Which business processes should you automate first? 15 workflow examples with a scoring sheet

The rest of this page gives you the scoring sheet, a catalog of 15 workflows by department, and an honest look at tools and prices. Every vendor number carries a link and the date it was checked.

Short answer: the five processes most teams start with

Most first projects come from the same short list. The reason is boring: these processes repeat hundreds of times a month and follow rules somebody could write on one page.

1. Lead routing and CRM enrichment. A form fills in, the lead lands with the right rep, the record gets company data.
2. Invoice capture. Supplier invoices arrive by email and end up as draft bills in accounting.
3. Order sync. Store orders flow into the ERP or accounting system without retyping.
4. Overdue reminders. Unpaid invoices trigger a polite sequence of emails on a schedule.
5. Onboarding checklists. A signed offer creates accounts, tasks and a first-week plan.

How much work is in play? McKinsey Global Institute estimated in June 2023 that current generative AI and other technologies could automate work activities absorbing 60 to 70 percent of employees' time. Its earlier estimate was about half (McKinsey, "The economic potential of generative AI," June 14, 2023, as of September 30, 2026). That is technical potential, and it says nothing about which of your processes pays back first. The sheet below does.

How to score a process

Pick ten candidates and rate each one from 1 to 5 on five factors. Multiply the scores. Multiplication is deliberate: one weak factor drags the whole number down, and that matches reality. A frequent process with no written rules still fails.

Use 2 and 4 for the in-between cases. The formula:

score = frequency × minutes × error cost × rule clarity × data readiness

The maximum is 3,125. Read the result in three bands:

• 600 and above: automate first.
• 200 to 599: second wave, after the first project runs cleanly.
• Below 200, or a 1 on rule clarity or data readiness: fix the process before you touch software.

In a worked example, lead routing at 300 leads a month scores 4 × 2 × 4 × 5 × 5 = 800, so it goes first. A twice-a-year bonus decision scores 1 × 5 × 3 × 1 × 3 = 45. Leave it with a person.

Count exceptions honestly. Pull the last three months of real cases, not the process doc. If a third of them needed someone to decide, the project is removing that ambiguity. Automation comes after.

Here is the blank template. Copy it into a spreadsheet and add a column for the owner's name.

Score with the people who do the work. Managers underestimate exceptions. The person processing invoices knows which vendor always sends the PDF sideways.

15 workflows worth automating, by department

The catalog below lists where each workflow starts, what it touches and whether plain rules are enough. The last column is the smallest useful first step.

Eleven of the fifteen run on rules alone, and a twelfth mostly does. That surprises people who expected an AI project.

On the sales side, the tool question often turns into a CRM question. If your pipeline lives in spreadsheets, sort that first. Run the build vs buy CRM math before wiring anything into it. When no packaged product matches how you sell, a CRM that fits the process removes much of the glue work.

Rules or AI: which processes need a model

A language model reads text written for people. That is its job here. An email with four questions in it. A scanned delivery note. A ticket with the order number buried in line three. Rules cannot parse those reliably, while a model can.

Models are poor at exact arithmetic and at rules that must never bend. Do not let one infer an invoice total. Compute it, then check it against the source document.

The clean split looks like this. The model reads and classifies. A rule decides and acts. Invoice capture shows it well: the model pulls vendor, amount and due date from a PDF, and rules match the purchase order, check the math and post a draft.

Where AI gets dangerous:

• Refunds and credits. A wrong yes costs money directly.
• Price changes. One bad inference repeats across every order.
• Anything sent to a customer. Tone and facts both need review.

For these, keep a person between the trigger and the action. Route low-confidence outputs to a review queue. Log what the model saw and what it returned. Once you know the real error rate, widen the lane.

Ask two questions for each step. What does a wrong answer cost? Would anyone notice it? Cheap and visible mistakes can run unattended. Expensive or silent ones need a human check.

No-code tools or custom integration

Zapier, Make and Microsoft Power Automate cover a lot of ground. They connect common apps through ready-made connectors, and a technical ops person can build a working flow in an afternoon. For the first three processes on most lists, start there.

The limits show up on volume and on logic. Zapier counts a task each time an action completes successfully. One order passing through four action steps uses four tasks, so 5,000 orders a month means about 20,000 tasks (Zapier pricing, as of September 30, 2026). Make bills in credits, and most module actions consume one credit (Make pricing, as of September 30, 2026). Power Automate prices by user or by bot (Power Automate pricing, as of September 30, 2026).

Custom integration makes sense in three cases:

• Volume. Per-task billing grows with every order, while code on your own server costs roughly the same at 1,000 or 100,000 runs.
• Logic. Pricing rules, multi-entity accounting or approval chains that become a maze of branches in a visual builder.
• Ownership. Your logic lives in your repository, under version control, and survives a vendor price change.

The honest middle path is common. Prototype in a no-code tool and watch it for a month. Rewrite the flows that hit limits or break often.

RPA is the special case. It clicks through a screen the way a person would. Use it only when the old system has no API. A moved button breaks the bot silently, so treat it as a bridge to a real integration.

What it costs

Tool prices, as listed on each vendor's site on September 30, 2026:

Tool fees are the small line; setup time, exception handling and upkeep are the large ones. Budget for someone to own each flow after launch.

A custom integration gets an estimate after a brief. The number depends on how many systems you connect, how many exception paths exist and how clean the data is. For timing, Toimi's automation page lists 2–4 weeks for an MVP of a single key process and 1.5 to 3 months for a full department.

Make failures visible

Automation fails quietly. A sync that stops sending orders raises no alarm. The first signal is often a customer asking where their package is, days later.

Every automated step needs three things:

1. A check that it ran. Better still, a check that it did the expected amount of work. A daily job that suddenly handles zero rows should fire an alert.
2. A place for failed items. Records that fail go to a queue a person can see. They never vanish.
3. A person who gets told. One named owner. A department does not answer alerts.

Add a one-page runbook for each flow: what it does, what it touches, how to switch it off and what the team does by hand while it is off. That last line is the one people skip.

Build for retries too, because networks time out and schedulers fire twice. Check whether an action already happened before repeating it, or you will send duplicate invoices.

Put a review date in the calendar. Once a year is enough for most flows. Connected apps change fields, tokens expire and processes move on. An automation nobody remembers still makes decisions in your business.