Software development

How we use AI coding assistants, and the limits we recommend

12 min readBy Sandeep Mahajan

We use AI coding assistants (Claude Code, GitHub Copilot and Cursor, plus chat tools such as ChatGPT and DeepSeek), and we keep one firm rule: client code and data go only into the approved business-plan tools. The evidence for a large productivity gain is weaker than the marketing: the 2025 DORA report finds that AI "amplifies what's already there" and still has "a negative relationship with software delivery stability"; METR's controlled study found experienced developers were 19% slower with early-2025 tools, and its 2026 follow-up could not reliably measure the improvement; Veracode found models chose an insecure implementation 45% of the time when given a choice. The safest working assumption: assistants speed up drafting, and nothing else gets faster unless review, tests and scanning keep pace. This article states what we use and our data rule, then sets out the controls we recommend.

What is practice and what is recommendation. The tools we use, the client-data rule and the disclosure in our proposals are Impinge practice. The uses, limits and review gates further down are recommendations drawn from the published evidence, for any team adopting these tools.

Key takeaways

  • Adoption is near-universal (90% of DORA's 2025 respondents use AI at work; 84% of Stack Overflow's 2025 respondents use or plan to use AI tools), but trust is not: 46% of developers distrust the accuracy of AI tools against 33% who trust it.
  • Measured effects on delivery are mixed. DORA 2025 found a positive relationship with throughput and a negative one with stability; METR's randomised study found a 19% slowdown for experienced developers in early 2025 and only "very weak evidence" of improvement a year later.
  • Security does not come for free: in Veracode's 2025 tests, models introduced vulnerabilities in 45% of tasks, with Java failing over 70% of the time.
  • Our data rule: client code and data go only into approved business-plan tools (Claude Code, GitHub Copilot, Cursor); DeepSeek and consumer chat tools are for non-confidential questions only.
  • The rules we recommend, in one line: assistants may draft, explain and test; only humans commit, review, approve, and touch credentials, customer data, payments or authentication.

Why we are publishing a policy

Buyers evaluating an agency in 2026 ask three questions about AI tools: is my code or data used to train a model, who owns what the assistant wrote, and how do you know it is safe. They deserve written answers. This page is linked from how we work and from our AI consulting offer, where the same discipline applies to the AI features we build for clients.

What the evidence says

We read the research before writing the rules. None of it is about our team.

Source (date) What it measured Finding
DORA 2024 report (Google Cloud) Survey; AI adoption against delivery metrics Over 75% "rely on AI for at least one daily professional responsibility"; as adoption increased, an "estimated decrease in delivery throughput by 1.5%, and an estimated reduction in delivery stability by 7.2%"
DORA 2025 report (24 September 2025; nearly 5,000 respondents) Same, one year on "90% of survey respondents report using AI at work"; a positive relationship with throughput and product performance; "AI adoption does continue to have a negative relationship with software delivery stability"; "AI doesn't fix a team; it amplifies what's already there"
METR (10 July 2025) Randomised trial: 16 experienced open-source developers, 246 real tasks, mainly Cursor Pro with Claude 3.5/3.7 Sonnet Developers took 19% longer with AI; they had expected a 24% speed-up and afterwards still believed they had been about 20% faster
METR update (24 February 2026) Follow-up with late-2025 tools Returning developers 18% faster (95% interval from 38% faster to 9% slower); new developers 4% faster (15% faster to 9% slower); "very weak evidence" of the size, because of selection effects
Veracode (30 July 2025) 80 coding tasks across more than 100 language models Models "chose the insecure option 45 percent of the time"; Java "security failure rate over 70 percent"; cross-site scripting missed in 86% of cases, log injection in 88%
Stack Overflow Developer Survey 2025 Developer survey "84% of respondents are using or planning to use AI tools"; "more developers actively distrust the accuracy of AI tools (46%) than trust it (33%)"; 66% frustrated by solutions "almost right, but not quite"

Two readings follow. The gain is real for drafting and unreliable for delivery; DORA's "amplifier" is the honest summary. And the risks are specific and measurable (security, stability, over-confidence), so they can be managed with specific controls.

The tools we use and our data rule

  • Tools in use: Claude Code, GitHub Copilot and Cursor for coding; ChatGPT and other chat tools, and DeepSeek, for general questions.
  • Client code and client data go only into the approved business-plan tools: Claude Code, GitHub Copilot and Cursor.
  • DeepSeek and consumer chat tools are used for non-confidential questions only: no client code, no client data, nothing under a confidentiality obligation.

Where assistants help: recommended uses

  • Drafting code against a written brief: scaffolding, data migrations, API clients, configuration, repetitive transformations.
  • Writing tests for human-written code and for its own drafts, subject to the tautology check below.
  • Explaining unfamiliar code, stack traces and legacy modules before a human changes them.
  • Refactoring with a test suite in place, so the suite is the judge.
  • Documentation and commit messages, reviewed before they are pushed, and second-opinion review of a pull request, with no power to approve.

What we recommend never letting them do

We recommend rules that do not depend on the tool and have no exceptions for deadlines.

  • Commit, merge or deploy without a named human reviewer who has read the change.
  • Receive production credentials, client secrets, API keys or customer personal data, or an unredacted production database. Use synthetic or anonymised fixtures for development.
  • Write or change authentication, authorisation, payment, cryptography or consent code without a second human reviewer working from a checklist.
  • Run with write access to production systems, or with unrestricted file-system and network access on a developer's machine. Run assistants in sandboxed, permission-prompting modes; Claude Code's documentation, for example, describes a manual mode that "starts with read-only permissions" and asks before edits or commands.
  • Install dependencies, extensions or MCP servers that nobody has vetted. Anthropic states it "does not security-audit or manage any MCP server"; nor does any registry.
  • Publish client-facing copy, legal text or privacy wording without human review, or accept a suggestion flagged as matching public code without a licence review.

Recommended review gates for every AI-assisted change

  1. A human writes the brief: the task, the acceptance criteria and the files in scope. The assistant does not define the problem.
  2. The assistant drafts in a sandboxed session with permission prompts on, inside the project directory, without credentials in its context.
  3. The author reads every line and rewrites anything they cannot explain. "The assistant wrote it" is not an answer in review.
  4. Tests run and are read. The existing suite must pass; new behaviour gets new tests; assistant-written tests are checked for tautology (a test that asserts whatever the code does is not a test).
  5. Automated checks run in CI: linting, type checks, dependency audit, secret scanning and static analysis. The Veracode results are why static analysis is not optional for AI-drafted code.
  6. A second human reviews the pull request. AI review tools may comment; only a person approves.
  7. Security-sensitive paths get a checklist review (OWASP's web and LLM Top 10 lists) by a reviewer who did not write the change.
  8. The pull request records AI assistance, for example a trailer line in the commit message or a note in the description, so the history is honest.

Infrastructure should follow the same shape: tests and type checks run in CI before anything deploys, and infrastructure is code reviewed like the rest (see our DevOps guide).

Tests, scanning and licensing

Secure development practice did not change because the drafting tool did. NIST's Secure Software Development Framework (SP 800-218, version 1.1) remains the reference, and AI-drafted code goes through the same gates as any other. What changed is the volume of plausible-looking code, which is why gates 4 and 5 are automatic.

Licensing is a specific risk: assistants trained on public code can reproduce it. GitHub Copilot, for example, can block suggestions that match public code or show "references to any similar code that is found" with the licence type, set in personal or organisation policies; other tools have equivalents. We recommend turning that setting on and treating any flagged match as a licence review.

Ownership is clearer than people assume. The US Copyright Office's January 2025 report concludes that outputs "will be copyrightable in whole or in part—where AI is used as a tool, and where a human has been able to determine the expressive elements they contain", and that "prompts alone, however, at this stage are unlikely to satisfy those requirements". Where every change is specified, edited and reviewed by a human, the usual contract clause assigning deliverables to the client can cover AI-assisted code too; have a lawyer word that clause for your jurisdiction.

Client code and client data

  • Tool plans whose terms exclude training. GitHub states that it "does not use Copilot Business or Copilot Enterprise customer data to train AI models", while individual plans may use "prompts (inputs), suggestions (outputs), and code snippets" unless the subscriber opts out. Anthropic states that "by default, we will not use your inputs or outputs from our commercial products" to train its models. Our recommendation: allow only plans with a written no-training default on client work, and recheck the vendor's terms when they change. This is why our own rule keeps client code in business-plan tools and out of consumer chat tools.
  • Content exclusion for sensitive repositories. Exclude directories holding secrets, customer exports or regulated data from the assistant's context where the tool supports it; GitHub notes that content exclusion is not supported in Copilot CLI or agent mode, so exclusion complements the data rules above rather than replacing them.
  • No client data in prompts. Reproduce bugs with synthetic data; if a real record is unavoidable, redact it first.
  • Let clients opt out. A client may require that no AI assistant is used on their code; the supplier should agree it in writing and say what it does to the estimate.

How clients are told

Our proposals and contract disclose that we use AI tools and the policy that governs them, including the client-data rule above. If you have questions about a specific tool or how your code is handled, raise them before the contract is signed.

How it changes estimates

Less than the headlines suggest, and not uniformly. Drafting, scaffolding, tests, migration scripts and documentation go faster. Requirements, integration with the client's systems, security review, QA and the reading time in gate 3 do not, and DORA's stability finding is a reason to keep those hours in. A blanket "AI discount" is the wrong model; we recommend pricing the work by the gates it must pass and showing where assistance reduces hours. See also cutting web development costs.

FAQ

Will AI make my project cheaper?

Parts of it. Drafting code, tests and documentation is faster with assistants; defining requirements, integrating systems, reviewing security and testing are not, and controlled studies have not shown a reliable overall speed-up for experienced developers. A fair estimate shows the effect task by task rather than applying a flat discount.

Is my code or data used to train AI models?

Client code and data go only into the business-plan tools we have approved (Claude Code, GitHub Copilot, Cursor); DeepSeek and consumer chat tools are used only for non-confidential questions. On training, GitHub states that it does not use Copilot Business or Enterprise customer data to train its models, and Anthropic does not use commercial-product inputs or outputs for training by default. Individual and consumer plans can differ, so check the plan, not just the product.

Who owns code written with an AI assistant?

Normally the client, under the contract's ownership clause, as with any other deliverable; the contract wording decides it, so have it checked by a lawyer. The US Copyright Office's 2025 report says AI-assisted work is copyrightable "where AI is used as a tool, and where a human has been able to determine the expressive elements", one more reason to keep a named developer specifying, editing and reviewing every change.

Can I ask you not to use AI tools on my project?

Raise it before the estimate. Our use of AI tools is disclosed in our proposals and contract, so that is the place to agree it; the main effect of working without assistants is more drafting time. The data rule applies either way.

How do you know AI-written code is secure?

Do not assume it is. Veracode's 2025 tests found models chose insecure implementations 45% of the time, so we recommend that every change pass static analysis, dependency and secret scanning in CI, a human review, and a checklist review for authentication, payments and other sensitive paths.

Impinge Solutions builds web platforms, online stores and AI features, with client code and data kept to approved business-plan AI tools. To talk through how it applies to your project, book an AI readiness call or request an estimate.

Sources

Are you ready to partner with us?

Our good-natured team is waiting to hear your business requirements.

Book a free project call

Stay informed with our newsletter