Patterns you can use today
Guest blog: By Jack Ducat, Co-founder of Gojee
At Xero, we’ve been posing a question lately: Is everyone a developer now? With the current pace of change, the lines between ‘technical’ and ‘non-technical’ roles are blurring. To explore this, we spoke to Jack Ducat, co-founder at Gojee, about how businesses can innovate inclusively. We looked at how ‘human-in-the-loop’ AI helps you grow your product and your team without losing the human touch.
Why this matters right now
AI gets better every week and the pressure to use it to stay competitive is real. While it can lift productivity, reduce toil, and unlock new user experience (UX),used recklessly, it creates technical debt that compounds fast. The ‘so what’ of human-in-the-loop AI is that it allows you to bypass this hype and focus on sustainable process.
When you keep humans in the loop, you open the door for everyone to be part of the development process. We call this inclusive innovation. By using AI to handle the ‘heavy lifting’ of drafting and documentation, your designers, product managers, and subject matter experts can contribute to the how of a product as much as the what.
This allows your team to move with pace and confidence. You get to ship faster and reduce daily toil, but because a human is always there to sense-check the output and maintain clear boundaries, you don’t trade off quality or safety for speed.
“AI is a capable companion, but the human is the navigator. Inclusive innovation happens when we use AI to lower the barrier to entry for the whole team while keeping a firm hand on the architectural wheel.”
Where AI fits — and where it doesn’t
To innovate effectively, treat AI as a partner for “zero to one” tasks. Use it for high-volume, low-risk drafting that helps your team move from a blank page to a first version:
- Drafting tickets: Turn rough notes into clear tickets with acceptance criteria.
- Proposing UI microcopy: Let AI draft tooltips or error messages for a designer to review.
- Building test cases: Generate initial test-case inventories and boundary lists before you start coding.
- Summarising PR threads: Create actionable next steps from complex discussions.
- Planning refactors or migrations: Draft initial plans from a brief, including risks and rollback steps for a developer to verify.
- Documentation: Generate first-pass release notes, change logs, or onboarding runbooks
When to keep AI out of the loop
Innovation requires professional judgement and empathy. There are areas where human logic must remain deterministic (fixed and certain). Keep humans firmly in charge of:
- Architecture: Design decisions for service boundaries, failure modes, and performance budgets.
- Data models: Schema and domain model changes (naming, constraints, invariants).
- Core business-logic: Anything that impacts money, permissions, or compliance.
- Security: Data policies, roles, visibility rules, retention, and residency.
- Irreversible operations and migrations: humans must design, review, and approve.
This split keeps trust high. AI proposes; humans and rules decide.
How AI enhances inclusivity (and how it hinders it)
AI enhances inclusivity when it acts as a translator. It can turn a designer’s rough idea into a technical requirement, or a developer’s complex code change into a clear summary for a stakeholder. It breaks down the “wall of jargon” that often keeps people out of the room.
However, AI can hinder inclusivity if it’s used as an autopilot. Over-automation can lead to “opaque decisions” where the team no longer understands why a feature works the way it does. This creates new silos and technical debt that can be difficult for a growing team to manage.
“The goal isn’t just to write code faster; it’s to build a shared understanding across the business. That’s how you grow without creating technical debt.”
The operating model: Build with AI, decide with humans
The biggest gains come when you apply human-in-the-loop thinking to how you build and ship software, not just where you drop AI into the product.Use this simple operating model:
- Build with AI, decide with humans. Ask an AI agent for options, trade-offs, and risks. A human picks the approach.
- Draft with AI, refine with humans. Let AI propose test cases, scaffolds, or doc outlines. A human edits for correctness and style.
- Review with humans, learn with AI. Humans own the review and the merge; your AI agent learns from the outcome.
Your first production-ready pattern: the conventions playbook
The key to balanced growth and inclusive innovation is consistency. Before you ask your AI agent to contribute, you need a single source of truth for how your codebase and your team work. Jack recommends capturing these as a ‘Conventions & Decisions’ manual.
Think of this as your team’s ‘shared brain’. By keeping it in your repository and treating it like code, you ensure AI follows your unique standards rather than generic patterns.
What to include
- Architecture and domain glossary: names, invariants, and “never break” rules.
- Coding conventions: file layout, error handling, logging, dependency boundaries.
- Review checklists: what ;good’ looks like for tests, security, and performance.
- AI working agreement: what the agent may draft, what it must never change, and how humans review its output.
- Change history: when conventions change, explain why and link to examples to help both humans and AI learn from the evolution.
How to make it durable
To ensure this manual remains a living asset rather than a forgotten document, Jack suggests three simple habits:
- Reference it constantly: Refer to this document in every PR discussion.
- Demand citations: Ask AI agents to cite the relevant section when they propose changes.
- Iterate monthly: Review and extend the document as the system and team evolves.
A low-risk experiment to accelerate learning
Create your team’s first draft of the Conventions & Decisions manual today. Give your AI agent a read-only token to your repo, scoped to PR comments only (no writes, no secret access). Ask it to read the entire PR history, synthesise the learnings, and condense them into a markdown document. Then your team reviews, edits, and adopts the best parts into your playbook. You get a distilled record of tribal knowledge, with minimal risk and a clear human checkpoint.
Measuring success without shipping blind
Track outcomes so you know when to tighten or loosen guardrails.
Product metrics
- AI suggestion acceptance rate and the reasons humans accept or reject.
- Median review time and exception backlog size.
- Post-release corrections linked to AI-assisted suggestions.
Engineering metrics
- Time from proposal to merge when AI assists versus when it doesn’t.
- Defects associated with AI-assisted changes and their root causes.
- Frequency of updates to your conventions playbook and the impact on review quality.
Use these trends to decide where to expand AI’s role and where to pause.
Common pitfalls (and fixes)
- Opaque AI decisions. Fix: always show a rationale and keep a safe human fallback.
- Prompt and pattern drift. Fix: version your conventions and require agents to reference them.
- Over-automation. Fix: keep deterministic controls on core logic; move AI inward only when accuracy and acceptance stay high.
Takeaway
Humans won’t be completely replaced, but the job shifts toward oversight as agents, guardrails, and processes mature. Start small: capture your conventions, put humans in the loop, and let AI do what it’s good at. That’s how you gain speed without inviting debt you can’t repay.
<hr /><p>Human-in-the-loop AI: Building your team’s innovation playbook was originally published in Xero Developer on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>