An AI trust scorecard is what Hansa Iyengar said enterprises need before a system ships, with a named owner and funded quality engineering. A QualiZeal and HFS Research study of 101 U.S. enterprise quality and technology leaders finds generative AI already in production while purpose-built testing is still rare. The report recommends ownership, broader trust scorecards, and QE investment; it does not publish mandatory pre-ship criteria.
Updated November 14, 2026
Key Takeaways
- 57% run generative AI in production; only 15% have purpose-built AI testing.
- 37% run production AI with no named function accountable for quality.
- 88% measure accuracy; 47% bias and fairness; 35% regulatory compliance evidence.
- Only 24% say QE budget and headcount have fully kept pace.
Why does the assurance gap block the next release?
A November 2026 QualiZeal and HFS Research study of 101 U.S. enterprise quality and technology leaders across nine industries finds adoption ahead of assurance. 57% run generative AI in production and 36% have agentic systems live, yet only 15% have a purpose-built approach to AI testing.
52% still test probabilistic AI with frameworks designed for deterministic software. Formal AI validation frameworks are in production at only 31%. Digital Terminal’s account says those frameworks remain largely in pilot stages, and that the study points to a shift from traditional quality engineering towards trust engineering.
Hansa Iyengar, practice leader at HFS Research, set out what enterprises need in the Morningstar write-up:
Enterprises need to name an owner for AI quality, score every system against a trust scorecard before it ships, and fund quality engineering for the expanded role it now has.
She said they “can show a model is accurate but not that it is fair, explainable, or compliant.”
Who should own AI quality before an agent ships?
Kalyan Konda, co-founder, executive director and CEO of QualiZeal, said ownership of AI quality “is no longer just a testing question; it has become a boardroom question.” The study, as he described it, shows 37% of enterprises running AI in production without a named function accountable for its quality.
Morningstar’s write-up reports 29% with distributed ownership and no single owner, and 8% with no one clearly accountable. The report recommends clear AI quality ownership. That exposure sits next to the overconfidence gap CxOs cannot sign away.
What belongs on an AI trust scorecard?
The study finds accuracy is proven, but trust is not. 88% measure model accuracy, but only 47% measure bias and fairness, 37% measure explainability, and 35% can provide regulatory compliance evidence.
| What enterprises measure | Share |
|---|---|
| Model accuracy | 88% |
| Bias and fairness | 47% |
| Explainability | 37% |
| Regulatory compliance evidence | 35% |
As Iyengar put it, “Accuracy gets AI into production, but trust is what lets it scale.” She said enterprises need to score every system against a trust scorecard before it ships. She did not list the scorecard’s line items. Pradeep Govindasamy, president and CEO of QualiZeal, separately said enterprises need “continuous evaluation of fairness, explainability, safety, and regulatory evidence.” The findings do not publish a safety share. The same measurement risk shows up when a scorecard rewards the wrong outcome.
How do you gate the release?
The study does not publish pass-fail cutoffs or a software procedure. It recommends clear AI quality ownership, broader trust scorecards, and QE investment aligned with the expanded mandate. Executives say enterprises need those steps. Use that as a release checklist, not as published criteria.
- Name an owner for AI quality. Iyengar said enterprises need to name an owner; Konda said 37% run production AI without a named function accountable for quality.
- Score every system against a trust scorecard before it ships, which Iyengar said enterprises need to do. Do not treat safety as a published scorecard metric.
- Plan continuous evaluation of fairness, explainability, safety, and regulatory evidence, which Govindasamy said enterprises need.
- Test the agent as a probabilistic system rather than with frameworks designed for deterministic software.
- Require that the system can be audited and explained. Iyengar said more test cases will not fix that failure, and that closing the gap does not require another tool.
- Fund quality engineering for the expanded mandate. Only 24% say budget and headcount have fully kept pace, even though 82% of enterprises expect the QE role to grow significantly. Madhu Murty, co-founder and head of India operations at QualiZeal, said the firm sees quality engineers elevated into trust engineers, not replaced by AI and automation.
Govindasamy said AI assurance “must become a distinct discipline—built on strong QE foundations, but designed for probabilistic and increasingly autonomous systems.”
Frequently asked questions
What should an AI trust scorecard cover?
The study does not publish scorecard line items. Iyengar said enterprises need to score every system against a trust scorecard before it ships, and the report recommends broader trust scorecards.
How common is purpose-built AI testing?
Only 15% of 101 U.S. leaders surveyed have a purpose-built approach. 52% test probabilistic AI with frameworks designed for deterministic software.
Who is accountable for production AI quality?
Konda said 37% run production AI without a named function accountable for quality. Morningstar reports 29% with distributed ownership and no single owner, and 8% say no one is clearly accountable.
Will more tests or another tool close the gap?
No. Hansa Iyengar said more test cases will not fix a system that cannot be audited or explained, and that closing the gap does not require another tool.