Goldman Sachs estimates 75% of companies overestimate the capabilities of AI being sold to them. 56% of AI POCs never reach production.
Model architecture, training dataset, independent benchmarks, real latency on your workloads, scalability, SLA. Request: sandbox access for real testing.
Where does your data go? SOC 2 Type II, ISO 27001, GDPR compliance (DPA, EU data location). On-premise available? Single-tenant vs multi-tenant?
System classification? Technical documentation available? AI register maintained? Audit trail for automated decisions?
TCO over 36 months. Pricing mechanism and vendor lock-in risk. ROI modelled on real use cases at comparable clients.
Change management planned? User training included? French/Dutch support for Belgium? Roadmap shared? Vendor financial stability? Exit strategy?
"Our AI achieves 95% accuracy", on what dataset? Compared to what? At what confidence threshold? Demand reproducible benchmarks on your own data, not marketing figures.
A vendor refusing contactable customer references or whose "clients" are all anonymised is hiding something. Minimum 3 clients in your sector, similar size, directly contactable.
Multi-year contracts without termination clauses, non-exportable data, non-transferable fine-tuned models = a trap. Negotiate: data portability within 30 days, shared fine-tuning IP, for-cause termination clause.
A vendor who will "handle compliance later" for a high-risk system exposes you directly. Deployer liability is distinct from provider liability. Demand full technical documentation before signing.
Molderez Consult SRL co-builds your 18-month AI roadmap.
AI strategy workshop