Why your SOPs get ignored by week two

Every operator has lived this. You spend a weekend writing it, walk the team through it on Monday, and two weeks later the old behavior is back. The document is still on the wall, unread and slightly greasy.

Most operating problems are not information problems. People usually know what they are supposed to do. What is missing is a system that makes the right behavior automatic when nobody is standing there.

Why does week two specifically kill it?

Because week one runs on attention and week two runs on design.

In week one you are present. You mention it at preshift. Somebody asks a question and you answer it. Compliance in week one measures your presence, not the procedure.

Week two, you are dealing with something else. A call out, a vendor, a broken piece of equipment, your actual life. The procedure now has to hold with no one advocating for it, and most procedures were never built to do that.

Then the first real conflict arrives. The room gets busy and the correct step costs thirty seconds that nobody has. Someone skips it, the night is fine, and nothing bad happens. That is the moment the SOP dies, and it dies with a positive outcome attached, which makes it very hard to argue with later.

A procedure that has been skipped once without consequence has been unofficially repealed.

This is not a restaurant phenomenon, which is worth knowing because it stops you taking it personally. Coval, a company that runs evaluation tooling for automated voice systems and processes roughly 3.1 million evaluation metrics a week, reports that about 95% of those systems work in a demo and about 62% survive their first week live. That is vendor data from a company selling the fix, so hold it loosely. But the shape is the same shape: the demo is week one, and week two is production.

What does a new procedure cost to run?

More than most owners price it at, which is exactly why they launch four at once.

Here is the arithmetic. Say the new procedure adds four minutes per person per shift. Nine people on the floor and in the kitchen, two shifts a day, six days a week. That is 432 minutes a week, or 7.2 labor hours.

At a loaded hourly cost of $22, meaning wages plus payroll taxes plus whatever benefits you carry, that is $158 a week and $8,216 a year.

On a restaurant doing $1.2M, $8,216 is about two thirds of a point of labor.

Now put that beside the benchmark. The National Restaurant Association's Restaurant Operations Data Abstract, 2025 edition, built on financial and operating data from more than 900 restaurant operators nationwide, puts full service salaries and wages including benefits at a median of 36.5% of sales for 2024. Operators reporting a pre-tax profit ran 34.2%. Operators reporting a loss ran 42.9%.

The whole distance between profitable and losing is 8.7 points. Which means your building can carry a handful of four minute procedures and cannot carry twenty of them, and every one you add has to buy back more than it costs.

That is the honest reason to launch one at a time. Not tidiness. Arithmetic.

Two implications follow immediately. Procedures that remove work are free and procedures that add work are not. A par sheet filled out in advance takes time off the line. A second signature adds it. And a procedure that pays for itself needs to be aimed at something expensive, which in most restaurants means food safety, portioning, stockouts, cash, or the close.

What separates a procedure that holds from one that does not?

Five properties. A procedure with all five usually survives. One missing any two of them usually does not.

It changes the physical work, not just the instruction. A portion scoop instead of a written ounce. A labeled bin instead of a labeling rule. A form that will not submit without a number. Physical beats verbal every time.

It is owned by a position. Not the team, not everyone, not whoever is closing. The closing line cook. Positions persist through turnover in a way that names do not.

It produces evidence. A number, a photo, a log entry with a value in it. Checkmarks verify nothing. If completing the step leaves no trace, you have no idea whether it happened, and neither does the person who skipped it.

It is short enough to be done in the time available. This is the most common failure and the least discussed. A procedure that cannot be completed inside the scheduled labor is a procedure the crew has correctly identified as optional.

Somebody looks at the output within a week. Not to punish. To look. A log that nobody ever reads teaches the crew exactly what it is worth, and they learn that faster than they learn the procedure.

That last one is where good operators lose most often. The system gets built and then never gets read, which is the same failure mode as a control that looks fine on paper and does not hold under a Friday night rush.

How do you introduce one so it survives?

One at a time, small, and with the reason attached.

One procedure at a time. Three new procedures in a month means none of them get the attention that week one requires. It also means when things improve you cannot tell which one worked.

Start with the smallest version. Do not launch a full inventory program. Launch a nightly count of eight items. Do not launch a full waste tracking system. Put a clipboard on the bin and define who writes on it, which is all waste tracking without a full inventory program actually requires to start producing useful information.

Say what it is for. People follow procedures whose purpose they understand and abandon ones that feel like paperwork. "This is so we stop running out of the item that generates our best margin" holds. "Because it is policy" does not.

Have the person doing the work help design it. Not as a courtesy. Because they know where the thirty seconds is going to come from and you do not.

Then check it in week two, deliberately. Put it in your calendar the day you launch it. Week two is when the procedure needs you, and it is exactly when you will have forgotten about it.

How do you tell whether it is working?

Carefully, because a restaurant produces very little data very slowly, and it is easy to fool yourself.

One bad shift after you introduce a procedure is not evidence the procedure failed. One good week is not evidence it worked. A restaurant generates a small number of shifts per month, and normal variation between them is large enough to tell any story you want to hear. That is the same reason small sample sizes will fool you if you let them in anything you measure.

Borrow the arithmetic rather than the folk number, because there is a real one. What a comparison can detect depends on how much the thing you are counting varies and how big a change you want to catch, not on a threshold somebody quoted. At 95% confidence and 80% power the constant is 2 × (1.96 + 0.84)², which is 15.68, call it 16. For counts, the smallest relative change you can see is roughly the square root of 16 divided by the events in each period.

Run it on stockouts. Six a week across a four week window is 24 events, and the square root of 16 over 24 is 0.82. You could only detect a change of about 82%derived, which means near elimination or nothing. Detecting a 25% reduction would need 256 events (derived), about forty-three weeks (derived) at six a week. A restaurant running two services a day, six days a week, produces about fifty shifts a month. You are not going to prove anything about a four minute procedure from twelve of them.

That is also why quoted numbers deserve checking. The 10,000 monthly visitor floor for a valid split test, which gets recycled into operations advice, is not a study. It appears as reference [29] in Kohavi, Deng, Longbotham and Xu's KDD 2014 paper, resolving to a QuickSprout blog post of 14 January 2013, and the four authors qualify it on the spot: "the guidance should be refined to the metrics of interest." Small samples cannot separate a real effect from ordinary variation. That part was always true. The number was never the point.

So do what operators can actually do instead of pretending to run an experiment.

Watch the leading indicator the procedure controls rather than the outcome you hope for. If the procedure is a nightly count, watch stockouts, not food cost. Give it four weeks before judging. And compare like weeks, because volume distorts the read: a procedure introduced during a slow month looks excellent and then breaks in the first rush. Staffing to a forecast instead of to a feeling gives you that comparison.

Then accept a lower standard of proof than you would like, and be explicit that you are doing it. "Stockouts went from six a week to one and the crew says the count is easy" is not a controlled result. It is enough to keep the procedure.

Where does procedure design stop being the answer?

Three places, and each one wastes a month if you miss it.

When the work does not fit in the hours. No design survives a shift that is understaffed for the volume in the room. The crew will drop something, and they will drop the thing with no immediate consequence, which is always the procedure. That is a scheduling decision wearing a compliance costume.

When the procedure is aimed at something cheap. Given that every procedure has a running cost, one aimed at an item that never breaks and never costs anything is a net loss even at full compliance. Price it before you launch it.

When conditions have moved under it. The Association estimates total expenses for an average restaurant jumped 36% between 2019 and 2026 and reports that 42% of operators said their restaurant was not profitable in 2025. A procedure written against 2019 costs, portions and prices can be followed perfectly and still be pointed at the wrong thing. Reread your own documents once a year and delete the ones that are answering a question you no longer have.

What to do this week

Pick the procedure you care most about and read it as though you were the person who has to do it during a rush. Time it. Most of them cost more time than the author assumed.

Then price it. Minutes, times people, times shifts, times your loaded hourly cost, times fifty two. Write the annual number at the top of the document.

Cut it until it fits in the time that actually exists.

Add one piece of evidence to the most important step. A number, a weight, a photo. Something that will still be true tomorrow.

Assign it to a position, write the position on the document, and put a note in your calendar for eleven days out to read whatever it produced.

Questions about a procedure that will not stick?

Email me at eric@seod.com with the one SOP that keeps failing, pasted in or photographed off the wall. I will tell you which of the five properties it is missing and the smallest version I would relaunch it as.

Sixteen years of multi unit restaurant operations and more than $54M in annual P&L means I have written a lot of procedures that died in week two, which is where most of this came from.

There is more on running the operation if you prefer to build it yourself.