Skip to main content

What are you looking for?

Explore our services and discover how we can help you achieve your goals

Client Workspace

Human review gates for AI-built software: the checklist a buyer should ask a partner for

Human review keeps AI-built software safe when the partner runs a named human gate at every stage where AI output can go wrong: requirements, design, code, tests, security, IP and release. This Netbase checklist lists the seven gates, the evidence each one should leave and the gates a buyer should never let a partner skip.

Book a solution review See quality engineering and testing

By Netbase Delivery and Quality Desk · Reviewed by David (CEO) · Updated · 10 min read

star

This guide is for founders, CTOs and procurement leads buying custom software development from a partner whose engineers use AI coding assistants. It belongs to the partner-selection series that starts with large or boutique Vietnam IT vendors.

In this guide

Why is "a human reviews the code" not enough?

A single code review at the end cannot carry AI-built software. AI tools repeat one misunderstanding across code, tests and documentation, so if the specification was read wrongly, the code and the tests that "prove" it share the error. A late reviewer facing a large diff tends to skim.

The fix is not more reading at the end but smaller human checkpoints in the places where AI output goes wrong. A Thoughtworks article on review gates describes the same idea: the AI pauses at defined points and waits for a person's explicit approval, with smaller inner gates during the work and an outer gate when a milestone completes.

For a buyer, the question to ask is therefore not "do you review AI code?" but "which gates do you run, who signs each one, and what record does each leave that I can read?"

Which seven review gates should a buyer ask for?

Each gate below catches a different way AI-built software fails. Ask the partner where each sits in their process.

Gate What AI can get wrong here Evidence to ask for Who signs
1. Requirements Fills gaps in a vague request with plausible guesses Written acceptance criteria per feature, agreed before build Your product owner and the partner's business analyst
2. Design Adds patterns, services or data models nobody asked for Short design note or decision record for each significant change Partner's solution architect
3. Code Large diffs that look correct, duplicated logic, invented APIs Pull request with a named human approver who can explain the change A developer other than the author
4. Tests Tests that pass because they share the code's wrong assumption Tests for critical paths written or checked by a person against the acceptance criteria QA engineer
5. Security Insecure patterns, secrets in code, packages that do not exist Scan results plus human review of authentication, payments and personal data code Security-aware reviewer
6. IP and licences Dependencies with incompatible licences, unclear origin of copied code Dependency and licence report; statement of which AI tools touched your code Partner's delivery lead
7. Release Changes merged without the earlier gates closed Release record linking each change to its gates and its milestone acceptance Partner's delivery lead and you

Gate 4 matters most with AI. The OWASP Secure Coding with AI cheat sheet warns that a passing test suite generated by the same agent that produced the code gives no independent assurance, and it asks for tests of authentication, authorisation, input validation and cryptography to be written manually.

Gate 5 has its own guide: the security review gate for AI-generated code sets out the four security checks inside it in engineering detail.

Which gates can be lighter, and which are never skipped?

Not every change needs all seven gates at full depth: a copy change and a change to refund calculation carry very different risk. A workable rule groups changes by what they touch:

  • Never lighter: payments, authentication and permissions, personal data, database migrations, infrastructure and pipeline configuration, and anything a regulator or auditor will ask about. All seven gates apply, and gate 4 tests are written or checked by a person.
  • Standard: new features and integrations with no sensitive data. Gates 1, 3, 4 and 7 run in full; design and security are reviewed when the change alters structure or adds a dependency.
  • Light: content, styling and internal tooling with no data access. One human approval at gate 3 plus automated checks is usually enough.

The trade-off: lighter tiers save review time, but a mis-classified change gets less scrutiny. Ask who decides the tier, and have it shown in each pull request so a sensitive change filed as light is visible.

How do you write the gates into a proposal and contract?

  1. Ask for the gate map in the proposal

    Request a one-page table like the one above, filled in with the partner's real roles. A partner that already runs these gates can produce it quickly; a blank answer is the finding.

  2. Agree acceptance criteria before work starts

    Gate 1 is the foundation for every later gate. The guide to what AI-native delivery changes for the buyer covers milestone acceptance terms.

  3. Make the repository enforce the gates

    Ask that the main branch require an approving review before merging and that a review is dismissed when new commits change the diff. GitHub's documentation on protected branches describes both settings, plus code-owner review for sensitive folders.

  4. Inspect one real change

    Before signing or at the first milestone, trace one change from requirement to release record.

  5. Write gates and disclosure into the agreement

    Name the gates, the tiering rule, disclosure of the AI tools used on your code and IP ownership of AI-assisted output, and tie milestone acceptance to the release record of gate 7.

How can you tell a review gate is only a rubber stamp?

A gate is only on paper when its records show approval without engagement. Warning signs in a sample of pull requests:

  • approvals that arrive within a minute or two of a large pull request opening;
  • the same person writing and approving most changes, or the AI tool listed as the reviewer;
  • review comments that never ask a question or request a change;
  • tests that were added in the same commit as the code and never fail in the history;
  • release notes that list features but cannot be traced back to acceptance criteria.

One sign proves little; several together mean AI-built code reaches your users with no independent human check.

Where AI fits

In a Netbase project, AI drafts code, tests, migrations and documentation from the agreed specification and runs first-pass checks such as dependency and pattern scans. Engineers still own every gate: they confirm the change meets the acceptance criteria, check that tests test the right thing, and review security-sensitive code, data access and integrations themselves. Netbase works with the major commercial and open-source AI tools and models, chosen per project, so the gate map stays the same whichever tool a project uses.

What delivery record exists, and what does not?

Netbase delivers its own projects with AI-assisted engineering under human review, in Agile increments with weekly reviews and security and API design from the start. Its published security practice includes secure code review and version control, role-based access control, vulnerability scanning and penetration testing, and contributors under NDA.

Most Netbase projects are delivered on fixed-price contracts, with milestone-based, monthly team retainer and KPI-linked models also offered. The delivery lifecycle runs from discovery through agile execution with outcome-based milestones to rollout and ongoing support, which is where acceptance at gate 7 is agreed. For custom development, the client owns the IP created for it; Netbase's productized modules are licensed, not transferred.

Netbase holds ISO 27001 certification and a SOC 2 Type II attestation for its own operations. Netbase also works as offshore development partner on a multi-tenant cloud ERP for a US client that is not named, a production SaaS system of the kind these gates protect. The delivered AI feature named with a client is 4over4's recommendation engine; beyond it, Netbase has delivered AI projects for clients that are not named, including retrieval-augmented knowledge assistants, document AI and MLOps pipelines.

What does not exist: no published figure for defects caught per gate, review time or speed gained from AI on any Netbase project, and no client named for the anonymised work. This page gives a method a buyer can apply to any partner, not results.

Alternatives: other ways to control AI-built software

Approach Strength Risk Choose it when
Partner runs all seven gates One accountable party, fewer hand-offs You rely on the partner's records You have no in-house engineering lead
Your engineers review every pull request Full control of code and design Your team becomes the bottleneck You have senior engineers with time to review
Independent QA or audit partner A second opinion separate from the builder Added cost and slower feedback The system is regulated or high-risk
Ban AI tools on the project Removes AI-specific risk Gives up speed; hard to verify A contract or regulator forbids AI use on the code

Whichever you pick, quality engineering and testing is where gates 4 and 7 live day to day. If you are still deciding what kind of partner to hire, the AI-augmented vs AI-native proposal checklist helps you read how each vendor uses AI before you ask for the gates, and the red-flag matrix for proposals, contracts and demos covers the rest of the paperwork.

Limits of this guide

  • It covers software built with AI assistance, not AI features inside the product; those need evaluation sets and model monitoring as well.
  • The gate map is a starting point. Regulated sectors add their own approval and audit duties on top.
  • Research on AI in delivery is moving fast. The DORA 2025 report on AI-assisted software development finds that AI amplifies an organisation's existing strengths and weaknesses, which is why weak review processes get worse with AI, not better.

Plan the next step with a Netbase consultant

Frequently asked questions

Yes. Every AI-assisted change should have a named human approver who is not the author and who can explain what the change does. Automated scans and AI review tools can run first, but they add signals; they do not replace the human approval.

They can help as a first pass that flags issues for a person, but they cannot be the gate. A model reviewing output from the same or a similar model tends to share its blind spots, so approval for merge and release should stay with a human reviewer.

Ownership follows the contract, so write it down. For custom development, the agreement should state that the client owns the code created for it, including AI-assisted output, and the partner should disclose which AI tools touched the code.

Next step

Bring a proposal or a sample pull request from your current partner, and Netbase will go through the seven gates with you: book a solution review.

AI-enabled custom software development built around your workflow AI-enabled custom software development built around your workflow

Netbase provides custom software development for companies whose workflow no off-the-shelf product fits, from discovery through architecture, build, testing and handover, with AI where it removes manual steps and the IP left with you. Since 2012 Netbase has worked with 500+ businesses across 80 countries; according to the published USticker case study, design-tool users converted 40% more often.

Learn More
line
AI-assisted QA and software testing that run inside delivery AI-assisted QA and software testing that run inside delivery

Netbase provides software testing and QA services for product and commerce teams to release often without breaking what customers use. QA runs inside delivery, not after it: acceptance criteria, automated regression with AI-generated test drafts, performance budgets and release readiness checks. Page speed shows the result: USticker's page load fell 21%, and PrintLeo's improved 35%.

Learn More
line
Application security assurance for web, SaaS and AI features Application security assurance for web, SaaS and AI features

Netbase provides application security testing for product and engineering leaders to find and fix vulnerabilities before release, including in AI features. We review code with AI-assisted triage, scan dependencies and test running applications, then rank findings and retest fixes. Netbase's own alignment with ISO 27001 and SOC 2 Type II covers Netbase's operations, never your application.

Learn More
line
SaaS product accelerator: launch an AI-ready SaaS on proven Netbase modules SaaS product accelerator: launch an AI-ready SaaS on proven Netbase modules

A SaaS product accelerator is a set of reusable Netbase modules for accounts, billing, roles and integrations that helps founders and product teams launch subscription software faster, with room for in-product AI from the first release. Reusing these modules can shorten development time, and the approach is proven on Printcart, Netbase's own web-to-print SaaS.

Learn More
line

Work with Netbase remotely

Every Netbase engagement is delivered remotely, so clients in any country and any industry can work with us; Netbase has worked with clients in 80 countries. Contracts and invoices are issued by Netbase JSC, and payment is made by bank transfer to Netbase JSC's company account or PayPal Business. Talk to the CEO, Huy Nguyen (David), on WhatsApp or by email.

Contact Netbase

Discuss a project

Netbase JSC helps organizations design, build, modernize, and operate digital products and AI-enabled business systems.
Project enquiries

project@netbasejsc.com

WhatsApp

+84 937 869 689

Office address

91 Nguyen Chi Thanh, Dong Da, Hanoi, Vietnam

Get in touch

Tell us what you want to build, modernize, or operate.

Tell us what you want to build, modernize, or operate.

Contact Netbase