AutomationProfessional services

Automating Repetitive Business Processes

The mapping step nobody wants to do, and why skipping it is why most automation projects disappoint.

This is an implementation example, not a client case study. It describes how this work is actually carried out — the method, the sequence and the reasoning. No client is named and no result is claimed, because inventing either would make it worthless as evidence.

Business process mapped before automation is built

Automation projects fail from bad process design far more often than from bad tooling. The software is the easy part. Working out what should actually happen, in what order, with which exceptions, is the part that decides whether the result helps.

Project type
Business process automation with AI assistance
Sector
Professional services

The challenge

Repetitive work accumulates informally. It lives in someone's head, varies by who is doing it, and has undocumented exceptions. Automating it as-is encodes all of that, permanently and at speed.

The objective

Remove genuinely repetitive work while keeping human judgement where judgement is actually required — and end up with a process that is clearer than the one you started with.

Approach

Map first, simplify second, automate third. Often the mapping alone removes steps that existed because nobody had ever written the process down, and that is worth more than the automation.

Implementation

1. Map what actually happens

Not what the process is supposed to be. What people actually do, including the workarounds.

For each step: who does it, what triggers it, what information they need, what they decide, what happens next, and what the exceptions are.

This is the unglamorous part and it is where the value is. Processes regularly turn out to contain steps that exist only because someone once did it that way.

2. Decide what should not be automated

Explicitly. Good candidates to keep human:

  • anything requiring judgement about a person
  • anything where being wrong is expensive and hard to reverse
  • anything that is the actual relationship with the customer

Automating the wrong step damages the thing the business runs on.

3. Decide where AI helps and where it does not

Good fitPoor fit
Summarising a long enquiry into a briefMaking the final decision on it
Drafting a first reply for reviewSending unreviewed replies
Extracting structured data from messy textAnything requiring guaranteed accuracy
Classifying and routingAnything where a confident wrong answer is costly

The pattern: AI is good at reducing effort on a task a human still checks. It is poor at being the last step before something irreversible.

4. Build incrementally

One step at a time, with the previous one running in production first. A workflow built end to end before any of it is proven is a workflow that fails end to end.

5. Keep a human in the loop where it matters

Draft, do not send. Suggest, do not decide. Flag, do not delete.

6. Make failures visible

The most important design decision in the whole build. An automation that fails silently is worse than no automation, because people stop checking the thing it replaced.

Every failure path needs to surface somewhere a person will actually look.

Technology

  • n8n
  • Make
  • OpenAI / Anthropic APIs
  • Webhooks
  • CRM integrations

Decisions worth explaining

Process mapping before any tooling decision

Automating an unclear process encodes the confusion and makes it faster. The mapping frequently removes steps on its own.

AI drafts, humans send

Keeps the effort saving while keeping the failure mode recoverable. An unreviewed wrong reply to a customer costs more than the time it saved.

Every failure path surfaces to a person

Silent failure is the worst outcome: the manual process has been abandoned and the automated one is not running.

What it produces

A documented process, with the genuinely repetitive parts automated, judgement kept where judgement belongs, and failures visible. No hours-saved figure is claimed — a saving is only a fact once it has been measured against a baseline that was actually recorded beforehand.

What it teaches

  • The mapping step is the valuable one. The tooling is the easy part.
  • Automation amplifies whatever process it encodes, including a bad one.
  • Silent failure is the single most damaging design flaw in an automated workflow.
  • 'Draft, do not send' preserves most of the benefit and nearly all of the safety.

Related services

Related guides

Worked examples

All case studies