ProBackend
Developer Tools
4 days ago5 min read

From Whiteboard to Working Software: Where AI Fits

From Whiteboard to Working Software: Where AI Fits

A practical guide to using AI in coding, from product requirements to reviewed changes, with lessons for developer tools and teams.

From Whiteboard to Working Software: Where AI Fits

A whiteboard sketch is a useful beginning, not a specification. Someone points to a screen, an engineer asks what happens when a user has no account, and the team writes down the answer before moving on. AI assistance can help organize notes, draft code, or suggest tests. The team still decides what the product should do and whether the output meets its needs.

The workflow and recommendations below are practical suggestions from the author, not findings validated by Deloitte. Deloitte’s article on AI in software development provides broad context for considering AI in this field; it does not establish that these steps will improve a particular team’s work.

Turn the sketch into questions and requirements

As a practical first step, use the sketch, meeting notes, and unanswered questions to produce a draft. Ask an assistant to sort rough notes into candidate user stories, suggest questions for discussion, or draft acceptance criteria. Keep generated text separate from decisions the team has made.

A prompt can say: “Draft acceptance criteria from this material. Put assumptions and unanswered questions in separate lists. Do not decide what the product should do.” Compare the draft with the sketch and resolve open questions. A list of questions can give the team a useful starting point alongside a draft story.

Treat the output as a draft, not an established requirement. If it introduces a user, permission, or behavior nobody discussed, decide whether to investigate it or discard it.

Move from requirements to a small change

Once the team agrees on a behavior, a practical approach is to give the assistant bounded context: the relevant function or component, expected behavior, language or framework constraints, and a testable definition of done. Request one change at a time. Avoid asking for a whole system from a sketch; doing so leaves many decisions to resolve and can make the result harder to inspect.

A suggested review sequence is:

  1. Clarify intent. Agree on a behavior and the edge cases relevant to the task.
  2. Request a bounded draft. Ask for a function, component, test, or explanation rather than an entire application.
  3. Inspect the diff. Compare the proposed changes with the task and the project’s conventions.
  4. Run relevant checks. Use the checks the team has chosen for the project.
  5. Resolve assumptions. Confirm or revise any assumption before accepting the change.

If a proposed change is difficult to review, narrow the request and assess the smaller revision instead. These steps are recommendations for the team to evaluate, not a workflow attributed to Deloitte.

Treat research as context, not a guarantee

Deloitte’s article offers broad context about AI in software development. It can inform questions about possible applications, but it is not evidence that a particular tool will improve work in a particular codebase or that a specific change is ready to ship.

As a practical suggestion, turn broad discussion into a local question: would assistance help draft a small component, explain an unfamiliar function, or produce an initial set of tests? Evaluate any such use in the team’s own setting. General discussion of AI does not establish a productivity result for a particular team or replace its engineering checks. For additional discussion of the evidence and human review, see AI coding tools and productivity evidence.

Consider review, testing, and security

As a practical recommendation, review AI-assisted code using the same criteria the team applies to other proposed changes. Check whether it matches the agreed behavior, fits the codebase, adds dependencies, or raises security questions. Treat generated tests as drafts and identify any additional cases needed to cover the requirements and agreed failure paths.

For each meaningful change, ask which acceptance criterion it addresses and whether tests exercise the intended behavior. When a dependency or permission appears in a proposal, check whether it is needed and approved. These are suggested review questions, not claims about a specific tool’s capabilities or risks.

Set clear rules for tool use: which tools are permitted, what project material may be entered, and how to handle sensitive or uncertain output. For consequential changes, keep relevant requirements and review notes with the work. Adapt these suggestions to the team’s existing policies and review process.

Help the team review AI-assisted work

A practical way to discuss these suggestions is to use a task the team considers suitable. Have an engineer request a small change, inspect the diff, and explain what the team kept, changed, or rejected. Ask whether the proposal assumed something the team had not decided, touched files outside the task, or needs a particular test to check the behavior.

Product colleagues can review drafted requirements against intended user behavior; engineers can compare a diff with the agreed scope; reviewers can trace changes to acceptance criteria. Adapt these responsibilities to the team’s existing roles. Focus the discussion on the work and its evidence, not on whether an explanation sounds confident.

Pilot a workflow if useful

As an author’s practical suggestion, choose a suitable task, such as adding a small component with agreed acceptance criteria, and run a limited pilot before expanding the workflow. Record measures that matter to the team—for example, time from ready-to-start work to an accepted change, review rounds, or rework. Ask participants to note where assistance helped and where it added work.

Compare observations with the team’s experience on similar tasks. Decide in advance what would justify further evaluation, considering whether the changes meet requirements and whether the review effort is acceptable. If the trial raises concerns, inspect the examples and decide whether to adjust, stop, or try a different approach.

Evaluate additional task types separately rather than assuming that experience with one task applies to every project. Adoption remains a local decision based on the team’s requirements and observations.

The route from whiteboard to working software depends on the team’s understanding of the problem, its implementation choices, and its assessment of whether the result meets the intended behavior. AI assistance can help draft along the way; the team remains responsible for its decisions and the finished work.

Editorial image of a team collaborating around a whiteboard and planning softwareEditorial image of a team collaborating around a whiteboard and planning software