Tahir Shahzad Product Manager & Community Builder
Book a Free 30-Min Review

AI Brought Back the Gap Between Requirements and Shipped Code

Tahir Shahzad By Tahir Shahzad 8 hours ago 7 min read
AI Brought Back the Gap Between Requirements and Shipped Code

The requirements gap is the distance between what a customer asked for and what the team shipped. We have spent decades trying to close it. AI did not create a new gap. It removed the delay that used to hide the old one. This post is for product managers, founders and engineering leads who ship with AI and wonder why things still go wrong.

You have seen the memes. The customer asks for a tree swing, and the team delivers something built from the same parts that does not swing at all. We answered with frameworks, methodologies and processes. Many of us added workshops and certifications on top. The real world still brings a new problem every few weeks. Now AI models have reopened the same gap, at a much higher speed.

A small, familiar failure

An essay on Hacker News this week describes a failure most product people will recognize. A teammate used AI to add install logging for the marketing team. The code worked. It also made a welcome screen take about 15 seconds to load, because each fallback quietly made several API calls.

Nobody was careless. The AI built exactly what the requirement said. What was missing was someone who would push back on the requirement itself.

What the research says about the gap

Fred Brooks wrote in 1987, in his essay “No Silver Bullet”, that the hardest single part of building software is deciding precisely what to build. Later research agrees with him.

  • Standish Group, CHAOS report (1995): lack of user input and incomplete requirements were the top two factors behind challenged projects.
  • PMI, Pulse of the Profession (2014): 47 percent of unsuccessful projects failed to meet goals because of inaccurate requirements management. PMI also estimated that organizations waste $51 million for every $1 billion spent on projects because of poor requirements management.

Both sources are old, and researchers advise reading the exact CHAOS percentages with care. The pattern is what matters. Across three decades and different methods, the problem sits in requirements and communication, not in typing code.

The frameworks we built to close it

Each major approach attacks a different part of the problem. The table shows what each one was designed to fix and where AI puts pressure on it. The second column is my own reading from delivery work, not research.

Row of blank index cards and sticky notes laid out on a desk in sequence
Row of blank index cards and sticky notes laid out on a desk in sequence
FrameworkWhat it fixesWhere AI strains it
Scrum (Schwaber and Sutherland)Short feedback loops and a Product OwnerSprints were sized around slow builds
Kanban (David J. Anderson)Visible flow and WIP limitsBuilds speed up, so review and decision queues become the bottleneck
User Story Mapping (Jeff Patton)Shared understanding of the user journeyAI writes stories fast, but not the shared understanding
Continuous Discovery (Teresa Torres)Weekly customer contact and assumption testsOutput looks finished, so teams skip testing the assumption
Behavior-Driven Development (Given, When, Then)Requirements as examples both sides can checkExamples rarely include performance or cost limits unless you add them
Definition of DoneAn agreed meaning of “finished”AI meets the letter of the criteria, not the intent
Working Backwards (Amazon)A press release and FAQ before buildingFits well, because it forces the cost question early

Most of these help a team agree on what to build and whether it matters. Very few name a person who checks what a requirement will cost once it is built. That was tolerable when building was slow, because the slowness gave someone time to notice.

What changes when AI writes the code

Typing code used to be slow. That slowness had a side benefit: while a developer wrote it, they had time to notice that something looked expensive, odd or risky. AI removes that pause.

The DORA research from Google shows both sides. In the 2025 report, AI use among developers rose from 76 percent to 90 percent, and more than 80 percent said it raised their productivity. The same report found that AI adoption still correlates with higher instability: more change failures, more rework and longer recovery. A year earlier, DORA had estimated a 7.2 percent drop in delivery stability for every 25 percent increase in AI adoption. DORA’s own summary is that AI is an amplifier. I wrote about the same effect in AI Made Shipping Fast. It Made Deciding Slower.

A team with clear decisions gets faster. A team with fuzzy decisions ships its fuzziness faster.

Where the question gets lost

In a team of 10 to 30 people with no product leader, the person who challenges a requirement is rarely named. Product writes the ask. Engineering builds it. AI now builds it faster. The question “what does this cost us?” falls into the gap between the two.

In the install logging example, marketing needed data and engineering delivered data. Nobody owned the question of what that data would cost the first-run experience. The Definition of Done was met. The gap was still there.

I think about this whenever I hear “vibe coding”. I have done hands-on delivery work, learned from workshops and earned certifications, and the real world still surprises me every few weeks. A tool that builds exactly what it is told does not remove that surprise. It makes the surprise ship sooner. I am curious how vibe coders will handle the part that was never about code. Knowing what to ask is a skill of its own, and discovery is not a prompt.

The fix is a written answer, not a new title

Hiring a “chief requirement skeptic” is not the answer, and many small teams cannot afford a product leader yet (I cover that choice in when to hire a fractional product manager). The fix is a clear answer to one question: who is allowed to say no to a requirement, and what context do they need to do it well? Write that down for each area of the product.

Whiteboard grid with one sticky note in each of four boxes
Whiteboard grid with one sticky note in each of four boxes

For each area (onboarding, billing, analytics, integrations), the page should say three things:

  1. Who can challenge a requirement. One name or role, not a committee.
  2. What they see before deciding. The performance budget, the cost of added calls, the effect on the first-run experience, and who asked and why.
  3. What “no” looks like. A smaller scope, a different approach, or a question sent back to the requester.

You can borrow from the frameworks above. From Working Backwards, state the customer benefit and the cost in one paragraph. From BDD, add examples for limits, such as “the welcome screen loads in under two seconds”. From the Definition of Done, add a performance budget, not just “works as described”. From Kanban, limit work in review so decisions do not pile up behind fast builds.

This takes about an hour of writing. It works the same whether a person or a model writes the code.

Who should not do this

  • Very junior engineers need guidance before they get decision rights. Without it, autonomy feels like being left alone.
  • Very small teams where two people already talk all day may not need a written map. A habit of asking “what does this cost us?” can be enough.
  • Regulated areas such as payments or health data often need a named owner and a sign-off, not a flexible “anyone can say no”.
  • Teams still searching for product-market fit should spend the effort on discovery first. A clear “no” depends on knowing what the product is for.

Name the person who can say no. Give them the context to do it well. Then let the AI build what survives that question.

If your team is shipping faster than it is deciding, get in touch and we can talk it through.

Tahir Shahzad
About the author

Tahir Shahzad

Tahir Shahzad is a Product Manager, Product Owner, and technology consultant with over a decade of experience helping startups and organizations build products people actually use. If you are working on a product problem worth solving, reach him at tahirshahzad.com/contact/

Let's find your real bottleneck in 30 minutes

Book a free 30-minute product review. You'll leave with a clear read on what's blocking delivery and what to tackle first, whether or not we end up working together.

Book a Free 30-Min Review 30 minutes. No pitch, no obligation.