What an AI Help Desk Actually Saves You in a Year
Every AI support pitch opens with the same number: the salary you will not have to pay. It is the wrong number, and any CFO will take it apart in about ninety seconds. Here is the honest version — where the money actually is, the arithmetic to check it yourself, and the point where this stops being worth buying.
The number everyone quotes is the weakest one
The standard pitch goes: you are hiring a support person at $90,000, that is really $112,500 once you count payroll tax, benefits, and a laptop, so buy our thing instead and keep the difference.
Two problems with it. First, it is arguing against a hire the company has already decided to make, which means you are telling someone their judgment was wrong before you have earned the right to. Second, and worse, it is not true. A support hire and an AI help desk are not substitutes. They are good at almost opposite things.
A support person is good at the ambiguous ticket, the angry customer, the one that needs a judgment call. Software is good at the fiftieth identical webhook question this month. Pretending one replaces the other is how you end up with a customer who is unhappy in month four.
The expensive part is not answering tickets
It is the interruption.
Most support tickets do not need a developer. But some do — a real bug, a data fix, a genuinely weird edge case in the API. Those get escalated, and when one lands on an engineer, the cost is not the twenty minutes they spend fixing it. It is the hour on either side.
Anyone who has been deep in a problem knows this. You are holding a dozen things in your head, a ticket arrives, and by the time you get back you have to rebuild the whole picture. That reconstruction is invisible on every timesheet ever filled out, and it is where the money goes.
So the number that matters is not the support salary. It is this:
- How many tickets a week end up in front of a developer
- How long each one actually costs, counting the return trip
- What that developer's hour is worth fully loaded
The arithmetic, worked
Here is a company shaped like a lot of the ones we talk to. Roughly thirty people, a technical product, one open support req. These are illustrative figures to show the method, not a case study. Put your own numbers in — that is the entire point.
- 8 escalations a week × 52 weeks = 416 a year
- × 45 minutes each, including the context switch = 312 engineer-hours
- × $95/hour loaded (about a $165k developer) = $29,640 a year
Then the front-line load. At 150 tickets a week and 12 minutes each, that is 1,560 hours a year of support time, which at a loaded $54/hour is about $84,375.
Call the two together roughly $114,000 of work sitting in the queue. A well-built help desk handles somewhere around 60% of it. Not 90% — anyone promising 90% is selling you a number they will have to explain away in month three. Sixty percent is about $68,000 a year.
Now the honest part
Those are not sixty-eight thousand dollars going back in the bank.
If you free up 312 engineer-hours, nobody writes you a cheque. You get 312 hours of engineering capacity back. That is genuinely valuable — it is a feature that ships a quarter earlier, a migration that finally happens — but it is capacity, not cash.
The only line in that math that is truly hard dollars is a hire you decide not to make. And that is a decision for the company to reach on its own, after watching the volume drop, not something to promise up front.
We would rather say that out loud than have you work it out afterwards. A vendor who inflates the easy number is telling you something about every other number they will give you.
When it is not worth buying
Some honest disqualifiers. If any of these are you, the return will not be there, and we would rather say so on the first call than six weeks into an engagement:
- Very few escalations. Under about five a week, the engineering interrupt is not large enough for the savings to clear the cost. Fix the docs first; it is cheaper and it might be enough.
- Every ticket is genuinely unique. Deflection depends on repetition. If your top three ticket types are not actually your top three, there is nothing to deflect.
- No test suite and no staging. Automated triage still works fine. Automated fixes do not, because there is no safe way to prove a change is correct before a human sees it. Scope down to triage and revisit later.
- Nobody willing to own the approval queue. Every change should be a pull request a human approves. If no one has room for that, the system has no gate, and a system with no gate should not be near your production code.
Take your estimated annual saving and divide it by three. If our price for year one is not comfortably under that, the deal is too thin and you should not buy it. A three-to-one return means you can be wrong by two-thirds and still come out ahead — and on estimates like these, you might be.
What it costs to run, honestly
People assume AI systems are expensive to operate because the models are expensive to build. At this scale they are not.
A help desk at the volume above costs roughly $190 a month in model calls and infrastructure. Cheap classification on every ticket, a mid-tier model to draft real answers, and an expensive model only for the reproduce-and-fix loop, which is a few dozen tickets a month rather than hundreds. Published rates, nothing exotic.
We mention it because it explains our pricing rather than hiding it. You are not paying for tokens. You are paying for the system being built correctly, the guardrails around what it may touch, and someone owning it when it breaks on a Thursday night. The compute is a rounding error and we would rather you know that than wonder.
Check your own numbers
You do not need us to do this. Three numbers, off the top of your head:
- How many tickets a week reach a developer?
- How long does one really cost, counting getting back on task?
- What is that developer's fully loaded hourly rate?
Multiply the three, take 60%, and you have the honest ceiling on what this is worth to you. If that number is small, good — you just saved yourself a sales call. If it is large, it was large before you read this, and it has been costing you all year.