Spend you can tie to the invoice.
Finance distrusts a vendor's savings math, and it should. So every saving the Gateway claims is tied to one request, priced against the baseline your own administrator set, and reconciled against your provider's invoice. A difference above the tolerance opens an item naming the window, the account, both totals, and the gap. You check the figure instead of believing it.
- window
- 1 calendar day, one provider account
- ledger
- $4,182.66 recorded by the Gateway
- invoice
- $4,194.10 billed by the provider
- variance
- 0.27 percent, inside tolerance
- savings
- $611.20 verified, attributed by lever
- kept
- $488.96 to you at the Growth share
How a saving becomes money
- Your baseline, not ours. Your administrator sets which model each kind of work would otherwise use, and every request is priced against that choice. We never rewrite the baseline to flatter the result.
- Graded evidence, then your decision. The Gateway samples real requests, and a judge model compares the current and candidate answers on a fixed sample sized to your plan's quality bar. You get the recommendation with the sample count and the date attached. Traffic moves when your administrator accepts, and not before.
- Attribution per request. Each accepted route's saving is split across named levers in a fixed order, so every total is the sum of lines you can read.
- Reconciliation to the invoice. The ledger is checked against the provider invoice for each account, each day, and our share accrues only on savings that pass that check.
What a finance review asks
Plan prices are printed, and planning figures are labeled.
| Question | Answer | Where to verify |
|---|---|---|
| What do we pay, and what is variable? | A fixed monthly fee by plan, $49 Starter, $199 Growth, $999 Scale, plus a share of verified savings at 25, 20, or 15 percent. Inference is billed to you by your own providers, never by us. No seats, no usage fee, no negotiated rate. | Pricing |
| Does the fee ever exceed the value? | The share is a fraction of verified savings, so every dollar of share sits beside at least three dollars you kept. The monthly fee comes on top, and the dashboard shows the point where a bigger plan starts to pay for itself. | Worked example on pricing |
| Is a saving verified or marketed? | The share accrues only on savings verified against your own provider invoice, and only on routes you accepted. In a month with no verified saving, the share is zero. | Your own savings export |
| What tolerance applies to reconciliation? | 1 percent per provider account per calendar day. A difference above it opens a reconciliation item with both totals and the gap. You can change the tolerance. | Reconciliation section |
| How is rounding handled? | Every cost is stored as a whole number of millionths of your currency, rounded once per line, with ties going to the even number. Totals round to cents only for display and for comparison with the invoice. | Ledger export |
| What about currency? | You set your currency when the organization is created, and every price is stored in it. The Gateway converts nothing, so when a provider invoices in another currency, you convert, and the item records the rate source and conversion date. | Reconciliation item fields |
| Where do period boundaries fall? | Each window is a calendar day in your chosen time zone, UTC by default. A cost belongs to the day its request completed, and a correction that arrives after a window closes lands in the next one. | Period boundary rules |
| Can we cap spend before an overrun? | Budgets act in order, warn, then throttle, then refuse, for each team, user, or key. A budget acts before the invoice rather than reporting after it, and it never quietly switches the model that serves a request. | Budgets in the admin API |
| Who owns the spend internally? | Every token and every screening catch reports by team and by person, so a chargeback comes from the record. | Accountability |
| What are the support terms? | Severity 1 is a 1 business hour standard response and 15 minutes on enterprise terms, with a status page entry inside 30 minutes and a post-incident report within 5 business days. A monthly SLA report arrives as a signed pack within 5 business days of month end. | Support tiers and severity tables |
| What does leaving cost? | Technically, one base-URL change, reversed. Commercially, your printed plan runs until renewal, annual billing is ten months paid up front, and the savings share is billed monthly in arrears, because it follows the statement. | Commercial terms |
The cost of ownership, said plainly
A cost hidden here surfaces in month three, and by then it sours the relationship.
- Policy has an owner, and it is not us. Someone owns which aliases exist, what each routes to, the quality tier each kind of work needs, each budget ceiling, and which attestations each data class requires. Providers change prices and retire models, so this is standing work rather than setup.
- Refusals need a reader. Refusals pile up. Some show a client doing something it should not, and others show a detector that is too broad. Expect a weekly review at first, heaviest in the weeks after a new screening rule.
- Onboarding is cheap, migration is work. Issuing credentials and mapping identity groups needs no ticket per person. Pointing each existing service at the Gateway is still a change and a test run per service, so budget for coordination across teams more than for engineering effort.
- The evaluation effort is yours. Reconciliation needs a few hours from a finance partner, and an evidence-pack review needs your compliance officer. That time is the cost of not taking our word for it, and it is worth spending.
A saving nobody can tie to an invoice is a marketing number. Finance should refuse it, including ours.
The risk of leaving it alone
- Spend with no owner is spend nobody caps. An AI invoice that arrives before anyone knows what it will say is the first problem finance reports. Budgets by team, user, and key act before the invoice, and the record names who spent what.
- Visibility alone is now table stakes. Many tools chart an AI spend trend, and switching providers can lower a bill. What stays rare is a figure that reconciles to the provider invoice, and a limit that acts before the overrun.
- One pilot, named as one pilot. One real deployment has measured how these savings work. One deployment shows the mechanism, not a typical rate, so nobody can hand you an expected percentage yet. Your own first workload is the only figure that will mean anything to your board.
Reconcile one month yourself.
We admit customers in cohorts, in waitlist order. Don't miss the next one. When your slot comes up, you run one real workload on the plan you chose, export the ledger, and compare it against your provider's own invoice for the same window. If we stop earning our fee, leaving is the same URL change, reversed.