Skip to content
Technology transformation expertise firm · ARDIANE Group Innovate, secure, succeed

Engineering

The questions to examine before rolling out AI-assisted development

In brief

AI-assisted development refers to developers' use of assistants based on language models that suggest code, tests or explanations. Management Square sets out here the questions a committee can examine before rolling it out: tasks assigned, review, tests and code review, data transmitted, ownership of the code and the measures for a trial.

Page contents

Take a road haulage company: an in-house team develops its route planning software. Two developers already use a coding assistant, on their own initiative, and the team lead asks the IT committee to extend this AI-assisted development to everyone.

The expression refers to the use of assistants based on a large language model, which suggest code, tests or explanations, from completing a line to an agent carrying out an entire task. Before rolling out this use, the committee can examine, in order, the tasks assigned, review, the role of tests and code review, the data that leaves the organisation, the ownership of the code and how to judge a trial. These questions and their order are our proposal; the public sources cited were read on 12 November 2026, and no gain is promised.

Which tasks should be given to a coding assistant?

We suggest classifying tasks according to two criteria: whether a means of checking the result is already in place, and the cost of an error that goes unnoticed.

Tasks whose result can be checked straight away lend themselves to assistance: repetitive code, conversion from one format to another, explanation of a module the developer does not know, a first draft of tests from a written rule.

Others rest on past reasons and decisions that the context passed to the tool does not always contain: the breakdown into modules, the data model, a business rule that nobody has yet written down. The assistant can suggest options; a person decides and records their reasons.

A third category calls for reinforced review: access control, authentication, encryption, calculation of an amount. When the assistant acts as an agent, the question of its rights is added; it is covered in our article on the security of AI agents.

Who reviews what the assistant suggests?

The developer who submits a change is accountable for it to the team, even when the tool suggested most of it. A team rule expresses this: only submit what you can explain.

The OWASP Top 10 for LLM and generative AI applications (genai.owasp.org) supports this requirement. Its “Misinformation” entry (LLM09:2025) describes the overreliance placed on generated content and cites, among its examples, the generation of unsafe code: the model suggests insecure or non-existent libraries. One of its scenarios features attackers who identify package names often invented by assistants, then publish malicious packages under those names. Any new dependency suggested by the assistant is therefore checked before it is added: existence, publisher, maintenance.

Review by a second person keeps the place given to it in our article on code review. Maintainability of the code by others, which our page on custom development treats as a design requirement, is not relaxed because a model suggested part of it.

Do tests and code review change with AI-assisted development?

If the number or size of submitted changes increases, code review can become the team’s bottleneck. Its rules on size and turnaround time are restated before the tool is opened up.

A test that the assistant derives from the code it has just written checks that code against itself, a limitation noted in our page on software quality and testing. The expected behaviour is therefore written first, from the rule or the acceptance criterion, by a person who knows the need; the assistant can then add further cases.

The automated checks in the integration pipeline (static analysis, secret detection, dependency analysis) apply to suggested code as to the rest, before any human review.

Which code and which data leave the organisation?

To suggest a continuation, the assistant sends context to the service that runs the model: depending on the tool, the open file, neighbouring files, the question asked. This context may contain an algorithm specific to the organisation, a password left in a configuration file, personal data in a test data set. OWASP’s “Sensitive Information Disclosure” entry (LLM02:2025) counts security credentials and confidential business data among sensitive information, and recommends clear rules on data retention, use and deletion.

The committee therefore asks, for the candidate tool, what is transmitted, where and for how long it is retained, whether inputs are used to train models and which subprocessors are involved. The answers are found in the contract and the documentation. The GitHub Copilot documentation (docs.github.com) describes a setting that makes the assistant ignore certain files, and states that it does not constitute a file system security boundary. The GitLab Duo documentation (docs.gitlab.com) describes, in certain use cases, the removal of secrets before code is sent to the models, and states that its model subprocessors may not use inputs and outputs for training.

As long as use goes through personal accounts, these answers are missing: this is the situation named by shadow AI. A model hosted by the organisation shifts these questions without removing them, a trade-off presented in our page on generative AI in the enterprise. The rules of use (authorised tools, data that must not be entered, reporting) belong in the corporate AI charter.

Who owns the code produced, and under which licence?

We state no rule of law here: these points are for the organisation’s lawyers.

  • Is code suggested by an assistant and then kept by a developer protected, and who owns it between the employee, the organisation and, where applicable, a service provider?
  • What do the vendor’s terms provide on the content produced and on a third-party claim?
  • Can a suggestion reproduce code published under a licence whose obligations would then apply to the software?
  • Do service providers’ contracts say whether the use of an assistant is permitted, and on what conditions?

On the third point, the Gemini Code Assist documentation (developers.google.com) states that Google Cloud products using Gemini attach citation information when they quote another source directly and at length, for example open source code, and that the citation may mention an applicable licence; a setting excludes suggestions matching cited sources. The Amazon Q Developer documentation (docs.aws.amazon.com) states that a suggestion may resemble public code, that references then give its source and that these suggestions can be disabled. These settings inform the lawyers without settling the matter.

How will we know whether the trial is conclusive?

We suggest running the trial as a bounded experiment: one team, one application, a duration set in advance, and measures taken on the same team before the tool is opened up.

Input measures describe what arrives in code review: number of changes submitted, size of each, share of those in which the assistant was used. Quality measures describe what becomes of them: time to first review, back-and-forth before merging, defects found in review, defects discovered in production in the code concerned. Added to these is a written record from the developers: the tasks where the help was useful, and those where it wasted time.

None of these measures amounts to a promise, and none is used to assess an individual. Before the trial, the committee writes down what would lead it to extend use, to restrict the tasks or to stop it. On the planned date, the team presents the record, any confidentiality incidents and the list of tasks it proposes to open to the assistant, to close to it or to keep under reinforced review.

FAQ

Frequently asked questions

Another question? Ask us

What is AI-assisted development?

AI-assisted development is a way of programming in which a tool based on a language model suggests code, tests, explanations or fixes from the context of the project. The help ranges from completing a line in the editor to an agent carrying out a whole task. The developer who submits the change is accountable for it to the team: the suggested code is reviewed, tested and maintained like the rest.

What is the difference between a coding assistant and a development agent?

A coding assistant suggests: it completes a line, proposes a function or answers a question, and the developer accepts or rejects each suggestion. An agent acts: from an instruction, it reads files, modifies several of them, runs commands and checks its work. Review then covers a set of changes, and the rights granted to the agent on the workstation and the repositories become a security question.

Should code written with an AI assistant be flagged?

Flagging code written with an AI assistant is a team rule to be written down. A mention in the description of the change informs the reviewer and feeds the measures for a trial, without weighing down the code. Whether a mention is owed to a client or a third party is for the organisation's lawyers to decide, in the light of the contracts.

A project to bring to completion?

A first conversation is enough to understand your challenge, identify the right people and tell you frankly what we can bring.