NextUp SoCal · AI Webinar Series
The AI Cost Checklist
A companion to Reducing AI Operating Costs. Everything here can be started without a budget request. Work top to bottom — the order matters more than the percentages.
Four moves that don’t need a budget
Start here. Two of these an executive can order without a technical conversation.
Meter it — you cannot manage what you’re not measuring.
Ask for cost per team, per month, and keep building your intuition for cost per workflow.
Most platforms have this reporting switched off by default; turning it on is an afternoon of work.
Done when: someone can tell you last month’s cost for one named workflow.
This week
Route it — name your three biggest use cases.
Work through two of them using the cheaper model for two weeks and compare the output.
If no material difference can be discerned, you’ve found savings.
Done when: a written rule says which tasks use which tier.
2 weeks
Automate one thing.
Pick the task that runs the same way every week — the Friday report nobody enjoys
building. Ask for it to be automated once, and for the recipe to live somewhere the team
can find it instead of one person’s inbox.
Done when: the recipe is in shared storage and someone else can run it.
1 month
Budget per outcome, not per month.
A monthly cap tells you when you are in trouble; tracking cost per task tells you
whether it was worth it.
Done when: your AI line item is dollars per outcome ($/report, $/ticket) — not a flat monthly figure.
Ongoing
If you only do one thing: ask for cost per outcome. Answering it forces the same
tier-and-volume breakdown that routing and metering need — the rest follows from there.
The five levers
Applied in order to a $525,000 baseline, these took it to roughly $137,000 — without removing a capability from anyone.
Right-size the model
The test: of all your business tasks, how many genuinely need the best model on earth?
Route the rest to a cheaper tier. The best open-weight models are within a couple of points
of the paid ones on routine work.
−40%
Script the plain steps
The test: if you can write the rule down, you don’t need to rent judgment to
follow it. Where the task genuinely needs judgment or language, the model earns its price.
−15%
Engineer the context
The test: ask what your agent re-reads on every step. Capping it — a short
rolling memory instead of the whole history — is a setting, not a rebuild. Ask your vendor
about prompt caching in the same conversation.
−25%
Stop redoing work
The test: where do finished prompts and outputs live? A prompt someone tuned
for four hours is an asset, and so is the artifact it produced. Both belong in shared,
versioned storage.
−15%
Run commodity work locally
The test: which workloads are routine, high-volume, and sensitive? Those are
the candidates. Two or three small specialist models share one GPU, and nothing leaves the
building.
−20%
Check the volume before you buy hardware. Local wins on volume, not by default —
against commodity-tier cloud pricing the electricity alone can cost more than the API call.
And the one that does need a budget: training. 63% of women report a lack of skills
and access to AI training at work, and the C-suite expects 46% of workers to need reskilling
within three years. Every lever above needs somebody who knows to pull it.
Slides, this checklistand how to reach me
1 / 2