Questions|Dr. Shashwat Bishwen

Technical Due Diligence Questions Investors Actually Need Answered

The questions a data room will not answer on its own, organized by the risk category each one is designed to surface.

Twenty technical due diligence questions, grouped by architecture resilience, key-person risk, cloud unit economics, security, and agentic AI liability, written for investors and acquirers who want evidence rather than reassurance before closing.

Most technical due diligence interviews produce confident answers, because the person answering wrote the system being evaluated. The questions below are built to require evidence instead of confidence: a log, a metric, a git blame output, a replay of what actually happened, not a description of what should happen.

Architecture & Runtime Resilience

What happens to this platform at 10x current peak traffic?

Ask for evidence from an actual synthetic load test, not a capacity-planning slide. Most targets have never tested beyond their last real traffic spike, which means the stated ceiling is a guess, not a measurement.

Where are the single points of failure, and does the team already know about them?

A team that can name its own SPOFs unprompted is a materially different risk profile than one that answers "we have not found any." The second answer usually means nobody has looked.

How does the system behave when a downstream dependency goes down?

Look for graceful degradation and circuit breakers, not a cascading outage. Ask for the last three incidents where a third-party API or database replica failed, and what actually happened to user-facing traffic.

What is the real replication lag under write-heavy load, not the dashboard average?

Averages hide tail latency. Ask for the p99 replication lag during the worst 10 minutes of the last quarter, not the number shown on a steady-state Monday afternoon.

Key-Person & Organizational Risk

Could a newly hired senior engineer deploy and recover this system using documentation alone?

This is the single most predictive question for post-acquisition integration risk. If the honest answer involves "they would need to call someone," that person is a structural dependency the cap table does not price in.

What percentage of the critical codebase was authored by one person who might leave?

Pull the git blame distribution for the modules that handle billing, auth, and the primary data layer specifically, not the whole repository average, which dilutes the real concentration risk.

What does the on-call rotation actually look like, and who gets paged most often?

A rotation on paper means nothing if one engineer is paged for 80% of Sev-1s because they are the only one who understands the subsystem. Ask for the raw pager data, not the rotation schedule.

How is access control and credential offboarding handled when someone leaves?

Shared root credentials and un-rotated API keys after a departure are a governance smell that predicts how the rest of the engineering org is run.

Cloud Economics & FinOps

What is the real hosting cost per transaction, and how has it trended over 24 months?

Unit economics that are flat or improving as volume grows indicate genuine operating leverage. Unit economics that worsen with scale indicate an architecture that was never built to be efficient, only to work.

How much of the cloud bill is reserved capacity versus on-demand?

A target running almost entirely on-demand compute at scale either never committed to a growth forecast or has unpredictable, spiky usage that makes forecasting hard. Both are worth understanding before close.

What is driving cross-AZ or cross-region network egress cost?

Egress line items are one of the most commonly underestimated costs in a data room. A chatty microservice architecture across availability zones can quietly add double-digit percentage points to infrastructure spend.

Is observability and logging spend proportional to the system, or is it inflating silently?

Unindexed, high-cardinality telemetry is a classic cost leak: teams add logging during an incident and never remove it, and the bill compounds month over month without anyone noticing.

Security & Code Quality

What known CVEs exist across the dependency tree and container base images right now?

Ask for the actual scan output, not a statement that "we patch regularly." A live CVE count with severity ratings is the only way to know whether the diligence team is looking at a clean codebase or a liability.

Does any open-source dependency carry a viral license (GPL/AGPL) that could expose proprietary code?

This is frequently missed in standard code reviews and can materially affect the IP value of the acquisition if the target has distributed software incorporating copyleft-licensed components.

What is the automated test coverage on the paths that handle money or customer data?

Overall coverage percentages are close to meaningless. Coverage on payment flows, auth, and data-deletion paths specifically is the number that predicts post-acquisition incident rates.

How many production rollbacks happened in the last year, and why?

Rollback frequency and root cause are a direct proxy for release discipline. A team that cannot explain its own rollback history clearly has not been tracking the signal that matters most for stability.

Agentic AI & Autonomous Systems

What are AI agents in this stack allowed to do without a human in the loop?

Map every autonomous action with write access to production data, APIs, or customer-facing systems. The liability is rarely in the model, it is in the scope of what the agent was given permission to execute unsupervised.

Is there an audit trail for every autonomous agent action, and can it be replayed?

Without deterministic logging, a post-incident investigation into what an AI agent actually did becomes a guess. This is a forensic gap, not a hypothetical one, for any target running agentic workflows in production.

How are retrieval-augmented generation systems isolated across tenants?

A shared vector store without strict row-level security is a cross-tenant data leakage risk that a standard security review, built for traditional databases, is unlikely to catch.

Is there a kill switch, and has it ever actually been tested?

A documented kill switch that has never been exercised in a drill is a policy, not a control. Ask when it was last triggered, even in a staging environment.

Need these questions answered on a live target?Direct pre-close engagement by Dr. Shashwat Bishwen for PE, VC, and acquirers worldwide.
Commission an Audit →