Certification is the floor. Production evidence is the proof.

Enterprise AI is entering its credential-building phase. Model providers and services firms are creating academies, certifications, centers of excellence, partner tiers, and large-scale training programs. This is necessary infrastructure. It is also easy to misread.

A certification can show that someone met the requirements of a defined credential at a point in time. It does not, by itself, show that the person diagnosed an ambiguous operating problem, built a reliable system, navigated a customer boundary, won adoption, or delivered an outcome in production.

The distinction matters because public announcements now describe populations at very different stages with numbers that look comparable when removed from context.

Access, training, certification, and proof are different states

In October 2025, Anthropic and Deloitte announced that Claude would be available to more than 470,000 Deloitte people and that the organizations would co-create a program to train and certify 15,000 professionals. In November, Cognizant announced that it would provide Claude to up to 350,000 associates. In December, Accenture and Anthropic announced that approximately 30,000 Accenture professionals would receive training, including forward deployed engineers.

Each statement describes a different commitment. “Available to” is not “trained.” “Trained” is not “certified.” “Certified” is not “deployed.” “Deployed” does not automatically mean an attributable outcome was measured.

The market's most useful precedent comes from Anthropic itself. Its June 2026 Claude Partner Network Services Track does not rank firms on certification alone. The entry Select tier combines at least 10 active certified individuals with at least two joint customers deployed in production during the prior 12 months and at least one public customer story. Higher tiers require larger certified benches, more production deployments, and more customer stories. Anthropic also counts recent Claude use and reviews standing on a published schedule.

That architecture makes an important admission explicit: knowledge credentials, current practice, production delivery, and external corroboration are separate signals.

The evidence ladder

A useful hiring or qualification system should preserve those distinctions:

1. Access: the person could use the tool.
2. Training: the person completed a defined learning experience.
3. Certification: the person passed a defined examination or credential requirement.
4. Deployment: the person contributed to taking a system live.
5. Proven outcome: the deployment produced a measured operating result with a known baseline and observation window.
6. Referenced contribution: an accountable third party can verify what the person actually did.

These states can reinforce one another, but they should never collapse into one badge. A recent certification may support current model knowledge. A production artifact may support engineering capability. A customer reference may support contribution and context. An outcome may support impact. None is a perfect substitute for the rest.

This is especially important for team-based work. An enterprise deployment may have an excellent outcome while an individual candidate played a narrow role. Conversely, a strong engineer may have done difficult, high-quality work inside a deployment whose business outcome was delayed by procurement, adoption, or organizational constraints. A credible evaluation must examine contribution without pretending that all outcomes belong to one person.

Questions for hiring and evaluation

When a candidate presents a credential, ask:

1. What did the credential assess? Was it knowledge recall, architecture judgment, hands-on implementation, or observed work?
2. When was it earned, and what has changed since? Model behavior, APIs, tools, and safety practices move quickly.
3. Where did you use the capability afterward? Ask for the environment, task, constraints, and the candidate's authority.
4. What artifact can be inspected? Prefer code, evaluations, decision records, observability evidence, launch criteria, or a sanitized operating document.
5. What reached production? Distinguish a course exercise, prototype, pilot, controlled release, and sustained operation.
6. What result was measured? Ask for the denominator, baseline, time window, missing data, and alternative explanations.
7. Who can attest to your role? The reference should verify contribution, not merely employment or project membership.

For an assessment program, add a recency question: is the evidence current enough for the decision being made? A credential should expire or lose weight when its content no longer reflects the operating environment. Work evidence should also carry dates, versions, and provenance.

LockedIn Benchmark implication

The LockedIn Benchmark should never convert a vendor badge directly into an FDE score. Certification belongs in an evidence portfolio as a bounded knowledge signal. It may reduce redundant testing where content alignment is documented, but it cannot waive observed work, production reasoning, or corroboration.

The benchmark should report the evidence ladder visibly. A candidate might be current-certified but not yet production-proven; production-experienced but missing a verifiable outcome; or strongly referenced with limited inspectable artifacts because of confidentiality. Those are different profiles, not one blended certainty number.

Where evidence is absent, the result should say “not observed” rather than infer incompetence. Where evidence conflicts, the conflict should be adjudicated and retained. The objective is not to make qualification punitive. It is to make claims legible.

Limitations

The cited announcements are first-party descriptions of planned and active programs. Large access or training commitments are not completion counts unless a source explicitly says so. Firm-level partner requirements do not validate an individual-level FDE assessment. This review provides no independent evidence that any particular certification predicts job performance, and LockedIn Labs has not completed criterion-validity, reliability, fairness, or transportability studies. The evidence ladder is a proposed measurement discipline, not a published finding about candidate quality.