GTP AI efficiency score: how we measure the effect of AI per workflow

Why

Why we measure before we automate

In AI projects the technology almost always works. The demo runs, employees save time, and still little reaches the P&L. The reason is rarely the tool. The baseline is missing: nobody recorded before the project how many hours a case costs, how often it is reworked and how long it waits. Without a baseline there is no proof, only a usage statistic.

That is why one order applies in every company we acquire: measure first, then automate, then measure again with the same definition. The numbers belong to the company and its team. They decide where AI helps first and whether it has helped.

Metrics

Three metrics per workflow

The score consists of three metrics collected separately for each workflow. A workflow is, for example, processing an incoming invoice, producing an inspection report or answering an owner's enquiry.

In addition we record per workflow what share of cases runs through without human intervention and what share lands with a person as an exception. That ratio shows whether the process learns or whether the exceptions stay with the team.

Step 1

Step 1: define workflows and tasks

First we define which workflows are measured and what counts as a case. We start where most of the hours are: recurring clerical work in administration, billing, documentation and customer communication.

Each affected role is broken down into its individual tasks using the Task Architect method. For each task it is fixed whether a person does it today and whether a machine should take it over after the rebuild. Only then is it clear which minutes per week can go away at all.

Step 2

Step 2: baseline before the first tool

The baseline is collected before the first AI tool is introduced. It covers a defined period that includes at least one full cycle of the workflow, for example a billing month or an inspection period.

We measure real cases, not the clean examples from the demo. That explicitly includes the cases where someone had to ask. We record the three metrics, the exception ratio, the measurement period and the person who measured. The owner, who usually stays 6 to 24 months as managing director, knows the exceptions and is the most important source in this phase.

Step 3

Step 3: measure afterwards, same definition

After the process has been rebuilt, we measure again with the same definition, the same notion of a case and a comparable period. The three metrics then stand side by side: before and after, per workflow.

We also answer the fourth of the four questions for every AI project: what happened with the time gained? More cases with the same team, lower costs, or better work the customer pays for. For that, one line in the plan states before the project which P&L item is meant to change, and after six months exactly that line is checked.

Scope

What the score is and what it is not

• A comparison per workflow, before against after, with a disclosed definition and a disclosed measurement period.
• Not a usage statistic. How often a tool is opened says nothing about whether work goes away.
• Not a company valuation. The score does not feed into the purchase price. The purchase price is based on today's earnings.
• An estimate before closing, a measurement after closing. Before closing we use the same task logic to estimate what share of the work runs at a desk, recurs and is poorly supported today. That is an estimate, not a measured value.
• A method, no results yet. We publish figures only once they come from an acquired company, with baseline and measurement period.

Limits

What we do not promise

Investor material on AI roll-ups names targets: traditional service businesses at 8 to 15 percent EBITDA margins (ProxDeal, updated 26 Jun 2026), target margins of 30 to 40 percent after the rebuild (Capital Founders, updated 11 Oct 2026). These are third-party targets, not measured results and not our plan.

We promise no margin, no percentage and no deadline. We promise a baseline before the first tool, the same measurement afterwards and disclosure of the definition. What the platforms' targets claim and what we underwrite is on AI-enabled roll-up.

Publication

Publication

With this page the method is published. Results per workflow will follow on this page once they come from an acquired company: with baseline, measurement period, definition and a note on who measured. Changes to the method are recorded here with a date.

• 24 October 2026: first version of the method with three metrics, exception ratio and three steps.

Sources

Sources

• Daniel Szabo: Task Architect
• Daniel Szabo, 6 October 2026: Four questions for every AI project
• Daniel Szabo, 6 October 2026: Ask-Sabine processes
• ProxDeal, updated 26 June 2026: AI roll-ups in DACH
• Capital Founders, updated 11 October 2026: What AI roll-ups are

Questions and answers

Frequently asked questions

What is the GTP AI efficiency score?

Generation Tech Partners' method for measuring the effect of AI in an acquired company per workflow: hours per case, error rate and turnaround time, collected before the first tool and afterwards with the same definition.

Which metrics are measured?

Three per workflow: hours per case, error rate and turnaround time. In addition the exception ratio, the share of cases that land with a person.

Why measure before the first tool?

Without a baseline there is no proof. Whoever measures only after the rollout cannot separate the effect from the starting point and ends up measuring only tool usage.

Are there results yet?

No. This page describes the method. We publish results once they come from an acquired company, with baseline and measurement period.

Does the score feed into the purchase price?

No. The purchase price is based on the company's current earnings. What AI delivers after closing is the buyer's risk and the buyer's work.

Who measures?

The team in the company together with the group level for AI. The owner, who usually stays 6 to 24 months as managing director, knows the exceptions and helps provide the baseline.