quote: an automation should cost meaningfully less over a year than the friction it removes, and now you know that number before they do.

Before automating anything, try deleting it

One step belongs between scoring and buying, and it's free:

For each shortlisted process, ask whether it should exist at all.

Some friction is best removed, not automated.

The weekly report nobody reads gets cancelled, not generated by AI.

The approval step that exists because of a mistake in 2019 gets retired.

The data copied between two systems sometimes stops being copied because one of the systems was redundant all along.

In our experience, an honest pass of eliminate, simplify, then automate removes one or two items from most shortlists.

And the automations you do build afterwards are simpler, because the process underneath them got cleaner.

The same pass sometimes runs the other way.

A process scores low on frequency only because your current way of working suppresses it.

Follow-up is the classic case.

A business that "rarely does follow-up" doesn't have a low-frequency process; it has an absent one whose right frequency is high.

Score the process as it should run, not as neglect currently shapes it.

Where AI changes the old answer

Classic automation advice said: automate the repetitive, forget the rest.

That advice is now incomplete.

The interesting shift of the past three years is that AI handles messy inputs to rule-shaped work.

A customer email in mixed Arabic and English.

A photographed invoice.

A rambling voice note asking whether you deliver on Fridays.

These used to disqualify a process from automation because the input wasn't structured.

They no longer do.

When you apply the Rules test, score the decision, not the input format.

The decision "does this enquiry need a price list, a stock check, or a human?" is rule-shaped, even when the enquiry arrives as a voice note.

That's also why we'd gently push back on treating "AI" and "automation" as separate purchases.

The useful question is never:

"Should we get AI?"

It's:

"Which process, and what's the cheapest reliable way to run it without a human?"

Sometimes that's a simple workflow rule.

Sometimes it's an AI agent.

The process decides, not the fashion.

Three questions we hear at this point

"Should we automate before making our next hire?"

Run the numbers side by side.

A hire brings judgment, flexibility and growth capacity.

Automation brings consistency, speed and coverage of the hours no one wants.

If the work you're hiring for is mostly high-frequency and rule-shaped — answering enquiries, chasing follow-ups, scheduling — automation typically delivers most of the value at a fraction of a salary, and your eventual hire starts with the clerical layer already handled.

If the work needs judgment or relationships, hire, and give the person automated support from day one.

"What if nothing scores above 60?"

Then you've learned something valuable at zero cost.

Your operation is either too small, too varied, or too judgment-heavy for automation to pay right now.

Revisit in six months — frequency grows with the business — and spend the budget on whatever the scoring exercise revealed as the actual constraint.

A framework that can say "not yet" is the only kind worth using.

"How long does a first automation take?"

For the single-process, customer-facing projects this article points toward: typically a few weeks from decision to live, plus a month of tuning while it meets real customers.

Anything quoted in quarters, for a first project, is scoped too big.

The one-page summary

Inventory where the time goes.

Score each process on Friction, Frequency and Rules.

Automate high scores.

Fix broken processes before automating them.

Keep genuine judgment human.

Sequence by visibility, data readiness and reversibility.

One process at a time, with a baseline and a measurement.

And if a vendor's recommendation happens to score terribly on this test — that tells you something too.

‍