Most agentic AI projects do not fail because of technology. They fail because the team deploying them never established who audits the agent’s decisions, what happens when it arrives, and who is accountable after launch. That is the pattern Ethan Pham, Founder and CEO of XNOR Group, examines in his latest Forbes piece on choosing an agentic AI consulting partner.
Why Agentic AI Projects Fail in Production
The numbers back this up. Gartner projects that more than 40% of agentic AI projects will be canceled before the end of 2027, not for technical reasons. S&P Global Market Intelligence found that 42% of companies abandoned most of their AI initiatives last year, up from 17% the year before. Autonomous agents layered onto fragmented data and legacy workflows do not resolve underlying inefficiencies. They execute them at a speed limit.
What Separates Partners That Compound from Partners That Deplete
Ethan Pham’s piece distinguishes consulting partners who deplete client capability and those who compound it. The firms that deplete sell hours and dependencies, leaving clients unable to operate the systems they were sold. The firms that compound engage in the strategy layer, translate between strategy and delivery, treat capability transfer as a deliverable, and bring their own operational learning into the engagement.
How to Govern Agentic AI Projects in Financial Services
For BFSI enterprises specifically, governance cannot be a post-launch concern. An autonomous agent operating on unready data does not simply expose risk; it acts on it at machine speed. IBM’s Cost of a Data Breach Report 2025 placed the global average breach cost at 4.44 million dollars. The compliance posture, audit trails, and human override mechanisms that regulated industries have always required are the same foundations that agentic systems demand today.
How to Audit AI Agent Decisions Before You Sign
Before engaging with a consulting partner, Ethan Pham recommends three questions.
Will They Challenge Your Agentic Strategy?
A partner unwilling to push back on agent scope, autonomy boundaries, or where human judgment must remain in the loop is selling build velocity, not consulting expertise.
What Capabilities Will Your Team Have After the Engagement?
Training on the system is not sufficient. The standard is a team able to audit, retrain, and confidently shut down an agent operating in production.
How Is Success Defined Beyond Launch?
Launch is a milestone, not an outcome. Revenue impact, adoption rates, agent reliability, escalation frequency, and incident response time are the actual outcomes.
Where This Governance Discipline Plays Out
XNOR Group has applied this same discipline to a PCI DSS-compliant payment platform engagement in Singapore’s fintech sector, embedding alongside the client’s technical leadership from the roadmap stage rather than arriving at implementation.
Ethan Pham’s complete argument, including the case that led him to this framework, is available on Forbes.
FAQ
Why do agentic AI projects fail?
Most fail for governance reasons, not technical ones. Autonomous agents deployed on fragmented data and legacy workflows exacerbate existing inefficiencies at machine speed rather than resolving them, and compliance controls are often treated as post-launch concerns rather than architectural requirements.
How do you audit AI agent decisions in banking?
Banking and other regulated environments need audit trails, human override mechanisms, and compliance posture built into the system from day one. The same foundations regulated industries have always required, now applied to autonomous decision-making.
What should a team be able to do after an agentic AI engagement ends?
The team should be able to audit the agent’s decisions, retrain it, and confidently shut it down if needed, not simply operate a system they do not fully understand.
