Enterprise AI ROI: how to measure return on investment in 2026

In 2026, no executive committee accepts funding an AI project without a quantified business case. The days when the simple promise "it'll save us time" was enough are gone — and that's healthy. Still, between initial enthusiasm and rigorous measurement, many teams stumble on the same obstacles: no baseline, underestimated costs, gains double-counted.

This article describes the method we apply on DevHighWay projects to calculate a defensible ROI. The goal isn't to inflate numbers but to produce an analysis that a CFO, a CIO and an operational lead all sign with the same pen. The result: fewer pointless debates, clear trade-offs, and projects that deliver on their promises.

AI ROI isn't a myth — it's a discipline

You often hear that AI is "intangible" and that its ROI can't be measured. That's false. What's true is that most projects don't give themselves the means to measure it: no baseline, no instrumentation, no post-deployment review. AI ROI follows the same laws as any industrial investment — only the variables (tokens consumed, adoption rate, response quality) are new.

Concretely, a well-scoped AI project breaks down into three categories of gain: direct gains (hours saved, additional conversions, automated support), indirect gains (time to market, perceived quality, customer experience), and defensive gains (market share preserved against competitors investing in AI). The first two are quantified in euros; the third is measured in risk avoided.

Step 1 — Establish the baseline before anything else

The baseline is the precise snapshot of your activity before the AI project. Without it, you can prove no gain. The KPIs to freeze depend on the use case: for customer support, it'll be average handling time, cost per ticket, first-contact resolution rate. For a sales funnel, lead qualification rate, cost per opportunity, average response time.

The classic mistake: grabbing numbers from the last steering committee. Not enough. You need at least 90 days of consolidated history, ideally 12 months to neutralize seasonality. And above all, the baseline must be validated by the business team — not by IT alone. Without that joint commitment, the numbers will be challenged the moment a gain is contested.

• Quantitative KPIs: volume processed, unit time, unit cost — the raw material of the calculation
• Qualitative KPIs: CSAT, NPS, escalation rate — to capture the experience
• Reference period: 90 to 365 days, dated, signed, archived
• Source of truth: a single tool per KPI (CRM, helpdesk, analytics) to avoid double-counting

Step 2 — Quantify the eligible scope

Not every process is a candidate for AI. The right reflex: cross volume and unit time to identify high-potential tasks. 500 tickets a month at 20 minutes each is 167 monthly hours you can automate — a credible scope. 50 ultra-specific 3-minute requests is not an AI project, it's a static FAQ.

For each eligible task, estimate the realistic automation rate. A well-built RAG chatbot absorbs 40 to 70% of tier-1 questions, rarely more. An AI agent on structured tasks can reach 80-90%. But a case showing 95% automation often hides an overly simplified scope.

Step 3 — Estimate full costs

The most common mistake: only counting implementation. In reality, an AI project costs across four buckets. Implementation (€15-60k depending on complexity), monthly LLM consumption (€50-2,000 per month depending on volume and model — Claude 3.7 Sonnet, GPT-4 Turbo and Mistral Large 2 don't price the same), technical maintenance (RAG to reindex, prompts to tune, models to upgrade) and change management.

This last bucket is systematically understated. Training 30 internal users on a new tool, documenting new processes, supporting the first weeks of adoption, managing resistance: count on 20 to 30% of the implementation cost, sometimes more. A technically perfect but poorly adopted project generates no ROI.

• Implementation CAPEX: design, dev, integrations, knowledge base, testing
• LLM OPEX: input/output tokens × monthly volume × model price (check the tiers)
• Maintenance: 15-25% of CAPEX per year, more for self-hosted models (vLLM, GPUs)
• Change management: training, documentation, adoption support — 20-30% of CAPEX

Step 4 — Quantify gains in three scenarios

A single gain figure is suspicious. Always produce three scenarios — pessimistic, realistic, optimistic — based on explicit assumptions for automation rate, adoption and quality. This triangulation forces clarity: a project that only pays off in the optimistic case signals a fragile business case.

A concrete example for a support chatbot: pessimistic 30% of tickets absorbed, realistic 50%, optimistic 70%. If even the pessimistic scenario crosses the profitability threshold, the project is solid. Otherwise, either rework the scope or accept that ROI will be marginal and decide accordingly.

Step 5 — Calculate payback and 12/24-month ROI

Two indicators are enough. Payback period (total cumulative cost / average monthly gain) answers "how many months until I recoup?". 24-month ROI ((gains - costs) / costs) answers "how much do I gain in total?". A healthy AI project shows payback between 6 and 18 months and 24-month ROI above 100%.

Beyond a 24-month payback, the project is exposed: models evolve, business needs shift, and obsolescence risk becomes real. That's one reason we recommend modular architectures, capable of absorbing an LLM or framework change without rebuilding everything.

Step 6 — Measure post-deployment and adjust

ROI is not a number frozen at kickoff. It's a living metric. At D+30, D+90 and D+180, take the baseline KPIs and compare. Document the gaps, identify the levers (insufficient adoption? response quality? scope too broad?) and adjust. This improvement loop is what separates an abandoned POC from a product that scales.

Useful tools for ongoing measurement: LangSmith or OpenAI Evals for response quality, a post-conversation CSAT dashboard, product analytics (Mixpanel, Amplitude) for adoption. Without these instruments, the ROI presented in committee remains declarative.

Which measurement horizon should you pick for ROI?

The choice of time horizon changes everything. Over 12 months, you prove immediate viability and unlock next year's budget; it's the minimum duration to make the project defensible in committee. Over 24 months, you capture compounding effects — broader adoption, lower LLM bills thanks to optimizations, scope expansion — that are rarely visible in year 1. It's the standard window we recommend for decision-making.

Beyond 36 months, caution. The pace of model evolution (major update every 6 to 12 months) and the continuous drop in token prices (-30 to -50% per year in 2024-2026) make any projection beyond 3 years unreliable. For projects requiring heavy hardware investment — self-hosting, dedicated GPUs — a longer horizon can be justified, but only if you build in a hardware reinvestment scenario at 36 months and revalidate the quality assumption every year.

The three traps that wreck an AI ROI calculation

In our audits, three mistakes come up systematically and invalidate entire business cases. Avoiding them from the scoping phase saves months of internal debate.

• Counting hypothetical gains without a baseline: "we estimate saving 30% of time" without initial measurement won't hold up in a post-project review
• Forgetting change management costs: the tech works, but 40% of users don't use it — real ROI cut in half
• Measuring on a non-representative sample: testing on the 10 simplest cases and extrapolating to the entire scope guarantees disappointment

What's next?

Calculating ROI for an AI project isn't a theoretical exercise: it's what makes the difference between an investment defended in the executive committee and a POC forgotten in six months. The method fits in six steps, but the rigor of initial scoping accounts for 80% of the final outcome.

• Start with a free audit to identify the highest-ROI AI scopes in your business
• Explore our support plans starting at €199/month to structure measurement
• Get in touch for 30 minutes of free scoping on your AI project

An AI project without measured ROI is a project that will die at the next budget review. Measure early, measure right, adjust often — that's the only way to turn initial enthusiasm into lasting value.