---
title: "AI workflow in a client project: Real numbers"
lang: en
type: blog
canonical: https://www.arocom.de/en/blog/ai-workflow-client-project
alternate: https://www.arocom.de/blog/ki-workflow-kundenprojekt-zahlen
author: "Axel Roth"
published: 2026-08-22
last_updated: 2026-08-22
---

# AI workflow in a client project: Real numbers

In one client project we switched an editorial workflow to an AI-assisted drafting process and measured the values before and after. Turnaround time per article dropped noticeably, pure writing time even more. At the same time, new work appeared: fact-checking, correcting invented sources and a fixed human approval step. On balance, a tangible relief remained, but smaller than the demo had promised. All figures stated here are anonymised orders of magnitude from this one project and explicitly illustrative, not a universal benchmark.

## An AI workflow in a client project: Numbers from practice

AI time savings get claimed often and measured rarely. "Saves 80 percent"
appears in the demo and never in the books. We wanted to know more
precisely for a concrete project, so we counted along.

The client stays anonymous, at their request and out of good manners. It
is a mid-sized company with a content-heavy web presence and a small
editorial team. Every figure in this article is an order of magnitude
from exactly this one project, anonymised and rounded. They are meant as
an illustration, not as a benchmark you can transfer one to one to your
own organisation.

## The starting point: a workflow without measuring points

Before the rebuild, article production ran the classic way. An idea was
discussed in the team, one person researched and wrote a draft, a second
edited it, then the text went through a subject-matter approval and
finally into the content management system. Nobody had ever measured how
long this took, because it never seemed necessary.

That was exactly the first step, and the most uncomfortable: for two
weeks, note the pure working time per article, split by research,
writing, editing and publishing. Without this before-measurement, we
would have ended up pitting gut feeling against gut feeling. Anyone who
wants to assess a workflow honestly cannot avoid this effort.

## What we rebuilt

The new workflow did not replace the human, it shifted their work. The
AI took over the first draft, and the human moved from first author to
editor and reviewer.

Three building blocks carried the whole thing. First, access to the
company's own knowledge, so the AI draws not from the general internet
but from the client's approved content (the principle behind this is
explained in our knowledge article on [RAG](https://www.arocom.de/en/knowledge/ai/rag)).
Second, a set of proven templates for the recurring article types, so
that structure and tone did not have to be renegotiated every time; why
the wording of the instruction matters so much is covered in the article
on [prompt engineering](https://www.arocom.de/en/knowledge/ai/prompt-engineering). Third, a
fixed, non-negotiable approval step by a human before anything was
published.

The third building block sounds obvious and is not. The temptation to
wave through a good-looking AI draft "quickly" is real. The step was
built precisely against that.

## The numbers: before, after, rework

After eight weeks of parallel operation, we had enough data points for
an honest comparison. The table below shows the orders of magnitude per
standard article. As said, these are anonymised values from this one
project and should be read as a range, not a promise.

| Phase per article | Before | After | Change |
| --- | --- | --- | --- |
| Research | 2 to 3 hours | 1 to 1.5 hours | roughly halved |
| Writing the first draft | 3 to 4 hours | 0.5 to 1 hour | strongly reduced |
| Editing and fact-checking | 1 hour | 1.5 to 2 hours | increased |
| Publishing and approval | 0.5 hour | 0.5 hour | unchanged |
| Total turnaround time | 6.5 to 8.5 hours | 3.5 to 5 hours | about a third less |

The most important value is in the third row, and it goes up. Editing
became more demanding, not easier. Anyone who leaves that out of the
calculation arrives at the dream numbers from the demos. Anyone who
counts it in lands at a real relief of about a third of the turnaround
time. That is far less spectacular and far more reliable.

## Where the AI slipped up

The new editing work had concrete reasons. Three types of error
appeared regularly, and the review step was aimed precisely at them.

First, invented evidence. The AI occasionally produced plausible-sounding
studies or figures that did not exist. Rare, but expensive every time you
miss it. Every external statement therefore had to be checked against a
real source.

Second, the wrong level of specificity with company knowledge. Despite
access to its own content, the system now and then mixed up details from
similar products. No prompt helped here, only a human who knows the
product range.

Third, a smooth, slightly generic tone. The drafts read cleanly but
interchangeably. The character by which readers recognise the sender had
to be written back in by the human. That is not a weakness of the tool,
it is a sensible division of labour.

## What we took away from it

The most important lesson is methodical: without the uncomfortable
before-measurement, we would have either over- or underestimated the
success, and in both cases made the wrong decision. The measured effect,
about a third less turnaround time at unchanged quality, clearly
justified the rebuild. The demo figure of 80 percent would have toppled
it at the first disappointment.

In terms of content, what we also see in other projects was confirmed,
and described in more detail in the article on
[AI content workflows in Drupal](https://www.arocom.de/en/blog/drupal-ai-content-workflows):
the AI is strong in drafting and weak in responsibility. It shifts work
from the front to the back, from writing to checking. Anyone who plans
for this shifted effort and obliges a human to approve gets a solid
relief. Anyone who saves it away gets fast text and a slow trust problem.

### Why do you not name the client?

Because the concrete name adds nothing to the insight, and the client wanted to stay anonymous. As a principle, we do not use client names or invented quotes. The figures stated are anonymised, rounded orders of magnitude from this one project and are meant as an example, not as a general yardstick.

### Can the numbers be transferred to our company?

Only as rough orientation. The effect depends strongly on how well your company knowledge is structured, how standardised your article types are and how strict you are about fact-checking. In an organisation with very individual, research-heavy content, the saving is smaller; with strongly recurring formats, larger. Measure your own baseline before you plan.

### Would more automation not have saved more?

On paper yes, in terms of responsibility no. Had we left out the approval step, turnaround time would have dropped further. At the same time, the invented evidence and the product mix-ups would have gone online unchecked. The human review step is exactly the insurance that makes this workflow defensible. Its effort belongs in the calculation, not outside it.

## The first step for your organisation

Before you talk about tools, measure the pure working time of a typical
content process for two weeks, split by phase. These baseline values are
the foundation for every later decision, and they cost nothing but
discipline. Only then is it worth asking which phase an AI draft really
addresses.

If you do not want to take this step alone, we will look at it together.
A [Future Check](https://www.arocom.de/en/future-check) maps out where a measured use of AI
genuinely holds up in your processes, and where the rework eats the
advantage.