Infrastructure due diligence is the part of an investment, acquisition or enterprise procurement review that asks a single question in several disguises: could someone other than your current team rebuild, operate and be responsible for this? Not whether the architecture is admirable. Whether it transfers.
That framing explains findings that otherwise look arbitrary. A reviewer who shrugs at an unusual database choice will stop cold at an environment that exists only as it currently runs, because the first is a preference and the second is an unbounded cost somebody is about to inherit.
Who is asking, and what they actually fear
#Three groups run this review with different incentives and, largely, the same checklist.
| Reviewer | What they fear | What settles it |
|---|---|---|
| Investor, technical diligence before a round | That the plan needs a platform rebuild the model has not budgeted for, and that the capability is concentrated in one or two people | Evidence the environment is defined in code, changed through review, and operable by a team rather than a person |
| Acquirer, pre-close and integration planning | Inheriting liabilities: stale access, unmet contractual commitments, an environment that cannot be separated from its founders or its accounts | An account and access inventory, the compliance posture in writing, and a clean ownership story for repositories and cloud accounts |
| Enterprise procurement and vendor risk | That you will fail in a way that becomes their incident, or that you cannot meet obligations they are contractually required to flow down | Recovery evidence, incident history, control mappings and the answers behind the security questionnaire |
Notice that none of the three are looking for perfection. They are pricing risk. A known, bounded weakness with a plan attached is a normal part of a deal. An unknown weakness discovered by the reviewer, after you said things were fine, is a different conversation entirely.
The six areas that carry the weight
#Reproducibility
#The question is blunt: if the cloud organization disappeared tonight, what would it take to have it back? A reviewer will ask when the environment was last built from source, and the answer “it has been running since 2022” is the finding.
What settles it is infrastructure defined in code with a plan that runs clean against an environment that does not exist yet. That last part is more specific than it sounds: a plan can only run end to end before anything exists if every dependency is expressible without reading live state. In the AWS Enterprise Baseline, that is one of six CI guard scripts — full mock outputs, so a brand-new environment can be planned end to end before a single resource is created. It exists because the alternative is discovering, during a rebuild, that half the configuration depends on the thing you are rebuilding.
Access control
#The reviewer wants the list of who can change production and the mechanism that keeps the list short. Two findings recur and both are expensive: credentials that outlive the person, and a former contractor whose access was never removed. Both are the same underlying problem — access granted to a stored secret rather than to an identity.
The defensible position is that human access runs through single sign-on and nothing else, that there are no long-lived keys anywhere because creating one is denied above the account, and that machine access is federated per job rather than stored. Zero long-lived cloud credentials covers the mechanism. In diligence, the value is that “who has access” becomes a query with a short answer instead of an archaeology project.
Recovery evidence
#Every company says it has backups. The reviewer asks when you last restored one, into what, and how long it took. The gap between those two sentences is where the finding lives, and it is a finding with a straightforward price attached, because an untested backup is an assumption about your worst day. We have written that argument at length.
What settles it is a dated restore record: what was restored, into which environment, how long it took, what broke, what was fixed. One page. The existence of the page is most of the signal.
Key-person risk
#This is the area founders consistently under-rate, and the one acquirers consistently weight highest, because it is the risk that does not show up in any system — it shows up in a resignation.
The fix is not heroic documentation. It is structural: the environment defined in code so knowledge lives in a repository, runbooks for the operations that matter, code owners so review is distributed by design, and at least one rehearsed handover in which somebody other than the expert performs the critical tasks unaided. Our own delivery ends with exactly that — a hands-on session where the customer’s team performs 11 tasks, unaided, because a handover nobody executed is a document rather than a transfer.
Cost governance
#Investors ask about unit economics and get an answer about gross margin. The follow-up is about infrastructure spend per customer, and whether the company knows which workloads drive it. A cloud bill nobody can attribute is a forecasting problem, and forecasting problems in diligence become discount rates.
The practical bar is lower than people fear: required tags applied consistently enough that spend maps to teams and environments, a policy that reports on missing tags rather than hoping, and cost visibility inside the change process so that the price of a change is known before it merges rather than at month end. In the AWS Enterprise Baseline, tagging is a policy with six required tags, running in report-only mode by default and enforceable with one flag, and pull requests can carry a cost estimate as a comment.
Change control and compliance posture
#The last area is the one that connects to everything else in this series. How does a change reach production, who approved it, and can you show the history? If every infrastructure change is a reviewed pull request applied by a pipeline, the answer is a link. If changes are applied from laptops, then no control you describe can be evidenced, because there is no record of what was actually done.
Attached to this is the compliance posture: which frameworks you map to, what is enabled, what your last audit or readiness assessment found. Reviewers are not impressed by an unqualified claim of compliance, and several of them will check. A precise position — controls mapped, evidence automated, certification described as the auditor’s decision — reads as competence. The SOC 2 version of that posture is the usual starting point.
The findings that cost you something
#None of these are fatal on their own. All of them turn into terms.
| Finding | What the reviewer infers | Typical consequence |
|---|---|---|
| The environment cannot be rebuilt from source | Unbudgeted platform work, and a migration risk on any move | Remediation as a condition, or a lower number |
| Former staff or contractors still hold access | Nobody owns the access lifecycle; other controls are probably also aspirational | Immediate remediation request, and a wider security review |
| Backups exist but have never been restored | Business continuity is an assumption; recovery time is unknown | A holdback, or a covenant with a deadline |
| One person is the only route to production | The asset has a notice period | Retention terms, escrow, or a lower number |
| Cloud spend cannot be attributed | Unit economics are estimates | More diligence, slower close |
| Customer contracts promise controls the environment does not implement | Contractual exposure being inherited | Legal review, indemnities, delay |
| Production changes are applied by hand | No change evidence exists, so no control claim can be verified | Findings across every other area at once |
The pattern is worth naming. A single finding is a negotiation. Three findings from different areas stop reading as items and start reading as a culture, and that is what actually moves a valuation.
The self-audit to run before anyone else does
#Run this on a normal Tuesday. Each question is a yes, a no, or the answer that should worry you most, which is “I would have to check.”
| # | Question | What a no means |
|---|---|---|
| 1 | Could we rebuild the whole environment from the repository into empty accounts? | Reproducibility is a claim, not a property |
| 2 | Has anyone run a plan against an environment that does not exist yet, successfully? | The rebuild would fail in the middle, and you would find out then |
| 3 | Can we list everyone with production access in under five minutes? | Access is not governed, it is remembered |
| 4 | Does any long-lived cloud credential exist anywhere, including in CI? | A password with no expiry is your perimeter |
| 5 | Have we restored a backup in the last 90 days, and do we know how long it took? | Recovery time is a guess |
| 6 | If the infrastructure lead left next week, would production be fine? | You have a person-shaped dependency |
| 7 | Is every infrastructure change a reviewed pull request applied by a pipeline? | You cannot evidence any control that touches change |
| 8 | Can we say which team and environment each dollar of cloud spend belongs to? | Unit economics are estimates |
| 9 | Do we know what our customer contracts commit us to, technically? | You may already be out of compliance with an agreement |
| 10 | Could we produce a quarter of access, change and configuration evidence in an hour? | Every review will be a project |
| 11 | Would we deploy on a Friday without anyone being nervous? | The nervousness is information about the system |
| 12 | Has anyone other than the expert performed the critical operations unaided? | The handover is untested |
Three or more answers that are not a confident yes, and the honest read is that a diligence process would find them before you did.
What to do with the result
#Fix in this order, because it is the order of cost-to-fix rather than the order of severity: remove stale access this week, run a restore drill this month, put change behind review this quarter, and treat reproducibility as a project with a date. Recording and evidence automation should start immediately regardless, since evidence has a clock on it and none of it can be backdated.
The deeper point is that all four of these are things you want anyway. Diligence readiness is not a costume put on for investors; it is the same property that makes a platform operable, an audit cheap and an on-call rotation survivable. That is why the BuiltForProd Standard treats them as one set of properties rather than several programs.
If a process is already scheduled and you want to know what it will surface, the BuiltForProd Assessment runs the same inspection against your environment and gives you the findings while you still have time to act on them. The situation itself has a page: due diligence readiness.