safeaiguy.com
Human card · charitarth-bharti Toronto, Canada

The Safe AI Guy

Charitarth Bharti

I design the governance, risk triage and guardrails that let organisations put AI agents into production and keep them accountable once they get there.

Get in touch AI regulatory tracker →
Now
Agentic Governance Team Lead, Responsible AI · Accenture
Also
Program Advisory Committee & guest lecturer, Certificate in AI Governance · York University. Research Fellow · Center for AI and Digital Policy.
Path
Delhi High Court→Parliament of India (LAMP Fellow)→LKY School, NUS→KPMG→EY→Accenture
Credentials
AIGP · FIP · CIPM · CIPP/C · B.A. LLB (Hons) · MPP
Recognition
IAPP Young Privacy Professional Award (2023) · Chair, IAPP Toronto KnowledgeNet (2023–25)
Current focus
Agentic AI risk, safety evaluation and frontier AI governance
§ Intended use

What I'm built for

Most of my work sits where a legal requirement has to become a control that an engineering team will actually run.

  • Agentic AI governance

    Frameworks, operating models and playbooks for agents: promotion gates from dev to prod, tool-use boundaries, observability and control-tower specs.

  • Use case risk triage

    Intake lifecycles with scoring rubrics, mitigation playbooks, failure-mode protocols and escalation paths that hold up across dozens of use cases.

  • Regulation into controls

    Turning NIST AI RMF, ISO 42001, the EU AI Act, SR 11-7 and OSFI E-21/E-23 into controls libraries teams can test against.

  • Teaching & policy research

    Curriculum and lectures on agentic risk and AI regulation, plus published research on AI, anonymisation and data protection.

Out-of-scope useSigning off on a launch nobody has assessed. I'll help you assess it quickly instead.
§ Evaluation results

Measured in production

Figures from client and firm engagements, rounded down.

ScoreBenchmarkSetting
15+Enterprise clients advised on AI governanceFS · Hi-Tech · Life Sciences · Healthcare · Retail · Insurance
50+GenAI and agentic use cases run through one standard risk-triage lifecycleAccenture
10,000+Employees enabled on Microsoft Copilot under risk-tiered policy, with full adoption in 8 weeksEY Canada
700+Legal professionals, executives and students trained in tech law and policyLawSikho
§ Current deployment

What I'm working on

Ten years across courtrooms, Parliament, policy school and the Big Four led here.

Agentic Governance Team Lead, Responsible AI

Accenture · since 2025

  • Built an agentic governance framework, operating model and playbook for a Fortune 100 asset manager building agents in Copilot Studio.
  • Designed the GenAI and agentic intake lifecycle that standardised risk triage across 50+ use cases.
  • Screens agent deployments for orchestration risk, tool-use boundaries, prompt injection and gaps in human oversight.

Program Advisory Committee & Guest Lecturer

York University · Certificate in AI Governance · since 2026

  • Shaping curriculum for an 18-week, three-course applied AI governance programme.
  • Lecturing working professionals on agentic AI risk, the EU AI Act and NIST AI RMF.
§ Monitoring

AI regulatory tracker

Refreshed daily by an automated job I run against official regulator, government and lab sources. Every item links to the original publication.

Loading the latest feed…
    Sources: European Commission, EDPB, UK AI Security Institute, UK Government, NIST, CISA, ISED Canada, OpenAI, Google DeepMind.
    § Live demo

    Ten-second use case triage

    Describe an AI use case and get a first-pass EU AI Act risk tier, plus the questions I would ask first. It is a keyword heuristic for fun and is not legal advice.

    Examples
    First-pass tier High-risk

    Matches Annex III employment use cases (recruitment and candidate screening).

    What I'd check first

      Simplified heuristic. Real classification depends on context, your role in the AI value chain and current guidance.

      § Citations

      Writing & talks

      Selected chapters, articles and podcasts. Linked titles open the source.

      § Known limitations

      Failure modes, disclosed

      Every good model card lists them.

      • Will ask who owns the risk before asking which model you're using.
      • Answers "it depends" and then tells you exactly what it depends on.
      • Thinks a good human-in-the-loop design has a human who can actually say no.
      § Escalation path

      Shipping agents? Let's talk.

      Governance for an agentic rollout, a speaker on AI risk, or a collaborator for policy research. Email is the fastest route.

      LinkedIn ↗