gothink.ai
Watch a run Scope a pilot
Home/GoSpend/The spend bridge
Deep dive / the arithmetic behind the 21%

Where the 21% comes from, before anyone signs anything.

A pilot is not a promise, it is a subtraction you can audit. Here is a representative estate and the exact route from $284k a month down to $223k.

Scope a pilot estate Back to GoSpend
$284k
Today, per month
−$61.2k
Identified
$223k
After
21.5%
Smaller
1
Human signature
01 / the bridge

Five subtractions, one signature, a bill 21.5% smaller.

Every bar is a specific reversible action with an owner and an evidence trail. No rate negotiation, no promises about behaviour.

Exhibit A· the spend bridge$284k → $223k per month
Waterfall chart: a $284k monthly bill falls to $223k through caching, model routing, idle shutdown, storage tiering and one committed savings plan. $200K AXIS STARTS AT $200K SO THE STEPS ARE READABLE — HEIGHTS ARE TO SCALE $284k today −18.4 caching context + semantic −14.9 routing right-sized models −9.7 idle scale to zero −6.1 tiering storage + retention −11.9 commit human gate $223k after
Only the commitment needs a signature — it is the one step that spends money rather than stopping it. Figures representative; your baseline is measured in week two.
02 / the five subtractions

Each one has an owner, a reversal path and a number.

ActionWhat stops being boughtPer monthApprovalReversal
Context & semantic cachingRepeated context and near-identical prompts, re-billed as fresh input.−$18.4knone neededflush the cache
Model routingLarge-model calls on traffic a smaller model answers to your quality bar.−$14.9knone neededrestore routing table
Idle shutdownThree GPU pools with no scheduled work outside business hours.−$9.7knone neededscale back up
Storage tieringHot-tier objects nobody has read in ninety days.−$6.1knone neededrehydrate
Committed savings planOn-demand rates on the demand the forecast is certain about.−$11.9knamed approverterm commitment
Four of the five are reversible from the trace and need no signature. The commitment is the exception, because it is the only step that spends.
03 / how the agent finds them

Predict, attribute, act — in that order.

The bridge above is an output. These three are what produce it, continuously.

Predict

A distribution, not a number

P10/P50/P90 across cloud and tokens together, per team and project. Budget talks start with a probability.

Horizonhours to quarters
Attribute

Every dollar has an owner

Untagged and shared spend allocated back to the workload that caused it. When the curve bends, it names who bent it.

Coverage~100%, automatic
Act

Cuts run under your credentials

Executed through your APIs and your IAM. Above your threshold, the step parks with a named approver.

Autonomya dial your team turns
Why a forecast interval is the product

A single-point forecast tells you nothing about risk. A distribution tells you when the expensive version of next quarter starts to become likely — which is the only moment an intervention is still cheap.

Attribution runs against the same numbers, so when the P90 moves the agent can already name the team, project and workload that moved it. That is the difference between an alert and a decision.

See the eight-week forecast

The ask / start with a scoped pilot

Turn the work you keep discovering by hand into a governed capability.

Three steps, one quarter, measured against your live numbers.

1

Pick a pilot estate

One business unit, one question answered by hand, or one document family.

2

Run a discovery sprint

Two weeks: baseline, target, and a deployment plan for your environment.

3

Prove the number

Measured in your environment, evaluated by your team, before anything scales.

Book a discovery sprint Back to GoSpend