Building a defensible ROI case for AI investments

Why telecom ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → cases collapse

Telecom AI projects (network optimization, churn prediction, RAN energy management) share a structural problem: benefits split into two buckets with very different reliability.

Hard savings are measurable, attributable, and show up on an invoice. Example: kilowatt-hours consumed per base station, before and after an AI sleep-mode algorithm goes live.

Soft benefits are plausible but diffuse. Example: "improved customer experiencecustomer experienceThe overall perception a customer forms of your brand across every interaction, from first touch to post-purchase support.View full definition →" from AI-based network optimization, which is supposed to reduce churn (customers switching providers), but churn has dozens of causes and isolating the AI's contribution is genuinely hard.

Most vendor business cases blend both into one payback number. That's the first thing to unbundle.

Step 1: Isolate hard cost savings

For a RAN (radio access network, the base stations and antennas connecting phones to the network) energy optimization project, hard savings typically come from:

• Energy consumption reduction: AI puts underused radio carriers into sleep mode during low-traffic hours (overnight, rural cells).
• Reduced truck rolls: predictive maintenance flags failing equipment before it causes an outage, cutting emergency site visits.
• Reduced manual RF (radio frequency) tuning labor: AI-based parameter optimization replaces engineers manually adjusting antenna tilt and power.

These are auditable against utility bills, dispatch logs, and headcount allocation. If a vendor can't tie a benefit to one of these three ledgers, treat it as soft until proven otherwise.

Worked example: energy savings payback

Assume (illustrative, not a market figure):

• Network: 10,000 macro sites
• Average site power draw: 2.5 kW continuous
• AI energy optimization reduces average consumption by 8% (a range commonly cited by vendors like Ericsson and Nokia for AI-based radio sleep features is 5 to 15%, GSMA industry estimates, treat as vendor-reported unless independently audited)
• Industrial electricity price: roughly $0.12/kWh in the US, roughly €0.20/kWh in much of Europe as of early 2025 (Eurostat and EIA estimates, varies significantly by country and contract)
• Software licensing plus integration cost: $4 million upfront (illustrative)

Annual energy cost before AI (US case):

10,000 sites × 2.5 kW × 8,760 hours × $0.12 = $26.28 million/year

8% reduction = $2.10 million/year saved

Simple payback period = $4,000,000 / $2,100,000 ≈ 1.9 years

In the European case at €0.20/kWh, the same 8% saving is worth roughly €3.5 million/year, shortening payback to just over one year. This is why European operators (Vodafone, Deutsche Telekom, Orange) have been aggressive early adopters of AI energy management: their electricity cost base makes the payback math more favorable.

This is a simple payback calculation, not a discounted cash flowdiscounted cash flowDiscounted Cash Flow (DCF) is a valuation method that estimates an asset's value by projecting future cash flows and discounting them to present value using a required rate of return.View full definition →. For an MBA-level business case, note explicitly that simple payback ignores the time value of money and risk-adjusted discount rates; it's a screening tool, not a final investment decision metric.

Step 2: Stress-test the soft benefits separately

Now take churn reduction, customer experiencecustomer experienceThe overall perception a customer forms of your brand across every interaction, from first touch to post-purchase support.View full definition → improvement, or "network quality perception" claims and put them in a separate line, with separate rules:

1. Demand a control group. If a vendor claims AI-driven network optimization reduced churn by 1.5 percentage points, ask whether that was measured against a comparable region or cohort that didn't get the upgrade. Without a control group, correlation with a broader retention campaign is impossible to rule out.

2. Apply a discount factor. Many operators internally haircut soft benefit estimates by 50% or more in year one, phasing up only after two or three quarters of confirmed attributionattributionA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.View full definition →.

3. Set a kill date. If the soft benefit hasn't shown up in the data within an agreed window (commonly two to four quarters), it gets removed from the business case, not carried forward indefinitely.

This is exactly where the opening example failed: the operator booked both hard energy savings and an estimated churn benefit into a single payback figure at project approval. When the churn benefit didn't materialize on schedule, the whole business case looked broken, even though the hard savings were roughly on track.

Step 3: Separate one-time from recurring costs

A defensible model also needs cost realism, not just benefit realism. Common underestimated costs in telecom AI projects:

• Data pipeline integration: connecting AI tools to OSS/BSS (operations support systems / business support systems, the software that runs network operations and billing) is often the largest hidden cost, frequently exceeding the AI software license itself.
• Model retraining and drift monitoring: RF environments change (new sites, new spectrum bands), so models degrade and need ongoing tuning. Budget this as an operating expense, not a one-time cost.
• Change management: RF engineers who've tuned networks manually for 20 years don't automatically trust an AI recommendation. Adoption friction has a real cost in delayed rollout.

A useful gut check: if a vendor's total cost of ownership (TCO) doesn't mention integration or retraining costs, the payback period is understated.

A simple structure for the business case document

For non-technical stakeholders reviewing an AI investment, structure the case in three explicit tiers rather than one blended number:

Tier 1: Hard, auditable savings (energy, truck rolls, labor hours)
        -> Full amount included in payback calculation

Tier 2: Soft, plausible benefits (churn, NPS, brand perception)
        -> Included at 50% haircut, with a kill-date review

Tier 3: Strategic/option value (competitive parity, future 5G/6G readiness)
        -> Noted qualitatively, NOT included in payback math

This tiering is the single most useful habit for anyone evaluating AI investment decks in telecom. It forces vendors and internal champions to be explicit about what they actually know versus what they hope.

For a broader framework on evaluating AI project value beyond simple ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition →, the OECD's AI system impact assessment guidance is a useful, vendor-neutral reference point on how to think about measurable versus attributed impact.

🎬 [VIDEO: "How AI Is Optimizing Telecom Networks" - youtube.com/results?search_query=AI+network+optimization+telecom - search for recent operator or vendor case study explainers (Ericsson, Nokia, or GSMA channels) on AI-driven RAN energy and performance optimization]

Key Takeaways

• Split every AI business case into hard savings (auditable against bills, logs, headcount) and soft benefits (churn, experience, brand); never blend them into one payback number.
• Use simple payback as a screening tool, and worked numbers (energy price × consumption × reduction %) as a sanity check on vendor claims, not as a final investment metric.
• Apply a discount (commonly 50%) to soft benefits and set a kill-date: if they don't show up in the data within two to four quarters, remove them from the case.
• Budget explicitly for integration with OSS/BSS systems and ongoing model retraining; these are usually underestimated and often exceed the software license cost itself.
• European operators often see faster paybacks on AI energy optimization than US peers due to higher industrial electricity prices, a structural factor worth naming explicitly in any cross-market business case.