Mechanisms Outlive Intentions

The Core Insight

On June 9, 2004, Colin Bryar sent one line to the S-Team at Bezos's direction. PowerPoint was banned from that day. The last fifty S-Team decks averaged 440 characters a page. A Word page in eleven point Arial holds 3,000 to 4,000 characters.

Working Backwards is the process manual, written by Bryar and Bill Carr after 27 years inside Amazon between them. Carr ran digital music, then Prime Video and Amazon Studios. Invent and Wander collects the shareholder letters from 1997 to 2019, which are the doctrine those processes carry out.

Most founders treat culture as a hiring problem. Write the values down, hire people who hold them, and repeat the phrases until they stick. The Amazon claim is colder. A repeated process that enforces a principle produces the behavior whether or not the room believes in it. The house saying is that mechanisms work and good intentions do not.

Scale forced the writing. Amazon had about 600 employees in 1997, 9,000 in 2000, and 100,000 by 2013. Day 1 is the posture underneath, and Day 2 is stasis, then irrelevance, then decline. Bezos gives four defenses against Day 2: customer obsession, a skeptical view of proxies, quick adoption of external trends, and high-velocity decision making.

The Framework

The letters carry the beliefs and the manual carries the machinery. The letters argue that earnings do not turn into cash by themselves. His example is a transportation machine costing 160 million dollars. Grown at 100 percent a year for four years, it earns 150 million dollars and produces negative 530 million dollars of free cash flow. The paradox he draws is that the slower this business grows, the better off it is.

Six mechanisms carry most of the operating load.

• The annual plan runs bottom up as a written narrative, OP1, which becomes the plan of record as OP2 in January.
• S-Team goals pull the most important input metrics from each plan, and about three quarters are expected to land.
• A trained Bar Raiser leads the hiring debrief and judges the bar.
• Every meeting opens with a six-page narrative read in silence, and a PR/FAQ defines a product before anyone builds it.
• One single-threaded leader owns one initiative, on a team separable enough to build and deploy without permission.
• The Weekly Business Review charts controllable inputs, ordered along the customer experience rather than the org chart.

A mechanism, in the manual's definition, is a repeated process that enforces a principle. In 2004 the head of HR asked Robin Andrulevich to codify the Leadership Principles. She expected two months and took nine, and the list now holds 14. Base salary caps at 160,000 dollars, no bonus system exists, and raises arrive as stock vesting 18 to 24 months out.

Key Ideas

The Memo Replaces The Deck

Bezos read Edward Tufte's essay on the cognitive style of PowerPoint. Tufte argues that the bullet list does more damage as analysis turns causal and comparative. The rule that followed is six pages maximum, no formatting tricks, and appendices nobody reads in the room.

The document goes out at the start of the meeting and never before. Everyone reads in silence for about twenty minutes. Reading runs at three minutes a page, so six pages fit a sixty minute meeting and leave two thirds of it for discussion. Someone other than the presenter takes the notes, and silence counts as agreement. Bezos read on one assumption: every sentence is wrong until it proves otherwise.

The 2017 letter names why the format is hard to run. High standards are teachable and domain specific, and they need two things: knowing what good looks like, and knowing how much work it takes. Most people believe they can master a handstand in two weeks, and a coach puts it at six months of daily practice. Memo writers make the same error, budgeting two days for a document that takes a week.

The Press Release Is Written Before The Work

Working backwards means defining the customer experience first and iterating back until the team knows what to build. The press release stays under one page and the FAQ runs five pages or less. Ten or more drafts and five or more leadership meetings is the normal cost of one.

The release names the product as the customer understands it and states the benefit in one subheading. Then come the launch date, the problem, the solution, and quotes. The internal FAQ holds what founders skip. Customer need and market size, per unit economics with a pro forma profit and loss, dependencies, and feasibility.

The reviewer tests every draft with two words, so what. If the product is not clearly faster, easier, or cheaper, the answer is to stop. The fictional example is a smart mailbox at 299 dollars against a 250 dollar bill of materials. That leaves 49 dollars a unit against a team of 77 people costing 15 million dollars a year. Most PR/FAQs never become products, and the manual calls that the point.

One Leader Owns One Thread

Between 1997 and 2001 Amazon revenue grew from 148 million dollars to over 3.1 billion. A dependency is something one team needs and cannot supply for itself. The monolith called Obidos ran on one database, and every change queued behind a three person cabal.

Bezos treated cross-team communication as a defect to remove. Teams talk machine to machine through documented interfaces, and the monolith was rebuilt service by service while the business kept running. New Project Initiatives failed first, running a quarterly global prioritization with scoping meetings of 30 or 40 people.

The two-pizza team came next. No more than ten people, owning design, technology and results, approved by the S-Team in advance. Fitness functions, which scored a team on weighted metrics, broke down after about a year. Amazon went back to input metrics with a goal on each. Size was not the predictor either. The predictor was a leader with the right skills, authority and experience, running a team that worked on nothing else.

Fulfillment by Amazon sat at coming soon for over a year. Jeff Wilke freed Tom Taylor of everything else in 2005, and it launched in September 2006. Dave Limp names the failure mode. The surest way to fail at inventing something is to make it somebody's part-time job.

The Bar Raiser Outranks The Hiring Manager

Amazon started the program in 1999 with twenty Bar Raisers, copying a Microsoft role whose final interviewer carried no penalty for an unfilled job. The loop runs five to seven hours with five to seven interviewers, and no future direct report of the candidate takes part.

Feedback is where the mechanism bites. Notes go in near verbatim within about fifteen minutes. The vote is one of four, from strongly inclined to hire down to strongly not inclined. Undecided is not on the list, and nobody sees another interviewer's feedback before submitting.

The Bar Raiser leads the debrief, reads every note into the room, then asks who wants to change a vote. Across interviewers who ran some 4,000 interviews over 15 years, the authors found three uses of the veto. The role carries no extra pay, only an icon next to your name in the directory.

The 1998 letter set three questions for every hire. Will you admire this person, will they raise the average effectiveness of the group they enter, and where can they be a superstar?

Manage The Inputs And Report The Outputs

Output metrics are orders, revenue, profit, and free cash flow per share. Controllable input metrics are the activities that produce them: selection, price, and convenience. Amazon first measured selection as the number of new detail pages. Teams added thousands of low demand items, holding costs went up, and sales did not.

The metric moved to detail page views, then to the share of views where the item was in stock. The last version added ready for two-day shipping, named Fast Track In Stock, with a 95 percent standard. Order matters in the sequence around it. Define, measure, analyze the root cause, improve, then control. Jumping to improve is the classic failure.

The Weekly Business Review carries the charts. Every metric runs trailing six weeks and trailing twelve months side by side on one axis. Growth rates sit on a secondary axis, which in one example shows growth down 67 percent since January behind a healthy looking chart. Nobody runs the meeting. Finance certifies the data and holds no stake beyond calling it as they see it.

The 2009 letter shows it in the goal sheet. Of 452 detailed goals for 2010, 360 touch customer experience directly. The word revenue appears eight times, free cash flow four times, and net income, gross profit, margin and operating profit not once.

Reversible Decisions Get Speed

The 2015 letter splits decisions in two. One-way doors are irreversible or close to it, and they earn a slow process. Two-way doors are reversible, and high judgment individuals or small groups take them fast. The failure mode is a large company running the heavy one-way process on everything. Companies that run the light process on one-way decisions go extinct before they get large.

Speed has a number attached. Most decisions get made at about 70 percent of the information you want, and waiting for 90 percent is slow. That works alongside fast correction, since being wrong costs less than being slow when you course correct well.

Disagree and commit closes the gap between a senior view and a team view. Bezos greenlit an Amazon Studios show he doubted by disagreeing and committing out loud, and he counts the cycle time saved. The senior person yields because the team holds more ground truth. Real misalignment gets escalated immediately, because the default dispute resolution mechanism in a large company is exhaustion.

The Fire Phone shows what a one-way door costs. Over a thousand people worked on it, and it launched in June 2014 at 200 dollars with a two-year AT&T commitment. Amazon discontinued it in August 2015. Bezos co-authored its PR/FAQ, and his verdict is that no process makes the decision for you.

Practical Applications

Ban slides for any decision carrying causal or comparative analysis. Write six pages of prose, hand them out at the start, and read in silence for twenty minutes. Budget three minutes a page.

Write the press release before the work starts. Keep it under a page and answer four internal questions in five pages or less: the market, the per unit economics, the dependencies, and feasibility. Set a not-to-exceed price so trade-offs land before engineering does.

Name one input metric per goal and one owner per metric. Chart trailing six weeks against trailing twelve months on one axis, and pair every number with an anecdote and an exception report.

Sort the next five decisions into one-way and two-way doors before discussing any of them. Take the two-way ones this week at about 70 percent of the information you want. Give the one-way ones the slow process. Put someone outside the hiring team in charge of your next debrief, and require written votes before anyone talks.

Who This Is For

Founders who have stopped being in every room get the most from Working Backwards. The mechanisms answer problems that appear once decisions happen in rooms they never enter. Invent and Wander rewards anyone deciding how fast to decide. Skip the manual if you are three people at one table, and take the PR/FAQ anyway, since it costs one document.

Both books are written from inside and by beneficiaries. Bryar and Carr spent their careers there, and the equity they hold came from the outcome they explain. The letters are investor communication from a founder raising money and defending a valuation, not a neutral record. The manual is honest in places. It records that fitness functions failed and that two-pizza teams worked only in product development. Prime itself ran without a single-threaded leader until late.

Scale is the real limit. A 95 percent in-stock standard, a corps of trained Bar Raisers, and four hour weekly reviews assume a company of hundreds of thousands. Read the mechanisms and rebuild them at your size.

The Decision

Take the next document you plan to present. Convert it to prose with no bullets, cap it at six pages, and hand it out in the room instead of sending it ahead. Then time the silence. A room that cannot read for twenty minutes has a meeting problem the format will not fix.

The next build gives you the second test. Write the press release for the thing you are about to start, keep it under one page, and read it to the team. If nobody can say what changes for the customer, you have the answer before you spend the engineering time finding it out.