Article
The Financial Case for an AI Developer, for the Finance Director
Most of the material written about AI in software development is addressed to engineers, and it is full of percentages that nobody can trace. A finance director reasonably wants something simpler: what am I buying, what does it commit me to, what is the bounded experiment, and what evidence would make me continue.
We have an AI Developer Platform, installed in the client's own Microsoft Azure and maintained by us. This post sets out the commercial shape without a single invented number, because a case that depends on a made-up saving is not a case.
What you are buying
You are buying capacity from the team you already have. The platform writes the tests, writes the code and builds a private preview for each story; your Product Owners decide what gets built and your reviewers approve it. The headcount does not change. What changes is how much of the queue clears.
That distinction matters financially. A tool that promises to replace salaries is making a claim about your people. A platform that makes the same team deliver more is making a claim about throughput, which is measurable in your own numbers.
The three cost lines, and who holds them
Azure. The platform runs in your subscription, so the infrastructure cost appears on your cloud bill, under your existing agreement and your existing discount structure.
AI usage. The model calls are made against your own AI account. Usage is bounded per story by a cost ceiling, so it behaves like a set of small commitments rather than an open meter.
Gyzer. A setup fee, then a monthly subscription per product. Installation takes two weeks.
Read together, this is unusual in a helpful way: you hold the two variable cost lines directly, so you can see and control them, and the one fixed line is ours.
The ceiling, in finance language
Every story carries a cost ceiling. At the ceiling, the work pauses and a person decides what happens next. That single mechanic converts a category of spend that usually arrives as a surprise line item into a bounded cost per unit of delivered work. It also gives you something to compare month to month: cost per story, and how many stories cleared. Three organisations run it today, 678 stories delivered, each story carrying its own cost line.
What the pilot is designed to prove
What the pilot measures in practice is set out separately, because that measurement is what the case rests on.
The buying sequence is short: a 30-minute AI Engineering Assessment, then a six-week paid pilot on one product at a fixed fee, credited against setup. Then the setup fee and monthly subscription.
The pilot is not a demonstration. It is a real product, real stories, real reviewers, with the ceilings, the logs and the limits all switched on. At the end you should be able to answer four questions with your own data: how many stories cleared, what each one cost in AI usage, how much reviewer time it consumed, and whether your reviewers were comfortable approving what they saw.
If the fourth answer is no, the first three do not matter. That is worth saying plainly, because the failure mode in this category is a technical success that nobody in the organisation trusts.
The honest limits of the financial case
Two of the seven limits are financial controls: a ceiling on every story, and a log that records what each action cost. Everything else in this section depends on them.
We will not put a number on the saving. Anyone who does, before seeing your repository, your story quality and your review process, is guessing. The variables that move the outcome are knowable but specific to you: how well your stories are written, how much of your codebase is already tested, how much reviewer capacity you can release, and how much of your backlog is genuinely well specified.
The pilot exists to replace that guess with your own figures.
How to build the case internally
The strongest version of this case is not made with a percentage. It is made with three lines your own team can produce.
The first is the size and shape of the queue: how many well-specified stories are waiting, and how many developers you have to clear them. The second is the reviewer capacity you can realistically protect, because that is the constraint that decides whether throughput changes anything. The third is the cost per story after the pilot, which you will have by week four.
Put those three together and the question becomes arithmetic rather than advocacy. Put a vendor's claimed percentage in front of a finance director instead and you will spend the meeting defending someone else's number.
What to ask before you sign anything
Ask which costs you hold directly and which are invoiced. Ask what happens at the cost ceiling. Ask what the pilot will measure and who signs off the measurement. Ask what the exit looks like if the pilot disappoints, and what you keep: the accounts, the repository, the logs and the test records are yours.
Ask, too, how the story log works. Every action is logged against its story with its cost. That is the artefact that lets you audit the spend after the fact rather than argue about it.
What to do next
Bring your finance lead to the assessment alongside your engineering lead. The numbers that matter are the ones you can measure in the pilot, and the pilot is designed so that you can.