Telekom's 43% Larger AI Target Excludes Token Costs • The Weighted Average

Finance teams should copy Deutsche Telekom’s separation of gross AI savings from the bill required to earn them. At its October 5 AI Investor Day, the company presented an approximately €1.0B target for 2027 indirect-cost savings, 43% above its earlier plan, with token costs explicitly outside the estimate.

The larger target still has an unpaid bill

The comparison starts with the correct denominator. Telekom’s 2024 finance presentation assigned approximately €0.7B of projected indirect-cost reductions to AI and automation. Its new CFO presentation puts that component at approximately €1.0B. Both concern 2027 relative to 2023, outside the United States. Divide €1.0B by €0.7B and subtract one: the increase is 42.9%, rounded to 43% because the inputs are approximate.

Telekom's AI savings target rises 43%, before token costs

2027 indirect-cost savings vs. 2023, €bn; DT ex-US. Gross, before token costs.

Deutsche Telekom CFO presentations · October 2024 and October 5, 2026

Keep capital expenditure out of that comparison. The new deck’s larger €1.1B total includes approximately €0.1B of capex savings; the earlier comparable total was €0.8B. Dividing the new total by the old indirect-cost component would manufacture a bigger improvement by changing the scope. The chart compares the same component at the same target date, rather than two conveniently adjacent headline numbers.

The CFO’s footnote matters as much as the bars: these gross figures exclude token costs. The presentation sets a token-cost objective in the low double digits as a percentage of gross savings, and leaves financial guidance unchanged while directing operating savings into further transformation. An upgraded efficiency ambition therefore does not automatically become additional cash available for distribution.

That distinction gives operators a better budget template. Maintain a ledger for avoided work, another for the cost of running the replacement, and a third for where the released capacity goes. Reinvestment can be productive without being a reduction in spending. Calling it a cash saving before the budget changes makes the next investment case harder to audit, even when the underlying workflow has improved.

Telekom’s announcement also targets approximately €2.5B of indirect-cost savings by 2030, compared with 2023. It expects AI-related business revenue outside the United States to rise from approximately €250 million in 2026 to €800 million by 2030. Those revenue ambitions belong in a separate column. Revenue earned from customers is neither an internal productivity saving nor evidence that a particular automation pays for itself.

The strongest reading is that a large operator is putting a more explicit financial framework around deployment. The weaker reading is that a forecast has been mistaken for a measured result. Procurement should distinguish them before using the numbers to justify a broader rollout. Our earlier analysis of AI billing and causation makes the same point: an impressive total still needs a defensible explanation of what caused it.

Make the workflow owner sign the savings bridge

Adoption is useful evidence, but it measures a different outcome. Telekom’s AI adoption presentation reports regular employee AI use rising from 70% in May 2025 to 77% that November and 83% in May 2026. A separate survey of 1,850 ChatGPT Enterprise users found 80% reporting substantial productivity improvements and 65% reinvesting saved time in work quality. These are different surveys, and reported productivity is not an audited cash return.

The same adoption deck assigns business teams responsibility for adoption and ROI, supported by shared tools, evaluations and a flexible model stack. That is the part smaller organizations can borrow without matching Telekom’s scale. A central platform team can provide access and telemetry; the manager who owns the workflow should establish whether the work disappeared, improved, or merely moved to another queue.

Start with a completed unit of work. For a support process, follow the issue through resolution and any subsequent reopening. For a coding process, include review and corrective work after delivery. Do not count a faster intermediate response as a saving if another team must finish it. The acceptance definition should remain stable when a model, prompt or supplier changes, or the comparison becomes an exercise in moving the finish line.

Then build the savings bridge from observed activity. Record the previous effort, the new effort, the volume actually processed and the costs attributable to the replacement. Include inference, integration, evaluation, supervision and exceptions where they occur. Have finance identify which benefits change invoices or staffing budgets and which create capacity for other work. This is a proposed operating discipline, not a claim that Telekom has published those measurements for every deployment.

The skeptical case deserves room. A workflow can deliver better service before management can remove costs, and forcing immediate budget cuts may prevent a useful system from maturing. Improved quality can justify investment on its own. The answer is to name that objective and measure it directly, rather than converting every saved minute into a hypothetical payroll reduction. A capacity gain and a cash gain can both matter without being interchangeable.

Today’s lead on advertising attribution examines a related accounting problem: the recorded outcome depends on the measurement boundary. Here, the boundary must include the resources consumed after an agent starts working and the people needed when it stops short. A cheaper model helps only if the workflow still meets its acceptance standard.

Expand deployments whose owners can demonstrate durable task completion and a credible use for the capacity released. Negotiate inference and platform commitments against observed demand, with room to change models when the economics deteriorate. Delay broad savings promises when the evidence consists mainly of adoption rates or self-reported time gains. A published reconciliation from gross benefits through operating costs to realized cash would strengthen Telekom’s case; rising exception work or token spending without corresponding outcomes would weaken it. The larger target is encouraging. The bridge from target to cash remains the operator’s job.