An AI Savings Target Is A Promise, Not A Result

Deutsche Telekom has put a number on what AI and automation will save it: about €2.5 billion in indirect costs by 2030.

On the same day, a deputy governor of the Bank of Japan said AI has so far shown up first as demand, with the profits still to come. A correction is possible, he added, if they do not follow.

One company publishes an AI savings target. One central banker warns what could happen across a whole economy if the expected profits do not arrive. Read together, they frame a question an AI lead hears from the board sooner or later. What is our AI savings target, and how would we know if we are missing it?

What Deutsche Telekom actually committed to

In its release for the AI Investor Day in Bonn, published on 5 October, Deutsche Telekom set out three figures. First, it expects AI and automation to cut indirect costs by about €2.5 billion by 2030, measured against 2023. Second, it expects gross savings of about €1.1 billion outside the United States in 2027, also against 2023. Third, it aims to grow AI-related revenue from business customers outside the United States. The company expects about €250 million for 2026 and about €800 million by 2030.

The release also says where part of the money goes. Additional savings in 2027 are to be partly reinvested in digital transformation and in fibre expansion in Germany. In other words, the AI savings target is partly a funding plan for other work.

The parts that make the AI savings target credible

Several details turn this from a slogan into something a finance director can check. There is a base year, 2023, so the saving has a fixed starting point. There is a scope, indirect costs, so the saving is tied to defined cost lines rather than the whole income statement. And the 2027 figure is labelled gross and limited to operations outside the United States.

The company also shows its working in operations. Its chatbot, Frag Magenta, handled about 2.6 million customer service calls in the first half of 2026. In its US business, customer service calls have fallen by 55 percent. AI agents there now handle 40 percent of customer contacts. Meanwhile, response time to network strain from large events dropped from several hours to about one minute. Those are operational measures a board can follow quarter by quarter, long before 2030 arrives.

The parts that are still a forecast

The same release is careful about what it does not promise. For 2030, it describes savings from “AI and automation” without separating the two. So the AI share of the target is not stated. Unlike the 2027 figure, the 2030 figure also carries no gross or net label.

Then there is the small print. Deutsche Telekom states that it can offer no assurance its targets will be met. It also names the progress of staff-related restructuring as one of the risks. That line is worth reading, because it tells you which assumption the whole AI savings target rests on.

Why the Bank of Japan reads the same story differently

Deputy Governor Shinichi Uchida was not talking about Deutsche Telekom. In his opening remarks at the ECONDAT 2026 Fall Meeting on AI, big data and monetary policy, he described what AI is doing to the economy as a whole. His tentative reading is that the demand side has come first. AI has boosted stock prices and, on balance, eased financial conditions. However, he sees a risk of correction if profits do not follow.

He was equally frank about the measurement problem. AI may raise productivity and capital accumulation, he said, but central bankers cannot yet say “to what extent and degree? In what time horizons?” He also suggested that conventional statistics may not keep up with the speed of AI adoption.

That macro picture has a direct equivalent inside any organisation with an AI savings target. The spend on licences, integration and training lands in this year’s budget. The saving lands later, if it lands at all. That gap is where a board’s patience runs out. It is also where an AI lead without a defensible number loses the argument.

Building an AI savings target your board can test

You do not need Deutsche Telekom’s scale to borrow its structure. You need the same few elements, written down before the work starts rather than reconstructed after it. A usable AI savings target answers five questions:

• Baseline. Which year, and what was the actual cost in that year for the lines you expect AI to touch?
• Scope. Which cost lines, which business units and which countries are in, and which are deliberately out?
• Label. Is the figure gross, or net of what the AI itself costs in licences, inference, integration and training?
• Attribution. What counts as an AI saving, as opposed to ordinary automation, a restructuring or a supplier renegotiation that happened in the same year?
• Destination. Does the saving return to the bottom line, or is it earmarked for reinvestment, and where?

The attribution question deserves the longest look, because it is where the arguments start. Deutsche Telekom does not split AI from automation in its 2030 figure. For a company of that size, that may be defensible. For a mid-sized organisation presenting its first AI savings target, it invites the obvious challenge: would this saving have happened anyway?

Who owns the AI savings target

A target without an owner is a forecast with nobody to answer for it. Deutsche Telekom describes its intended way of working plainly: employees set goals, review results and keep responsibility for decisions. That principle applies to the target itself. Someone named reviews the figures each quarter, and someone named decides when the target is revised.

Ownership also has a cost line attached. Deutsche Telekom says it has trained more than 100,000 employees in using AI. Now consider a target that assumes staff will adopt the tools but carries no training budget. It has quietly moved part of its cost out of view.

What to take into the next budget round

Do not copy Deutsche Telekom’s number. Its base, scope and size bear no relation to yours. A borrowed figure is also the first thing a sceptical board member will test. Copy the shape instead: a base year, a defined scope, a gross or net label, an attribution rule and a named owner.

Then take the Bank of Japan’s caution seriously at your own scale. The spending comes first and the profit is expected later. The AI savings target that survives is the one that shows progress before the deadline, in measures the board already trusts.

So, before the next budget round: if your board asked for your AI savings target tomorrow, which of those five questions could you answer with a figure?