Process Automation Strategy: A Practical Guide

A process automation strategy is a short, ranked list of the processes worth automating, the method each one needs (fixed rules, AI, or a person), the tool it will run on, and the numbers you will use to judge it. Most strategies fail because they start with a tool and go looking for uses. Start with the work instead: find where the repetitive hours go, score each candidate, automate one, measure it, then decide on the next.

The rest of this guide is the working version of that: the scoring criteria, the rule-based versus AI decision, the tools, a rollout plan, an honest ROI calculation and the ways it usually goes wrong.

Step 1: list where the repetitive hours actually go

Do not begin with a workshop on "automation opportunities". Ask the people doing the work what they spend Monday morning on, and write down each repetitive task with:

• what triggers it (an email arrives, a form is submitted, the month ends);
• how often it runs and how long one run takes;
• which systems it touches, and whether those systems have an API;
• how often it goes wrong, and what happens downstream when it does;
• who owns it today.

A few days of this usually produces a list of ten to thirty candidates. Typical ones: copying figures from PDF invoices into accounting, moving enquiries from an inbox into a CRM, routing support tickets to the right person, reconciling two lists that should match, producing the same weekly report by hand.

Step 2: score which processes to automate

Score each candidate from 1 to 3 on six criteria. The point is not precision; it is to stop the most visible process winning by default.

Criterion	1 (weak candidate)	3 (strong candidate)
Volume	A few times a month	Many times a day
Time per run	Under two minutes	Fifteen minutes or more
Rule clarity	Judgement on every case	Clear rules, few exceptions
Input structure	Free-form, varies wildly	Same fields, same format
Stability	The process changes monthly	Unchanged for a year
System access	Data only on screens	APIs or exports on every system

Then apply two filters before you rank anything:

• Cost of being wrong. If an error touches money, contracts or a customer relationship, the automation needs a human review step. That does not rule it out, but it changes the design and the payback.
• Is the process itself sound? Automating a broken process produces wrong answers faster. If people disagree about how it should work, fix that first. Sometimes that is the whole job, and it is a cheaper one.

Rank by total score, then sanity-check the top three against the hours they consume. A process scoring 16 that takes two hours a month is a worse first project than one scoring 14 that takes ten hours a week.

Step 3: rule-based or AI automation?

This is the decision most strategies get wrong in both directions.

Use rule-based automation when the input is structured and the logic can be written as "if this, then that": a form submission creates a CRM record, a paid invoice updates a spreadsheet, a new order sends a Slack message. It is cheap to run, predictable and easy to test. Most of a typical automation list is this kind of work.

Use AI (a language model) for the one step that rules cannot handle: reading an unstructured document, classifying a free-text email, extracting fields from a supplier invoice in an unexpected layout, drafting a reply for a person to approve. In good designs the model does one small part and ordinary integration work does the rest.

Keep a person where the cost of a mistake is high and the volume is low. Full automation suits some tasks and is reckless for others.

Two practical points about AI steps. They make mistakes, so log the model's output and its confidence, and send low-confidence cases to a review queue. And model usage is an ongoing cost that scales with volume, so it belongs in the ROI calculation, not as an afterthought.

Step 4: choose the tools

Pick the tool after you know the process, and prefer what you already run.

• Zapier. The quickest way to connect common SaaS apps with little technical skill. Billing is by task, so costs climb with high-volume or many-step workflows.
• Make. A visual scenario builder with more control over branching, iteration and data mapping than Zapier, billed by usage. A good fit for operations teams comfortable with some logic.
• n8n. A workflow tool you can run on n8n's cloud or self-host, with code steps in JavaScript or Python when the visual nodes run out. Self-hosting keeps data on your own infrastructure, which matters for some European businesses, but someone has to maintain the server.
• RPA (robotic process automation) tools such as UiPath or Microsoft Power Automate drive applications through their user interface, clicking and typing as a person would. They are the answer when a legacy system has no API, and they break when the screen changes, so treat them as a bridge rather than a foundation.
• Custom code. A scheduled script or a small service in Python or TypeScript. It wins when the logic is complex, the volume is high, or per-task pricing stops making sense. It needs a developer to change it.

A common pattern is a workflow tool for the orchestration, with a model call for the one unstructured step and a database (PostgreSQL, for example) holding the audit trail. If a scheduled script and one integration solves the problem, that is the right answer, even if it is less impressive.

Step 5: a rollout plan that survives contact with reality

1. Baseline. Before building anything, measure the process as it is: hours per week, runs per week, error rate, backlog. Without a baseline you cannot judge the result.
2. Build the smallest useful version. One process, end to end, in production. Not a pilot that runs on test data.
3. Design for exceptions. The happy path is a fraction of the work. Decide what happens when the input is in an unexpected format, an API is down, or two records match equally well. Exceptions should land in a queue a named person checks, not disappear.
4. Run in parallel. For two to four weeks, let the automation run alongside the manual process, or with every output reviewed. Compare results.
5. Hand over. Write down what it does, where it logs, who gets alerted when it fails and how to switch it off. Give it an owner inside the business.
6. Measure the same numbers again, then decide whether the next candidate on the list is worth doing.

A win somebody can point at buys more support inside an organisation than a roadmap does.

Step 6: measure ROI honestly

The arithmetic is simple. The honesty is in including every cost.

• Monthly saving = hours saved per month × the loaded hourly cost of the people doing it now, plus the cost of errors avoided, if you can price them.
• Monthly running cost = tool subscription + model and API usage + hosting + maintenance time.
• Payback period = build cost ÷ (monthly saving − monthly running cost).

A worked example, with invented round numbers: a task takes 8 hours a week, about 35 hours a month. At a loaded cost of €30 an hour that is €1,050 a month. If the automation removes 80% of the work and costs €150 a month to run and maintain, the net saving is about €690 a month. A €4,000 build pays back in roughly six months.

Two rules keep this honest. Count only hours that are genuinely freed for other work, not hours that dissolve into the day. And if the payback is longer than about a year, or the process runs too rarely for it ever to arrive, do not automate it. Some processes are cheaper done by a person.

For what an engagement costs with us, see our pricing: €25 to €45 an hour with a €1,000 minimum, and AI automation is quoted after a review of the process.

Common failure modes

• Tool first, process second. Buying a platform and then hunting for uses. You end up with several tools and no idea whether any of them helped.
• No baseline. Nobody measured the process beforehand, so nobody can say whether the automation worked.
• Automating a broken process. The errors arrive faster and look more official.
• Ignoring exceptions. The demo works; production fails on the first odd input, and nobody notices because failures are silent.
• No owner. The person who built it leaves, a password changes, and the workflow stops without anyone being told.
• Unstable processes. If the process changes every month, the automation needs changing every month too.
• Forgetting running costs. Per-task pricing and model usage grow with volume; the business case should survive the volume you hope to reach.
• Skipping adoption. People route around a tool they do not trust. Show them the reasoning, keep the review step, and let confidence be earned.

Frequently asked questions

What are the main process automation strategies?

In practice there are three, and a good plan mixes them per process: rule-based workflow automation for structured, predictable tasks; AI-assisted automation where one step needs to read or classify unstructured input; and RPA as a stopgap for systems with no API. The strategy is the order in which you apply them, driven by a scored list rather than by the tool you happen to own.

What should a process automation strategy include?

An inventory of repetitive work, a scoring method, a ranked shortlist, the method and tool chosen for each item, a baseline for each, a rollout plan with an owner, and the ROI test each automation has to pass.

Which processes should be automated first?

The ones with a knowable cost today, high volume, clear rules and stable inputs. Usually that means document handling, moving data from email into a system, ticket triage or reconciling lists. They are unglamorous, and they are where the hours are. If you are weighing AI projects specifically, our guide to choosing the first AI use case covers that decision in more depth.

Is RPA the same as process automation?

No. RPA is one technique within process automation: software that operates other software through its screens. Where systems have APIs, integration through a workflow tool or code is usually more reliable and cheaper to maintain.

How long does a first automation take?

It depends on the systems and the exceptions more than on the main logic. Plan for the parallel-running period in Step 5 on top of the build, and treat any estimate that skips it with suspicion. For a wider plan across a business, see our 90-day AI adoption plan.

Where does our website fit in?

Many processes start with a website form. If enquiries are being lost before they ever reach the automation, fix the form first; a free website roast will tell you whether that is the problem.