Thirty-five dollars. That is roughly where most independent restaurants land when they finally write down the amount a phone complaint can resolve without a manager, and the number matters far less than the fact of writing it. Until it exists, every complaint call is negotiated from scratch by whoever answered, under time pressure, with no way to be consistent.
The reason to write it now, before you evaluate any phone system, is that automation makes an unwritten policy visible. A voice agent will do exactly what the rule says, every time, at three in the afternoon and at nine on a Saturday. If there is no rule, the agent either gives away nothing or transfers every complaint, and both of those look like a product failure when they are actually a policy failure.
A ceiling is the first line of the policy
Pick a dollar figure that a complaint can resolve without approval. Anchor it to your average ticket rather than to a feeling. If your typical phone order is twenty-eight dollars, a ceiling somewhere around thirty-five covers the whole order plus a little, which is what most legitimate complaints are asking for.
Above the ceiling, a person decides. Not because large complaints are more serious, but because they are the ones worth two minutes of a manager's judgment and the ones fraud aims at.
The ceiling should be a number the staff know without looking it up. If you have to check a binder, it is not operational.
Refund, remake, and credit are three different products
Policies collapse when they use these words interchangeably. They have different costs and different risks and the policy should name which one applies.
A remake costs you food cost, roughly a third of the ticket, and it satisfies the caller who wanted dinner. It is the default for anything reported the same day where the customer still wants to eat.
Store credit costs you nothing today and brings the customer back, which is why it is the right instrument for a complaint reported the next morning about food that was already eaten. It is also the one customers resent when it is offered instead of a refund they explicitly asked for, so the policy should say when to stop pushing it.
A card refund costs you the full ticket plus the processing you already paid, and it is the one that leaves the building. It should require the original transaction to be matched, which means a person with access to the payment record. Handing that capability to a phone agent creates an obvious attack surface, and the patterns people use against it are laid out in call fraud and chargebacks.
The frequency rule is where the money actually leaks
Ceilings get most of the attention and frequency limits do most of the work.
A rule of once per phone number per ninety days, resolved automatically, with a manager flag on anything beyond that, catches nearly all of the repeat activity without making a rule that punishes a family that genuinely got two bad orders in a year. The flag matters more than the refusal. When a manager sees that this number has called four times since February, always about a missing item, always on a large order, they can look at the ticket history and make a call. Without the flag, four separate hosts each handled a first-time complaint.
This only works if complaints are attached to a phone number rather than to a name, which is one of the practical advantages of a phone channel over a walk-in counter.
Say what proof you require, and mean it
Most restaurants require nothing, which is a defensible choice and should be written as a choice rather than left as a gap. Requiring a photo of the wrong item raises friction on legitimate complaints considerably and reduces claim volume noticeably, and you should decide which of those effects you want.
If you do require something, require it only above the ceiling. Asking a caller for a photo of a missing side of ranch costs more goodwill than the ranch costs money.
Write down what you accept as evidence of an order, too, since that is the check that actually matters. A caller who cannot produce an order number, a name on the ticket, or a phone number matching the one on file is not describing an order you can find, and the policy should say what happens then. The answer can be generous, but it should be decided in advance and not at nine on a Friday.
The time window most policies forget
An unstated window is an infinite window. Someone will call on Wednesday about Saturday, and the person answering will feel obligated because nothing says otherwise.
Same day for temperature, missing items, and anything about how the food arrived. Twenty-four to forty-eight hours for quality complaints where the customer ate some of it. Anything older goes to a manager and usually gets store credit, because at that distance you are buying a relationship rather than correcting an error.
Third-party orders need their own paragraph
Orders that came through a delivery platform are not yours to refund, and a policy that ignores that will have your staff issuing credits against tickets they were never paid for in full.
The rule is short: on a platform order, the restaurant remakes food when the complaint is about food, and directs the customer to the platform for anything about money. Note the order number so the platform charge can be disputed later if the fault was not yours. What you should not do is refund a platform order out of your own drawer to end a call quickly, which happens constantly and is invisible until someone reconciles a month of comps. The rest of that call flow is in the five-minute script for a lost order.
Handing the policy to a phone system
Once the policy is numbers and categories, a voice agent can run it, and running it identically every time is the actual benefit. The agent checks the order exists, checks the amount against the ceiling, checks the phone number against the frequency rule, offers the remedy the policy names, and escalates anything that fails one of those tests. That is a decision tree, not judgment, which is why the categories in the complaint escalation matrix have to be settled first.
What the agent should never do is improvise a remedy the policy does not name, or apologize in a way that concedes fault on a call involving illness. Those cases go to a person immediately, with the context carried across.
Give the policy a version date and re-read it after the first month with real comp data next to it. If the ceiling is never reached, it is too high to be doing anything. If half your complaints escalate on amount, it is too low and you are spending manager time to save nine dollars of food. The tracking that tells you which, and what the complaints keep being about, is in tracking comps and remakes back to their cause.