The question we get most often is not "whether" but "where to start". That is the right question — and most answers to it begin at the wrong end: with a tool, rather than with a process that eats time today.
This text is our order of operations. It is sorted by the ratio of effort to effect, not by what currently sounds impressive.
The ground rule: models prepare, rules decide
Before the concrete workflows, a distinction that saves a lot of trouble. There are two kinds of automation, and they suit different jobs:
Rule-based — if A, then B. Inquiry arrives, gets assigned by postcode, a task is created. Deterministic, checkable, boring. This is the part allowed to carry business-critical processes.
Model-based — a language model condenses, pre-sorts, drafts. Strong at anything that understands or produces text. Weak anywhere an answer goes out unchecked.
The combination works: the model prepares, the rule decides, a person signs off where it goes outside. Whoever flips that and lets a model decide builds a system that sends a wrong reply to a customer at three in the morning on a Tuesday.
Workflow 1: pre-sorting and assigning inquiries
The first lever is almost always the same: the stretch between "form submitted" and "someone takes care of it". In many companies every inquiry lands in a shared inbox, and from there everything depends on who looks first.
What belongs automated here: entry into the CRM with source and campaign data, assignment by traceable rules — region, service, size —, confirmation to the person who wrote in, a task with a deadline for the team, a reminder if nothing has happened after 24 hours.
A language model can pre-classify before that: is this a real inquiry, spam or a job application? Roughly what is it about? That assessment is written into the CRM as a suggestion, not a decision.
Effort: low. Effect: response time drops from days to hours, and no inquiry falls through because someone was on holiday.
Workflow 2: preparing recurring replies
A large share of communication repeats: appointment confirmations, requests for documents, answers to the same five questions. Here a model that generates a draft from the inquiry and your templates is worth it.
The key word is draft. The text lands in the outbox with a sign-off step, not directly with the customer. The time saved is not that nobody reads anymore, but that nobody starts from a blank page anymore.
Workflow 3: speeding up content production, not replacing it
The area with the most hype and the worst execution. Fully auto-generated articles are recognisable by three features: they say nothing twenty others do not also say, they contain not a single verifiable detail, and they sound like everyone else. Search engines and readers now recognise that reliably.
What is worth it: condensing research, suggesting outlines, turning a call transcript into a first draft, generating variants for headlines and ad copy, shortening long texts for different formats. The core — your own experience, the concrete number, the uncomfortable statement — comes from you. The model takes over the work around it.
Workflow 4: following up by behaviour instead of calendar
Most follow-up automations run on a schedule: day 3, day 7, day 14. That is better than nothing and worse than what is possible.
More useful are triggers by behaviour: someone opened the proposal but did not reply. Someone was on the pricing page three times. Someone booked the appointment and then cancelled. Each of those moments calls for a different message — and it can be prepared and triggered by rule.
The prerequisite is tracking that sees those signals at all. How that is set up is in Analytics and tracking.
Workflow 5: reporting that writes itself
The weekly report for which someone copies numbers from four tools into a spreadsheet is a candidate for full automation — rule-based, no model. Data is pulled via API, brought into a fixed structure, sent on a fixed date.
A model can write a summary on top: what changed, what stands out. That too is a suggestion someone reads before walking into a meeting.
What you should not automate
- First conversations. The moment someone decides whether to trust you.
- Replies to complaints. A model apologises in a way that makes it worse.
- Anything legally binding. Proposals, contracts, deadlines.
- Processes that do not work yet. Automation makes a bad process faster, not better.
Data protection is not a footnote
As soon as personal data — and an inquiry is one — passes through a language model, that is a processing activity that belongs in the record. Which provider, hosted where, under which contract. We settle that before the first prototype, not after. In most cases there is a provider with EU processing that solves the same task.
The honest order
Starting from zero: workflow 1 first, then 4, then 5. All three are mostly rule-based, easy to check and show results within weeks. Models come in where text has to be understood or drafted — as preparation, never as the last instance.
What makes sense in your case is settled fastest against your real process. Write to us with where the most time is lost today. How we go about it is on the Automation & Systems page.
