Scan for the slides, the cost checklist, and contact details

upskilled.consulting/nextup-socal/

Grab the slides

NextUp SoCal · Virtual Event

Reducing
AI Operating Costs

What the tokens actually cost, why the invoice grows faster than the usage, and the five levers that bring it back down.

Nick McCarty

Upskilled Consulting

Fri Aug 21, 2026 · 12:00 PM PST

Part two of the NextUp SoCal AI series

Where we left off · June 5

In June we mapped the levels.
Today: what each one costs to run.

Level 1 — Assistant. One prompt, one answer. Cheap enough that nobody checks.
Level 2 — Collaborator. The tool knows your context. More reading, more tokens.
Level 3 — Chief of Staff. It works without you. It reads constantly.

Where today picks up

74% of companies plan to deploy agentic AI within two years. So the useful question now is a budgeting one: what does Level 3 cost to run for a year?

The hour

From “it’s cheap” to a number you can budget.

01 · The unit

What is a token, and what does one actually cost?

Pages, meetings, emails — converted into the thing you're billed for.

By the end of it you can price any document your team touches.

02 · The multiplier

Why a fraction of a cent becomes half a million dollars.

Three scenarios, same 600 people, same price list, 47× apart.

Same tool, same headcount — three very different invoices.

03 · The levers

Five ways to cut the bill without cutting the capability.

Right-size the model
Script the plain steps
Engineer the context
Stop redoing work
Run commodity work local

What you don't spend is also a lever

What you keep out of the prompt, and what you keep out of your stack.

And the cost of standing still

The most expensive line item is usually the work you didn't automate — and who in the room never got the tool.

01 · The unit

You’re not billed per prompt.
You’re billed per token.

01 · The unit

Your work, converted into the billable unit.

Every artifact your team touches has a token weight.

01 · The unit

Here is what one of everything costs.

Read it in, hand a summary back. Common artifacts, priced.

01 · The unit

For perspective: what everyday work costs.

The unit is cheap. The bill is a volume problem — and volume is a decision we have influence over.

02 · The multiplier

Nobody's budget breaks on price.
It breaks on volume.

Scenario 1 of 3 · Level 1

The assistant.
Chat, summarize, draft.

Scenario 2 of 3 · Levels 2–3

The agent.
Agents read far more than they write.

Scenario 3 of 3 · Past Level 3

Always-on automation.
Agents that never stop watching.

02 · The multiplier

Same 600 people. Same 250 days.
Same price list. 47× apart.

03 · The levers

Five levers. Same capability,
a fraction of the bill.

Lever 1 · Right-size the model

We’re paying frontier prices
for a 1.7% edge.

Lever 2 · Script the deterministic steps

Most of what you're automating
doesn't need to think.

This is the least popular lever and the most valuable one. Lead with the honesty: where the work needs judgment or language, the model wins and you should pay for it.

Beat 1–2: the two lines. A model call has no economy of scale — every run costs the same as the first. A script has almost nothing but economies of scale.

Beat 3: the crossover — around 7,500 runs a year. "That's thirty runs a working day. One recurring report, run over and over across the company — not thirty different jobs."

Beat 4: at 150,000 runs: $27,000 vs $1,350.

"If the task has the same shape every time — pull the file, reformat, check the threshold, send it — you are renting a Ph.D. to read a thermometer."

Practical test to give them: if you can write down the rule, you don't need a model to follow it.

Lever 3 · Engineer the context

Your agent is re-reading the same page
all day long.

Lever 4 · Stop redoing work

The chat window is a shredder
with a very good memory of nothing.

Lever 5 · Run commodity work locally

Smaller models, your hardware,
nothing leaves the building.

04 · What you don't spend

What you don't put in the prompt
is a line on the budget too.

06 · Who gets the lever

The most expensive line item
is the one that isn't on the invoice.

Pivot here. Everything so far has been “spend less.” This is “don’t mistake spending less for winning.”

Beat 1: the compounding gap. An hour a day, per person, over a year. The teams that automate pull away from the ones that don't, and the distance grows.

Beat 2 — deliver this carefully, it is the most important slide in the deck for this room. The old story was that women were not using AI. That gap has essentially closed: 50% of men and 47% of women have used a chatbot. Two years ago that gap was eleven points.

What has not closed is intensity. Daily use, 27% against 20%. Using it for work, 40% against 35%. Believing it makes you more productive, 35% against 25%. The door is open. The gap is in how far through it people walk.

If the room is right for it, add the Lean In finding: only 30% of women say their manager encouraged them to use AI, against 37% of men — and men were 27% more likely to be praised for using it. Same behaviour, different credit. That is worth naming out loud in a room like this one.

Beat 3: the tie-back to June 5. The line on screen is Ashika’s closing slide quoted in full — read it as hers, then add your half.

"That was Ashika’s last slide in June. Multiplying your output is the whole game."

Sources, if challenged: Pew Research Center, 5,119 U.S. adults, fielded Feb 17–23 2026. The workplace and recognition figures are Lean In’s 2026 survey, reported in Forbes. Both are on the slide footnote.

06 · Who gets the lever

The appetite is already there.
The training is the missing line item.

This is the constructive answer to the slide before it. That one showed a gap in intensity. This one shows what actually closes it — and it is not enthusiasm, because the enthusiasm is already there.

Beat 1 — the demand. AI is the number one topic women say they want to learn about. And 63% report a lack of skills and access to training on the job. Say it plainly: "the constraint is not interest. It is access."

Beat 2 — the scale. C-suite executives expect, on average, that 46% of their workforce needs reskilling within three years. Nearly half. That is not an HR programme, that is a budget line — and most of the rooms I sit in have not costed it.

Beat 3 — the tie-back. Every lever in this talk needs somebody who knows to pull it. Routing, context, scripting — none of it happens because you bought a tool. It happens because someone was taught what to look for.

"You are going to spend money on AI this year either way. The cheapest line on the whole page is teaching the people you already employ what these levers are."

For this room specifically: if 46% of the workforce needs reskilling and women already report less access to training, then the reskilling budget is where the adoption gap either closes or gets locked in. That is a decision someone in this audience will personally make.

Where to start

Four moves you can make without a budget request.

01 · 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.

02 · Route it

Name your three biggest use cases. Work through two of them using the cheaper model.

Run the tasks on the cheaper model for two weeks and compare the output. If no material difference can be discerned, you’ve found savings.

03 · 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.

04 · Budget per outcome

Set a price per thing, not per month.

A monthly cap tells you when you're in trouble; tracking cost per task tells you whether it was worth it.

Live demo

This is what “it reads constantly” looks like.

Video plays in the live deck only —
see upskilled.consulting/nextup-socal

25×

reads for every write

Reads (in)500,000/day

Writes (out)20,000/day

Same shape as Scenario 2, a few slides back. This is what that ratio looks like actually running.

Wrapping up

The takeaways worth keeping.

Tokens are cheap. Volume drives the bill. Summarizing a hundred-page report costs about sixteen cents — that alone won’t break a budget. Agents that ingest your business artifacts all day and reason in circles about what to do with them can cost far more than anticipated.
Agents read far more than they write. As we demonstrated, always-on agents can easily cost 47x more than you’re accustomed to paying for ad hoc chatbot sessions.
Roughly three-quarters of cost is decision-driven, not a set price. Routing, context and scripting get you most of the way back — without removing capabilities.
Ask for cost per outcome, not cost per month. Continue building intuition around cost per task and partnering across departments to begin analyzing monthly spend by task and workflow.

Questions?

Nick McCarty

Upskilled Consulting

nick@upskilled.consulting linkedin.com/in/nicholasmccarty

Scan for the cost checklist, the slides, and contact details

upskilled.consulting/nextup-socal/

Scan for the cost checklist, the slides, and how to reach me.

Beat 1: the four takeaways. Read them; don’t elaborate. They are the deck compressed.

If you only get one sentence out: “Ask for cost per outcome, not cost per month. Keep building your intuition for cost per task, and bring other departments into that conversation.”

Landing line for this room specifically: fluency with these numbers is what turns “asking for budget” into “presenting a plan.” That shift matters more here than the dollar figure does.

Beat 2: Questions. Hold this slide for the whole Q&A — the takeaways stay on screen, which keeps questions anchored to something concrete.

Likely questions and short answers:

“Which model should we use?” — Wrong first question. Which tasks, and which of them are routine? Route after that.
“Is open source safe for us?” — For commodity work on your own hardware it is often the safer option, because nothing leaves.
“How do we start metering?” — Admin console, usage export by API key, one key per team. An afternoon of work.
“What if our vendor won’t break out cost?” — A vendor who can’t or won’t break it out is telling you they don’t know either. That’s worth knowing before you sign.

Before Aug 21: scan the QR yourself from a phone that is not signed in to anything, and confirm it lands where you expect. It is live on this slide now, so a wrong link is a wrong link in front of the room.

ReducingAI Operating Costs

In June we mapped the levels.Today: what each one costs to run.

From “it’s cheap” to a number you can budget.

What is a token, and what does one actually cost?

Why a fraction of a cent becomes half a million dollars.

Five ways to cut the bill without cutting the capability.

What you don't spend is also a lever

And the cost of standing still

You’re not billed per prompt.You’re billed per token.

Your work, converted into the billable unit.

Here is what one of everything costs.

For perspective: what everyday work costs.

Nobody's budget breaks on price.It breaks on volume.

The assistant.Chat, summarize, draft.

The agent.Agents read far more than they write.

Always-on automation.Agents that never stop watching.

Same 600 people. Same 250 days.Same price list. 47× apart.

Five levers. Same capability,a fraction of the bill.

We’re paying frontier pricesfor a 1.7% edge.

Most of what you're automatingdoesn't need to think.

Your agent is re-reading the same pageall day long.

The chat window is a shredderwith a very good memory of nothing.

Smaller models, your hardware,nothing leaves the building.

What you don't put in the promptis a line on the budget too.

The most expensive line itemis the one that isn't on the invoice.

The appetite is already there.The training is the missing line item.

Four moves you can make without a budget request.

You cannot manage what you’re not measuring.

Name your three biggest use cases. Work through two of them using the cheaper model.

Pick the task that runs the same way every week.

Set a price per thing, not per month.

This is what “it reads constantly” looks like.

The takeaways worth keeping.

Presenter controls

Reducing
AI Operating Costs

In June we mapped the levels.
Today: what each one costs to run.

You’re not billed per prompt.
You’re billed per token.

Nobody's budget breaks on price.
It breaks on volume.

The assistant.
Chat, summarize, draft.

The agent.
Agents read far more than they write.

Always-on automation.
Agents that never stop watching.

Same 600 people. Same 250 days.
Same price list. 47× apart.

Five levers. Same capability,
a fraction of the bill.

We’re paying frontier prices
for a 1.7% edge.

Most of what you're automating
doesn't need to think.

Your agent is re-reading the same page
all day long.

The chat window is a shredder
with a very good memory of nothing.

Smaller models, your hardware,
nothing leaves the building.

What you don't put in the prompt
is a line on the budget too.

The most expensive line item
is the one that isn't on the invoice.

The appetite is already there.
The training is the missing line item.