0%
of GenAI pilots show no measurable P&L impact.
MIT NANDA, 2025FactAI runs your AI portfolio through TEARI, a stage-gate from idea to impact, so every use case is prioritised, proven with evidence, EU AI Act compliant, and accountable for its € and its CO₂.
Pilots multiply, budgets scatter, and few use cases ever prove their business case or clear compliance. FactAI replaces gut-feel with a governed pipeline where evidence, not intention, decides what scales.
of GenAI pilots show no measurable P&L impact.
MIT NANDA, 2025of firms are AI high performers, with 5% or more of EBIT coming from AI.
McKinsey State of AI, 2026EU AI Act transparency duties in force: governance is no longer optional.
EU AI Act maximum fines up to 7% of turnoverFive tollgates, each with a clear owner and a pass/stop decision. A use case cannot claim impact it hasn't proven. Evidence over intention, at every step.
TEARI (Theoretical, Empirical, Applicable, Replicable, Impactful) adapts the standards-of-evidence hierarchy for evidence-based decision-making, applied at NASA, the OECD, the World Bank and the Max-Planck-Institut für Plasmaphysik. Ruggeri et al. →
A portfolio cockpit for Business, Data Science and Legal, aligned by the same evidence-driven method.
ROI, payback, Time-to-Value and a risk-adjusted Value-at-Risk across every use case, recomputed as evidence arrives.
A guided intake classifies each use case (Prohibited, High, Limited or Minimal) with the obligations and mitigations attached.
Backtests and live A/Bs are scored to an evidence level, and the probability of success is derived from that evidence, not a guess.
Allocate a fixed budget by return-on-risk, spread risk across the portfolio, and see today vs the optimised proposal.
Every use case reports net abatement, after its own AI compute, measured against your science-based (SBTi 1.5°C) target.
A grounded assistant interviews, scores and advances use cases through the funnel, reading your live portfolio and files.
Transparent formulas, an audit trail on every change, and a method that treats a confident number without evidence as a red flag.
TEARI tollgates from idea to realised impact.
EU AI Act risk tiers classified at intake.
lenses on every case: value, risk & carbon.
industry starter portfolios, ready to demo.
Interview-led business case scores Impact × Feasibility against a calibrated benchmark, with the EU AI Act risk assessed up front.
Audit the backtest, then run a controlled live experiment. Value is risk-adjusted, causally confirmed, and reported against baseline.
Move to production with control charts on value, drift and KPIs, so impact is measured, not assumed, and CO₂ is on the ledger.
A multinational pharmaceutical company receives 20 AI use cases a month for three years. The same stream of proposals goes to two boards: one with no process and no indicators, and one with risk-based tollgates, a Responsible AI committee and a delegated fast-track.
Illustrative queueing model · not client data36-month present value at an 8% discount rate, illustrative € millions. Value generated is almost identical; the gap is incident losses and spend on projects that never deliver.
Nine typical pharma use cases, from medical information to pharmacovigilance and GMP. Each bar is 100% of that type: top row Board 1, bottom row Board 2. Right column: incidents per run.
A third run puts the same Responsible AI committee in charge of every case, with no triage and no fast-track. Risk control alone is not enough: the queue congests and shadow AI comes back.
Each case has exponential patience (mean 3 months). If the wait exceeds it, the case leaves the queue (Erlang-A). Depending on the type, 10% to 70% of abandoned cases are deployed anyway without controls: shadow AI. Board 2 keeps an AI inventory that detects 60% of it and shuts it down after 2 months. Offered utilisation: Board 1 ≈ 20/14 = 1.43, stable only because cases abandon; Board 2 triage ≈ 0.67, committee ≈ 0.5.
Mean of 100 paired runs, with the 2.5–97.5% range across runs underneath. Green and red mark ranges that do not overlap Board 1.
Actual traces from run 1, one case per type. Values are the true ones, which neither board sees: sponsors report inflated value, and Board 2's business case sometimes reveals that a project will not deliver.
NPV advantage (Board 2 − Board 1) with its 2.5–97.5% range when one parameter moves and the rest stay fixed; 100 paired runs per point. The hollow point is the base case.
Assumptions, not estimates. Value and cost in € millions (medians, lognormal dispersion per case). P(incident) is the 36-month unmitigated probability. Risk tiers follow the EU AI Act, and the triage policy also treats GxP-critical uses (pharmacovigilance, GMP batch release) as high risk. Diagnostic software is high risk through the medical-device route (MDR/IVDR); emotion recognition in the workplace is prohibited (art. 5).
Every parameter here is an assumption. Four series from your own portfolio (arrival rate, how long business units wait before deploying anyway, the share of pilots that never deliver, and incident severity) are enough to calibrate the same model to your organisation.
Run it on your portfolio →Start from a curated portfolio of the flagship AI use cases for your industry: from Customer Lifetime Value to Predictive Maintenance to Grid Load Forecasting.
FactAI is built by a small team that has lived the gap between AI ambition and governed, provable impact, and got tired of watching value leak.
“The question is no longer whether we can build it; it's whether we can prove it's worth building, and govern it once we do.”
A 30-minute walkthrough on your own use cases: value, risk, EU AI Act and carbon, end to end.
Book a demo →