Short answer. Ask for one number: the share of the vendor’s AI projects that reach production. Ableneo runs at 4 of 5, roughly 80%. Then verify it. Call two named production references under the same regulator, ask to see the audit trail of a live system, and write the production claim into the contract with exit terms. A vendor that answers with a slide deck instead of a production rate has already told you the answer.
Most AI vendor pitches lead with capabilities. Verification works the other way: you start from what already runs in production and trace it back to the people who signed it off. The claim you are testing is not “can this vendor build a model” but “can this vendor get a model past the last mile into daily use inside a regulated bank.” Those are different skills, and the second one is rare.
A verifiable track record has numbers attached. In 2025 Ableneo shipped 34 production AI projects for 20 clients, roughly 4 of 5 of its projects reach production, 94% use large language models, and the average client runs 1.7 projects over the relationship. One named, client-approved example: at T-Mobile Czech Republic, document validation went from 7 hours to 30 minutes, 14 times faster, at over 90% accuracy, after a previous supplier had not delivered. Each of those is a number you can repeat back to the vendor and ask them to prove.
The verification itself is a short list of checks, not a feeling:
In financial services, verifying a vendor is not optional diligence, it is a legal duty. The Digital Operational Resilience Act (DORA), in force for financial entities since January 2025, requires a bank to assess an ICT third-party provider before it signs, when the service supports a critical or important function. That assessment covers the provider’s ability to deliver reliably and securely, and a separate preliminary check of concentration risk if you already depend on the same provider elsewhere. The responsibility stays with the bank. You cannot outsource it to the vendor’s marketing.
The EU AI Act adds a second layer. Credit scoring and insurance risk assessment and pricing are high-risk uses under Annex III, with obligations that apply from 2 December 2027, and the Act’s value-chain rules mean a bank that puts its own name on a system, or changes how it is used, can take on the provider’s obligations itself. That makes the vendor’s production discipline your exposure. If their monitoring, logging, and rollback are weak, the regulator asks you about it, not them.
Lead verification with one number: the share of the vendor’s AI projects that reach production. Ableneo’s is 4 of 5, roughly 80%.
Four numbers separate a track record from a brochure. First, the production rate: of every ten AI projects the vendor started, how many are live today. Second, the count of production deployments that look like yours, same industry, same regulator, same scale, because a model running at a small fintech does not prove capacity at a national bank. Third, time from pilot to production, which exposes whether the vendor actually crosses the last mile or stalls at the demo. Fourth, repeat business, the share of clients who bought a second project, which is the market’s own verdict on whether the first one worked.
A vendor that can answer all four with specifics, and name the base each number is measured on, is describing reality. A vendor that answers with a feature list, an award, or a partner logo wall is describing ambition. Both are useful, but only the first is a track record.
References are the vendor’s best-case deployments, so treat them as a starting point, not proof. Ask for two, both under oversight comparable to yours, and call them. On the call, skip the testimonial and ask what broke, who fixed it, how long the pilot took to reach production, and whether the system is still running today. Ask who owns it now, the vendor or the client’s own team, because a clean handover is itself a production signal.
The most useful question is the one a slide cannot answer: “What would you do differently if you ran this project again?” A reference who answers in concrete, slightly unflattering detail has lived through a real production deployment. A reference who stays glossy either had an unusually smooth project or was not close to the work.
DORA sets out the duty across three articles. Article 28 lists the general principles for managing ICT third-party risk, including that the bank assesses the arrangement before entering it. Article 29 requires a preliminary assessment of ICT concentration risk, whether the provider is easily substitutable and whether you already rely on the same or a closely connected provider. Article 30 fixes the minimum terms the contract must carry, including the bank’s right to access, inspect, and audit the provider.
In practice that means your verification work has a regulatory home. The production references, the audit-trail review, and the exit plan are not extras you add to be thorough, they are evidence you can show a supervisor that you did the assessment the law requires. A vendor familiar with regulated clients will expect these requests and have the documentation ready. Hesitation here is itself a finding.
A pilot and a production system read almost identically on a slide, so test for the things a pilot does not have. Production has a go-live date and a number of months running. It has real users rather than a test cohort. It has monitoring, an owner on call, an incident history, and a rollback plan for the day the model behaves badly. A pilot has a demo environment and a promising result.
Ask for the go-live date and the current status in one question. If the case study is a production system, the vendor gives you a date and a monthly volume. If it is a pilot, the answer turns vague, shifts to future tense, or moves to a different example. The OECD AI Principles point the same way: a system you can trust is one whose behavior is traceable and whose actors are accountable for how it performs across its whole lifecycle, which is exactly what a pilot has not yet proven.
The last step of verification is to make the vendor’s numbers enforceable. If the vendor told you 80% of projects reach production and quoted a target accuracy, those belong in the contract as service levels, not in the pitch as adjectives. Add the right to audit the live system and its logs, data access on your terms, and an exit clause that returns your data and models if the vendor fails. The EU AI Act’s value-chain rules make this concrete: the contract is where you allocate who holds which obligation when something goes wrong.
A contract that reflects what the vendor proved in diligence, and no more, is the cheapest insurance in the project. It also filters vendors by itself. A vendor confident in its production record signs terms that reference it. A vendor selling ambition asks to keep the language soft.
Ableneo publishes the number most vendors leave out: roughly 4 of 5 projects reach production, measured across 34 production AI projects delivered in 2025 for 20 clients in regulated industries across four countries. That rate is only credible because it is verifiable, named references, dated go-lives, and audit trails a bank’s own risk team can inspect. Ableneo’s work in financial services and telecom is built for exactly the due-diligence a regulated buyer has to run, from the first reference call to the DORA assessment file. You can start from the Ableneo AI Transformation FAQ and ask for the production rate before anything else.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.