メインコンテンツへ移動 / Skip to main content

How Big Does a Company Have to Be to Get Slow?

Why can a feature spend a year in the queue? Amazon, Google, Toyota and Japan Airlines offer different ways to connect authority, evidence and action.

Folders waiting for approval on a vacant desk in a porcelain office
Technology
Published on: November 16, 2026
Read time: 10 min
Author: Pochang Lab
Read time: 10 min

A company has a bureaucracy problem when people see a customer’s needs change, yet decisions repeatedly arrive too late because authority and the means to test and revise those decisions sit somewhere else. This is a working definition. Imagine a company of 3,000 that fits it and a division inside a company of 30,000 that does not. The numbers describe a thought experiment. To find out where your organization stands, trace one feature that finally reached a customer.

The feature we asked for last year shipped today

Imagine a billing clerk copying line items from a SaaS product and rearranging spreadsheet columns every month. Sales records a request for a download in the right order. The feature launches a year later. By then the customer has changed accounting systems and needs an integration. Engineering kept its promise. The account manager knew about the migration. Who could have retired the request, and when?

A year spent reconciling historical invoices during a migration can be time well spent. A quarter when nobody touches a ticket and nobody checks whether the customer still needs it is a different expense. Both projects can generate exactly one line in the release notes.

Mark four dates: receipt of the request, the decision to pursue it, start of implementation and first actual use. Ask the people involved what they learned between those dates and who could have changed the plan. A ticket marked “in development” can cover both a day writing code and a month waiting for a reply.

How much of the year can an AI agent save?

Suppose implementation takes ten working days and planning and approval queues consume another ninety. The stages are sequential. Halving implementation cuts elapsed time from 100 days to 95, a 5% improvement. If the review board meets on a fixed date, finishing on Monday instead of Thursday may leave the launch date unchanged. These are illustrative assumptions, not measured gains.

Even the assumed coding gain deserves measurement. In a randomized study of early-2025 tools, METR assigned 246 tasks to 16 experienced developers working in familiar projects. Allowing AI increased average completion time by 19%. The finding concerns a particular period, population and kind of work. In a follow-up, METR described how reluctance to work without AI and task selection complicate estimates for newer tools. Compare similar changes in your own setting and count review and revision time. [1][2]

Google’s DORA research program examined external approval boards and alternatives involving informed peer reviewers and automated checks. It reported poorer delivery performance with heavyweight approval and found no evidence that such procedures reduced change failure rates. An observational finding cannot tell you which review to remove. Look at the last review that caught a defect. Who spotted it, and when could that person first have seen the change? [3]

When Amazon and Google send a decision upstairs

In Amazon’s 2016 shareholder letter, Jeff Bezos advocates lightweight procedures for reversible decisions. He also recommends escalating genuine disagreement early. Suppose finance prioritizes final invoice accuracy while sales promises delivery the morning after close. When neither team can change those commitments, another meeting merely tests which manager is more persuasive. The letter records a management principle. It supplies no release-speed measurements for Amazon’s individual services. [4]

Google’s SRE Workbook gives an example of an error budget policy. Teams agree in advance how much unreliability a service can tolerate. In the example, exhausting the budget over the previous four weeks freezes changes with exceptions including the highest-priority problems and security fixes. Operations gains a way to stop releases, and developers can see what it takes to resume them. [5]

For the billing product, a team might trial a reversible column order using existing permissions. Adding personal information to the download changes the decision: once customers distribute copies, rolling back the software cannot retrieve them. Delegation becomes usable when the team knows the eligible customers and data, the stop conditions and who settles a dispute.

What happens to the person who spots a problem?

A porcelain worker pulls a green call cord and the team leader turns toward the line

Toyota describes a worker pulling a cord to light an andon and summon the person in charge when quality or timing goes wrong. One worker has a means to signal. Someone else has a job to do when the signal appears. That pairing allows intervention before a fault travels downstream. A suggestion discussed at month end arrives on a different clock. [6]

In Japanese workplaces, informal consultation before a meeting, often called nemawashi, can bring an absent customer’s knowledge into a decision. It can also keep a proposal off the agenda until everyone has privately signaled assent. In the latter case neither disagreement nor delay appears in the minutes. Ask for one recent example. Did the consultation change a requirement? Or did the proposer simply establish that everyone would vote yes? Nationality cannot answer that question.

Amy Edmondson studied 51 teams within one US manufacturing company. A shared sense that interpersonal risk-taking was safe was associated with learning behaviors such as asking questions and discussing errors. Teams differed within the same company. A manager can examine the last time a junior colleague asked to pause a release. Who replied, and how quickly? If the only entry in the performance record is the delay, consider what the next colleague will report. [7]

What can a software business learn from Japan Airlines?

Japan Airlines applied for corporate rehabilitation in 2010. The Enterprise Turnaround Initiative Corporation of Japan invested ¥350 billion. Creditors waived claims, shareholders faced a complete capital reduction and the company withdrew from unprofitable routes and cut jobs. JAL also introduced a corporate philosophy and a divisional profitability system. Structural costs contributed to its problems. Several interventions happened together, at considerable cost to employees and other stakeholders. These records cannot isolate the effect of philosophy or divisional accounting. [8][9]

JAL described an ongoing intention to examine profitability route by route. What is the corresponding unit of observation in SaaS? Total company revenue tells the billing team little about its export. How many eligible customers used it again the following month? Did correction requests rise? How much manual work remained for support? Return these figures to the people authorized to decide what happens next. They can then review the consequences of their previous decision.

A divisional metric can reward moving work across a boundary. If engineering closes the ticket and support rebuilds the file every month, the customer’s task is still unfinished. Follow the work until the customer completes the close.

Which clock makes a company look fast?

Imagine one company announces ten changes every week and another announces one a month. Before ranking them, ask what is in the denominator. One may bundle updates across ten products while the other serves a single accounting workflow. Ten revisions to pricing copy and one change that lets a clerk finish the monthly close in a single operation also return different amounts of time to the customer. Release notes are a starting point for questions.

Inside a business, compare changes serving similar customers with similar scope and risk. Time both implementation-to-production and request-to-first-use. A short first interval and a long second one suggest investigating selection before coding or activation after launch. If both intervals are long, a change might be too large to test in pieces. These are hypotheses to check against actual histories.

Adding many quick changes can improve the average while last year’s request still waits. Look at the oldest open request alongside the average. If a customer wants to wait until the next fiscal year, record a date and a contact. If three departments are each waiting for another to decide, appoint one decision maker. A deliberate pause and an invisible queue call for different responses.

Three ways delegation can go wrong

Authority without capacity. A team can choose the column order, but the only person able to modify the shared exporter may be busy closing another product’s books. Removing a signature leaves the engineering queue intact. If separating the code is expensive, reserve a block of that person’s time through the trial. Repeated dependencies can then justify an investment in independent changes.

Consultation that requires unanimous consent. Sales may know that fixed columns would hurt another customer. That information matters. Waiting for every salesperson to reply makes the customer’s timeline depend on travel and vacations. Consult the relevant account manager and specify when the team can proceed with its one consenting customer. An objection involving access rights triggers a pause and specialist review. Identify both whose knowledge matters and which conditions block the trial.

Counting cancellation as failure. Suppose sales learns that the customer has migrated and retires the request before coding starts. A target based solely on the number of promised features shipped records a miss. Keep a line about what changed the decision and which task received the freed time. There is no need to invent a monetary saving. When another customer asks for the feature, the team can check whether the original reason to stop still applies.

A founder delegating a product for the first time can encounter all three problems. In a larger business several teams may make commitments to the same customer. A trial that engineering can roll back becomes hard to reverse if sales has promised the output in a contract. Time spent agreeing on the scope with the account owner buys the option to withdraw. Skipping that conversation can put the rework on someone else’s schedule.

Change one queue in four weeks

The following is a proposed trial. Select one shipped change, one abandoned change and one still waiting. Sampling only successful releases hides the requests nobody ever picked up.

  • Week one. Bring sales, engineering and the reviewer to the same timeline. Find the longest gap. Did anyone know who was expected to move next? Exclude days spent on necessary checks from waiting time.
  • Week two. Delegate one class of change, such as a column-order trial with one consenting customer. Name the owner and peer reviewer, document rollback and identify a support contact. Bring in the appropriate specialist early for changes affecting permissions or invoice amounts.
  • Week three. Observe a real close. A successful download event cannot tell you whether the clerk still rearranges the columns. Ask why. If the customer has moved on, establish who can retire the work.
  • Week four. Decide whether to continue, revise or withdraw. Compare decision waits, time to first use, rework and support contacts against similar cases. Examine what the trial missed before widening it. With few cases, extend the observation window.

In a company of thirty, one founder’s inbox can hold the queue. In a company of thirty thousand, a customer-facing team can have room to try a change, stop it and retire it when circumstances shift. A customer who asked for something last year may no longer want it today. When the first employee hears that, what can they decide before the day is out, and with whom?

References

  1. [1]METR, randomized trial (2025) ↩
  2. [2]METR, experiment design update (2026) ↩
  3. [3]DORA, Streamlining change approval ↩
  4. [4]Amazon, 2016 letter to shareholders ↩
  5. [5]Google SRE Workbook, Example Error Budget Policy ↩
  6. [6]Toyota Production System ↩
  7. [7]Edmondson, Psychological Safety and Learning Behavior in Work Teams (1999) ↩
  8. [8]Japan Airlines, management message ↩
  9. [9]Japan Airlines, investigation report (2010) ↩