Leading in the AI Era Module 3 · Lesson 3.1 · Recognizing organizational barriers

Where the gains go

Updated

Beginner

What you'll learn

~14 min
  • Name four places a useful change waits, and the purpose each control serves
  • Compare the cost of a decision that waits with the cost of a rollout that moves ahead of its evidence, in your own calendar days
  • Record "we cannot estimate this until we measure the waiting time" as a finding, with an owner and a date

You leave this lesson with: a record of waiting days against building days for your team’s last three changes, and a two-column note, “cost of waiting” and “cost of proceeding without evidence”, each in days or marked “not yet measured”.

Four controls, each with a reason

A faster draft of a program, a report or an analysis only helps if the change can then get through access, review and approval. Between “this works” and “people use it”, a change can wait in four common places.

ControlThe purpose it servesThe cost when nobody owns it
A request queue for access and environments (test systems, accounts, data access)Keeps track of who can reach which systems and dataA request with no named owner waits without anyone being late
Sign-offs in sequence (security, privacy, legal, budget)Each reviewer sees a risk the others do notEach step waits for the one before it, and no step has a time limit
Sandbox (a separate test area that cannot touch live systems) access restricted by defaultProtects sensitive data and limits connections, credentials and spendingThe restriction also covers work that touches no sensitive data, with no one deciding when it can relax
Decisions escalated to a senior levelKeeps accountability with someone who can answer for itThe person who knows the work is far from the person who decides

The reviewers in this table are doing their jobs. The question for a leader is narrower: for each control, who owns the queue, and how many calendar days is a request allowed to wait?

Unowned queues and serial sign-offs

A queue is unowned when a request enters and nobody is responsible for how long it stays. Nothing is refused. Nothing is approved. The request is simply open.

Serial sign-offs add to this. Four reviews in sequence, each reasonable, each without a service level (a stated time within which the reviewer answers), mean the change waits the sum of all four. A reviewer who is busy is not the problem. A reviewer with no stated time limit and no deputy is a gap in the process, and a leader can close it.

Withheld sandbox access and distant decisions

A sandbox is a separate environment where a team can try something without touching live systems. Restrictions protect data, and they also limit connections, credentials and spending. Public or invented data may justify a lighter review, but the reviewer must still check what the tool can reach and change.

The practical test covers both the data and the tool’s reach: what sensitive information, systems, credentials and spending could this trial affect? The reviewer decides whether those boundaries justify a lighter review.

Ask your team whether work is moving outside approved systems while a request waits, and what approved access would meet the need.

Decisions made far from the work follow the same pattern. A senior approver sees the whole portfolio. A team lead sees the task. When every question climbs, the approver decides with less detail than the team had.

Measuring your own queue

Record the calendar days your own team’s last three changes spent waiting for access, review or sign-off, against the days spent building. Three changes give an initial local comparison; further requests confirm whether the pattern holds.

If waiting time is small, investigate another constraint. If it is large, identify where the wait occurred and who owns the request. Agree a response time with that owner using the team’s recorded waits, the work required and available review capacity. Revisit it when local evidence shows whether it is workable.

💬Decision: sandbox request stuck open

A state agency’s forms team has built a prototype that rewrites public help-page text in plainer language. Its request for a test environment has been open since the start of the quarter, in a queue that no one is assigned to watch, and the only data involved is the agency’s published web pages.

  1. Wait for the queue to reach the request.
  2. Name an owner for the queue, ask for a decision on the request within an agreed number of days, and state that the data class is public.
  3. Tell the team to run the prototype on personal accounts so the work continues.

Pick one before you open What happens.

ℹWhat happensclick to expand
  1. Wait for the queue: It keeps the process intact and the work idle; the waiting days grow, and nobody is accountable for them.
  2. Name an owner for the queue: It keeps the control and gives it an owner and a time limit; the security reviewer still decides, with the data class stated up front.
  3. Run it on personal accounts: It keeps the work moving this week, but it moves agency work outside the control the queue exists to enforce, and the next piece of work may not be public text.

Choice 2 holds up, because it keeps the purpose of the control and removes only the open-ended wait.

Two costs: waiting, and moving ahead of the evidence

A decision that nobody owns has a cost: the useful work sits idle, and the people who built it stop expecting an answer. The opposite also has a cost: a tool rolled out to a whole department before anyone has shown it works on the department’s own tasks. Careful reviewers exist to prevent the second kind. Neither cost is larger by default. The comparison has to be made for the decision in front of you.

Stability is the other risk. DORA (Google’s research program on software delivery, September 2025, a survey of nearly 5,000 technology professionals) linked AI adoption with more software delivered and better product performance, but also less reliable delivery. It does not tell us what caused those differences or what your team should expect. Measure the complete workflow locally before crediting AI with either.

The cost of waiting is countable too, once someone counts it, in the calendar that applies to you. In a university research group, a postdoctoral researcher’s funding period is fixed, so each waiting week is a week of the project that cannot be recovered. In a company, it may be a release window. In an agency, it may be a reporting date. Name that calendar and count in its units.

“We cannot estimate this yet” is a finding

Sometimes the honest answer to “what does the delay cost?” is that nobody has recorded the waiting time. A leader who hears that should not push for a number. A guessed figure can be repeated, planned around and defended.

Accept the finding, assign the measurement, and give it a date. That is a decision with an owner. It is also the point at which the formal question of what a system’s use, and its non-use, will cost belongs to the assurance team. That team checks that a system is safe and works as claimed before release, and AI Assurance (Lesson 5.1, a separate training in this series) shows how that assessment is written.

💬Decision: lab asks for wide rollout

A university research group’s data analyst has used an AI tool to clean survey files in a limited pilot. The department head is asked to approve its use for all projects in the lab this month, and no one has recorded how long the pilot’s files waited for data-use review or how many corrections reviewers made to its output.

  1. Keep the tool within the pilot’s approved scope, name the accountable reviewer, and measure corrections and waiting days before a dated decision on wider use.
  2. Approve use for all projects now, since the pilot went well and every month of waiting holds up the lab’s other projects.
  3. Pause all use of the tool until the department updates its general data-use policy.

Pick one before you open What happens.

ℹWhat happensclick to expand
  1. Keep the tool within the pilot’s scope: It keeps wider use on hold while the reviewer obtains the missing evidence within the approved scope.
  2. Approve use for all projects now: It takes the cost of waiting seriously, and the pilot may well be sound, but it expands a result nobody has measured.
  3. Pause all use of the tool: It protects the data while the policy is settled, but it pauses the work without a named reviewer or a decision date, so the question stays open with no one accountable for it.

Choice 1 holds up, because it waits for the missing evidence and puts a date on the decision.

KNOWLEDGE CHECK

A director asks what the delay on a pending tool request is costing, and the team has never recorded waiting time. Which response is the most useful?

Optional deeper reading, for federal programs: Secure Software Delivery for Federal Environments leadership brief, one page, a separate training in this series.

Key takeaway

LEADERSHIP DECISION
Name an owner and a time limit for each queue that stands between a pilot and its first users.
IMMEDIATE ACTION
Ask the team for the waiting days and the building days of its last three changes, the queue where the waiting happened, and both costs for the decision now pending.
SUCCESS MEASURE
Calendar days from request to decision for the next access request, against the waits recorded for the last three changes.
Search lessons