Deutsche Telekom's €2.5B AI Savings Target Excludes Token Costs

Deutsche Telekom's headline AI number, about €2.5 billion of savings by 2030, is a gross figure that excludes what the company pays for tokens. The CFO deck from its October 5 AI Investor Day footnotes every savings chart "excl. token costs" and sets a separate target that token spend should not exceed a "low double-digit percentage of gross savings." If you are building an AI savings case for your own board, DT's disclosure is one of the most detailed templates a large incumbent has published this year. Copy its function-by-function structure and its fixed baseline. Do not copy the gross headline.

The rest of the disclosure splits into two kinds of number. Some are measured: US care contacts per account more than halved, chat containment near 80%, a network agent that responds in about a minute. Others are targets built on a bottom-up model nobody outside the company can check. Telling them apart is most of the work when you read a deck like this, or write one.

What Deutsche Telekom Actually Promised Investors

Deutsche Telekom promised gross AI and automation savings of about €1.1 billion in 2027 and about €2.5 billion by 2030 outside the US, both measured against 2023. The press release says the 2027 figure is gross. The 2030 figure appears there as "savings of approximately 2.5 billion euros in indirect costs," and the release does not say gross or net. The CFO's deck does: the chart is titled "Gross IDC AL savings from AI & automation," and its footnote reads "excl. token costs, incl. cap. labour."

The deck adds the detail most coverage skipped:

• The 2027 number went up. DT's 2024 Capital Markets Day target was €0.8 billion of AI and automation impact for 2027. The updated view is about €1.1 billion, roughly €1.0 billion of indirect costs plus €0.1 billion of capex.
• The 2030 number is a percentage first. DT aims for about 15% of its ex-US indirect cost base, up from about 4% in 2026 and about 6% in 2027. The €2.5 billion is that 15% converted to euros, which implies a base of roughly €17 billion.
• Every function has its own range. For 2030: network 10-15% (automation), IT 15-20% (AI coding), sales and service 20-30% (call volume reduction), G&A and other about 10% (AI-native processes).
• The EBITDA target did not move. DT confirmed, rather than raised, its 2027 ex-US guidance of about €16 billion adjusted EBITDA AL. The deck says opex savings "are reinvested in accelerated digital transformation." The extra €0.3 billion does not reach the 2027 guidance.

The CFO slide calls the 2030 target a "medium term outlook based on bottom-up analysis." That is an honest label for a plan.

Why the Gross Framing Matters to Your Board

A gross savings target with token costs carved out tells your board what AI removes from the cost base and leaves out what running the AI adds back. DT does at least name the offset and cap it. The CFO deck lists the levers: model tiering and routing, "harnesses to minimize model lock-in," vendor management and tiered token budgets. The target is that token costs stay under a low double-digit percentage of gross savings.

Run the arithmetic yourself, because DT did not. If "low double-digit" means 10% to 20%, the 2030 token bill would be €250 million to €500 million a year, and the net figure would be €2.0 billion to €2.25 billion. The gap, up to €500 million a year, compares with the roughly €800 million of AI revenue DT is targeting for 2030.

The footnote names only token costs. The deck does not say whether seat licences, the platform engineering team or integration work are netted out. DT's release says employees have ChatGPT Enterprise, Microsoft Copilot and its own AskT assistant, and more than 100,000 have been trained. Those seats cost money every year, and nothing in the materials says which line absorbs them.

Outside readers caught the framing. SiliconAI News, citing Reuters, reported that the cost of using the AI is not counted, so the figures are not net. AI News repeated the €2.5 billion without the qualifier. Secondary coverage like that is where most board members will first meet the number.

This is the same problem we flagged when Sysco tied executive pay to an AI cost-out target without defining what it was net of.

The Measured Numbers Use Three Different Baselines

The operational results in DT's disclosure are real measurements, but they start from different years, so they cannot be added up or compared with the 2023-based savings targets. That point is easy to miss when they share a slide deck.

The T-Mobile US deck carries the strongest numbers:

• Call reduction of more than 55%, measured as annual contacts per account, from 7.0 in Q4 2021 to 3.1 in Q2 2026. The baseline is late 2021, which predates generative AI in production, so self-service apps and digital channels share the credit.
• AI chat containment of 78% in Q2 2026, up from 25% in Q2 2024.
• Net Promoter Score of 46 in Q2 2026, up from 39 in 2023, per HarrisX Mobile Insights. T-Mobile's stated ambition is 50 or more.

The group press release compresses that into "customer service calls dropped by 55 percent" and "AI agents there now handle 40 percent of customer contacts." The 40% and the deck's 78% measure different things. One is a share of all contacts. The other is containment inside the chat channel. The T-Mobile deck also gives a third figure the release does not: 32% of postpaid contacts were "served by an AI agent" in Q2 2026. The materials do not explain how that squares with the release's 40%. Containment also differs from resolution, as we covered in the AI contact center buyer's guide: a contained chat can still end with a customer who calls back tomorrow.

In Europe, the release says the Frag Magenta chatbot handled about 2.6 million customer service calls in the first half of 2026, and that early use cases in new-connection setup cut complaints by 30%. It gives no baseline for the 30%.

T-Mobile published a before-and-after pair for its metrics: a starting value, an end value, a date for each and a definition of the unit. DT's European figures mostly lack the starting value, which is the first thing an auditor would ask for.

RAN Guardian and the Network Case

DT's network numbers are the most concrete example of an agent doing a job a person used to do, and they come with no euro figure attached. The release says the RAN Guardian Agent detects impending network strain, such as crowds at large events, and cut response time "from several hours to about one minute."

The network deck sizes the pool it works against: €4.3 billion of 2025 network operations opex and €5.8 billion of network capex outside the US. It also puts a number on fiber: €100-150 million a year of savings in fiber roll-out "via digitalization and AI." The vision slide says "humans act by exception."

On the T-Mobile side, the deck cites more than 30,000 automatic antenna adjustments during Winter Storm Fern and 11,000 adaptive network optimizations at 12 FIFA World Cup fan events. Those are activity counts: they show the system ran and say nothing about what it saved.

For a reader in a utility or another network-heavy business, response time is the copyable metric here. A shorter response time is measurable from your own incident logs before and after, and it does not depend on a cost model.

---

What DT Left Out: Headcount and AI Costs

Neither the release nor the CFO deck gives a headcount target, a workforce number tied to the savings, or a total for AI spending. SiliconAI News notes the missing job figures. A 20-30% savings target in sales and service, driven by call volume reduction, is a staffing number whether or not it is labelled as one.

T-Mobile's deck says a little more. Its IT organisation moved from "a contract workforce that managed other people's software" to its own staff, measured on whether they actively code, with a 200% year-on-year increase in the share of staff building software by Q2 2026 and a 50% cut in release cycle time. The care slides describe AI "arming our experts, not replacing them."

The last company to publish a big contact-center substitution number learned how it reads a year later. In May 2025, Klarna's CEO Sebastian Siemiatkowski told Bloomberg the company would hire human agents again, after earlier claims that its AI assistant did the work of 700 employees. His explanation: "I just think it's so critical that you are clear to your customer that there will always be a human if you want." DT's choice to report NPS next to call volume is the better instinct. It puts a quality measure beside the efficiency claim.

On AI investment, the CFO deck lists projects without totals: an Industrial AI Cloud in Munich (tens of megawatts, on balance sheet, B200 capacity "sold out"), a possible bid for an EU AI Gigafactory in a consortium, and fund investments in data centers. None of it is netted against the savings.

How to Build Your Own Version

The parts of DT's disclosure worth copying are the fixed baseline year, the per-function savings ranges and the separate token cost cap. The parts to avoid are a gross headline and metrics that each start from a different year.

We saw the same pattern at DBS, where a 30% figure for agent-assisted credit memos was a goal, and at M&T, where three Copilot counts were published and none was weekly actives. The companies whose numbers hold up, like Airbnb's support cost per booking, report a unit metric with a start and end date.

This Week:

1. Pick one baseline year for every AI savings figure you report and write it on the first slide. DT uses 2023 for savings. If you have metrics on other baselines, label each one.
2. Pull your current monthly token, seat-licence and AI platform spend into one number. If it takes more than a day, tag API keys by function first; our AI FinOps tools guide covers how.

This Month:

3. Break your savings target into functions, each with an owner and a range, the way DT split network, IT, sales and service, and G&A. A single company-wide percentage hides which team is behind.
4. Set a token cost ceiling as a share of gross savings and put a number on it. "Low double-digit" works for an investor deck. Your finance team needs 12% or 15%.

Before Your Next Board Review:

5. Present gross savings, AI run costs and net savings as three separate lines. If you can only defend the gross line, say so.
6. Pair every efficiency metric with a quality metric from the same channel and period: containment with repeat-contact rate, call volume with NPS or complaints.

The Bottom Line

DT has published more function-level detail on AI economics than most large companies, and its token cost cap is a discipline other CFOs should copy. Its headline number is still a gross figure stated against a 2023 baseline, set next to operating metrics that start in 2021 and 2024, while the EBITDA guidance stayed where it was. The Q3 results on November 5 are the first chance to see whether DT reports the 2026 progress (about 4% of indirect costs, per its own deck) with the token costs alongside.

Put gross, run cost and net on three separate lines before your board sees the gross figure alone.

Continue Reading

• Sysco Tied Equity Pay to Its AI Cost-Out Push. Net of What?
• Best AI Contact Center Platforms: Containment Isn't Resolution
• DBS Put 1,500 Bankers on Agents. The 30% Is a Goal.
• M&T Published 3 Copilot Counts. None Was Weekly Actives.
• Airbnb Shipped 80% More Features. One Number Is Auditable.
• Best FinOps Tools for AI Spend: No Dashboard Fixes an Untagged Key