An Integrated QA Engagement for a Complex, High-Risk Data Platform

About project

Client

A mid-market SaaS vendor with 200+ enterprise customers across the EU and UK, building an enterprise database management & observability platform — the tooling DBAs and platform teams use to monitor, query, optimize, and govern production databases across an organization.

• Deal stage: post-Series B, scaling into regulated enterprise accounts
• Deployment: multi-tenant SaaS, sitting directly against customers’ production data
• Markets: finance, healthcare, and enterprise SaaS — all with auditors attached

To stay competitive, they’ve shipped an AI assistant on top of the platform:

• AI index/optimization advisor — the assistant recommends schema and index changes.
• Natural-language-to-SQL — users describe a query in plain English; the assistant generates and runs the SQL.
• AI anomaly detection — the platform flags abnormal database behaviour (slow queries, unusual access patterns, capacity risks).

Project overview

Let’s make your product release and audit-ready.

Challenge

The AI shipped fast, and now the team is exposed on three fronts at once:

• They can’t prove the assistant is correct. It sometimes generates SQL that runs but returns incorrect data, or references columns that don’t exist. There’s no baseline to measure against and no way to catch quality slipping after a model update.
• They can’t prove it’s safe. The assistant is a brand-new attack surface — a crafted prompt might coax it into reaching data outside a user’s permissions or across tenant boundaries. Their existing security testing was built for the classic app, not for an LLM capable of performing database actions.
• They can’t prove it’s compliant. Their enterprise buyers’ auditors are now asking how AI decisions are logged, traced, and governed — and with EU AI Act enforcement landing August 2026, “we’ll figure it out later” is no longer an answer.

Why a tool alone doesn’t close this. They’ve already bought scanners and monitoring. The gap isn’t more tooling:

• A scanner finds known CVEs. It doesn’t decide what to test on a non-deterministic AI, build a golden dataset, or recognize a prompt-injection path specific to their schema and business logic.
• A monitoring platform shows a dashboard. It doesn’t give them a release verdict — a defensible ship / don’t-ship call — or remediate what it finds.
• No tool maps findings to the EU AI Act, SOC 2, or GDPR, attaches governance context, or produces an audit-ready evidence pack an external auditor will accept.
• The three problems interact. A tool tests each in isolation and misses exactly the failures that live in between — where an AI correctness bug becomes a data-leakage incident becomes a compliance gap.

They don’t need another tool to run themselves. They need someone to own the testing, do the work, and hand back proof — across all three surfaces, in one engagement.

Why One Combined Engagement

Solutions

Workstream 1 — AI Validation

Testing the assistant as a system that takes non-deterministic action against real data.

What we tested

• NL→SQL semantic correctness — does the generated query mean what the user asked? Schema-aware validation, not just “does it run.” 
• Golden-dataset benchmarking — a curated set of natural-language prompts mapped to expected query intent and result sets, giving the model a fixed accuracy reference instead of a moving target.
• Hallucination checks — does the assistant invent columns, tables, or joins that don’t exist in the schema?
• Prompt stability & reproducibility — same intent, different phrasing → consistent, correct output.
• Model-drift monitoring — an automated regression pipeline that re-runs the golden set on every model/version update and flags accuracy regressions before release.
• Anomaly-detection accuracy — false-positive / false-negative rates measured against a labelled incident dataset.
• Edge & adversarial inputs — ambiguous prompts, oversized schemas, prompts engineered to confuse the model.

Results

• NL-to-SQL semantic accuracy reached 93%, exceeding the 90% acceptance threshold (the threshold is set with the client during scoping, against their schema and their risk tolerance — it is not a fixed industry number)
• Zero hallucinated schema objects in the regression suite
• Drift gate: any accuracy drop beyond tolerance blocks the release

Workstream 2 — Cybersecurity Testing

Testing the classic surface and the new AI attack surface.

Classic surface

• Penetration testing — BlackBox / GreyBox / WhiteBox across application, APIs, infrastructure, and cloud
• Secure code review — manual + tool-assisted, targeting injection risks and logic flaws
• Vulnerability assessment with a prioritized remediation roadmap
• RBAC & multi-tenant isolation — verifying roles, permissions, and that no tenant can reach another tenant’s data

AI attack surface (Red Team / Blue Team)

• Prompt injection — crafted input that tries to make the assistant ignore its instructions and dump or exfiltrate data
• Jailbreaks & guardrail validation — can a user push the assistant past its allowed actions?
• Data leakage via the AI layer — does the assistant respect row-level security and tenant boundaries, or can it be coaxed across them?
• Model abuse — resource exhaustion, unsafe query generation

Mapped against the OWASP LLM Top 10.

Results

• All critical/major findings remediated and retested (retest included by default): 18 findings across all severity levels, 3 critical, 5 major, 6 minor, 4 trivial. 
• Zero cross-tenant data leakage, including via the AI layer
• No prompt-injection path that reaches data outside the user’s permissions
• Every finding is severity-rated (CVSS) and delivered with a reproducible proof-of-concept — the client knows exactly how many issues exist, at what severity, and how to close each one

Workstream 3 — Compliance & Audit

Turning the work above into evidence an auditor and a regulator can act on.

Framework mapping — selected according to the buyer’s market, not applied as one list

We scope to the buyer’s actual regulator.

What this produces

• A gap analysis against the relevant frameworks with a remediation roadmap
• An audit trail / logging review for AI decisions — can every AI action be traced and explained?
• Governance context attached to every finding, produced by a certified governance specialist working alongside the testing team — so findings arrive ready for auditors, not as a raw bug list

Results

• EU AI Act readiness gap analysis delivered with prioritized remediation
• Every AI-driven database action is logged, traceable, and explainable
• Compliance evidence pack structured to survive an external audit
• The number of major and critical issues found in the production environment dropped by 80% 
• Within a month after the completion of testing, the number of user complaints dropped by 50%

Timing lever: EU AI Act enforcement lands August 2026. AI-product teams in regulated markets need this evidence in place now.

How we deliver — what you actually get back

A scanner or a monitoring platform hands you a dashboard and leaves the interpretation, the judgement, and the remediation to you. We hand back the opposite: a verdict and the evidence behind it. You get a clear answer to “is this safe to ship,” not a list of signals to triage yourself.

• Golden datasets — a fixed accuracy reference for AI behaviour
• Release verdicts — a clear ship / don’t-ship call backed by reproducible runs
• Evidence packs on every engagement:
  • Executive summary — business-risk view for leadership, non-technical
  • Technical report — reproducible evidence, CVSS ratings, step-by-step remediation
  • Free retest — validation that critical issues are actually closed

How we work with your team

• Flexible engagement — full-time or part-time; no fixed commitment on team size
• Scales up and down with your release cycle
• Embedded, not siloed — we work alongside your dev team, not over the wall

Our working stack

• Test management: QASphere, TestRail
• AI-assisted engineering: Claude Pro (incl. Claude Chrome Extension), ChatGPT Plus, GitHub Copilot, Gemini — used to accelerate test design, coverage analysis, and reporting, with human review on every output

Same principle across all three workstreams: the work is done for you, and you walk away with documented, auditable proof — for your own release decision and for your customers’ auditors.

Credentials

Company: ISO/IEC 27001:2022 · CMMI Dev Level 3 · ISTQB Certified Testers · Cyber Essentials · European Software Testing Awards 2021–2023

Governance & Compliance Lead

• AIGP — AI Governance Professional (IAPP)
• CIPP/E — Certified Information Privacy Professional / Europe (IAPP)
• CIPM — Certified Information Privacy Manager (IAPP)
• CAMS — Certified Anti-Money Laundering Specialist (ACAMS)
• CAFCA — Certified AML FinTech Compliance Associate (ACAMS)

In short

When a data platform ships an AI feature into a regulated market, it doesn’t get one new testing problem — it gets three, and they overlap.

A wrong answer from the assistant is a correctness bug. The same path can be a security exposure and neither is something an auditor will take on trust. Tested separately, the failures that live between those surfaces go unseen.

This is the engagement model that closes all three together: AI validation (golden datasets, drift gates, hallucination checks), security (classic pentest plus AI red-teaming against prompt injection and data leakage), and compliance (framework mapping, AI decision logging, governance context).

One embedded team, one shared evidence trail, one release verdict — and an evidence pack your auditors and your customers can both read.

Whether your product is a database platform, a DevOps tool, or a data pipeline, the shape is the same: own the testing, do the work, hand back proof.