Questions to settle before you ask

It is the System of Record for the AI your company runs.

One record per agent: who owns it, what it costs, what it returns, and which risks are open. Each number keeps its source attached, so the figure you hand a CFO or an auditor arrives with the evidence behind it.

Oabo builds that defensible record from the inputs and sources your team records. It does not verify your value claims against your books — a person still has to name the general-ledger event behind a claim and sign for it.

A cost dashboard tells you what you spent. A system of record keeps that spend beside the evidence.

It holds value, risk, ownership and performance next to the cost, and it still has that evidence when the number is challenged. Four properties do the work:

• One record per agent with a named owner, so a question has someone to go to.
• A stage on every number, set by the server and never by the number’s author, so nobody promotes their own claim.
• Each application write adds a linked event to its organization’s hash chain, preserving reviewable change history without claiming database-enforced immutability.
• Four evidence classes kept apart: Realized cash, capacity hours, assisted estimates, and spend stay separate, so unlike records are not collapsed into a vendor ROI claim.

Put the two side by side and ask each one to show its work on a single number.

A record you can defend. Not a return we can promise you.

The gap it aims at has been measured by other people:

Read those carefully, because they measure reporting rather than proof: they count what people said about their own organizations.

It is also a cross-sectional survey, so the cost-visibility figure is an association and not evidence that visibility causes the return.

What Oabo changes is what you can show — and the worked example in our method does not flatter us.

Meridian — a composite example, not a customer — reports $495K of realized cash against $620K of AI spend, which is a realized ROI of –20.2%.

Its 19,200 capacity hours and its modeled value sit beside that figure and outside it, because hours are not dollars and a model is not a result. A record that can only produce good news is not a record. See what the method covers.

You ask with an email address, and nothing else.

The founder reads every request and decides. If the answer is yes, one link arrives by email, and it does three things:

• Sets up your workspace.
• Gives you your own copy of the whitepaper.
• Opens a read-only sandbox you can walk through.

You choose your password when you set the workspace up, not before.

Asking for access costs nothing. So does the trial it opens, and neither asks for a card.

Prices are on the pricing page. If you want terms in writing first, call 551-222-0087 or send us a message.

A fully populated Oabo workspace running on invented data.

Your access link opens it, and it opens read-only: you can walk the whole loop — agents, the cash ledger, baselines, rate cards, evaluation results, the audit pack — and change nothing.

Everything drawn from that data is labeled "Sample data", so you always know what is real.

It used to be open to anyone in one click. That version was closed in August 2026.

There is no implementation project and no professional-services engagement to buy.

Your access link sets up the workspace and you choose your password. From there, four steps:

• Register your first agent — it enrols itself by sending one telemetry event, so there is no SDK to adopt.
• Record what it costs, either by connecting a provider or by entering the figures.
• Enter your first value claim, and name the person who will check it.
• Set a rate card and have finance countersign it, because a rate nobody signed cannot harden a claim later.

The part that takes real time is not ours. It is finding the person who owns each agent and the person who will review each claim.

The trial runs 30 days.

In the first hour you can register an agent, record what it costs, and enter your first value claim.

On day 30 the workspace turns read-only. Nothing is deleted at that moment, and your records stay readable. About a week after that, access is switched off — the records themselves are still not erased. If you want to keep writing to it, tell us before then.

Nothing is verified because you typed it. A number moves through four stages and skips none:

• Potential — someone believes the agent helps.
• Claimed — someone states a figure and signs their name to it.
• Verified — a second person checks the evidence behind it.
• Hardened — it reconciles to a specific general-ledger event, with a named signer behind that event.

Which stage a claim sits at is set by the server, never by its author — the person who writes a claim cannot promote their own work, so a move up always means a second person looked.

Alongside it, the Grade reports how complete the evidence is on two separate scores, count coverage and rate provenance, always as a pair and never averaged into one letter. Walk one claim through it.

Three checkpoints, and only the first is ours to hit.

• Week 1: AI spend and workflow counts are visible.
• Day 30: finance countersigns the rate card, and the first cash backed by a general-ledger entry appears.
• Day 90: an independent reviewer can rebuild every value claim from your own records.

What no timeline can shorten is evidence grade. It rises only as support is added and a second person reviews it, and the Board Readout says plainly which parts are still self-reported.

If your team cannot supply sources and reviewers, those dates move — that half depends on you, not on how fast we ship. What exists today is the inventory, owners, spend, evidence, reviews, audit packs and the readout itself; what is coming is on the roadmap.

Four ways, and you pick per machine or per account:

• Paste an admin key from your AI vendor’s billing account.
• Have your fleet manager deploy one settings file to managed seats.
• Set two headers on calls routed through an API gateway.
• Paste one command into Terminal on a developer machine.

The admin key is the one most teams start with. OpenAI and Anthropic each publish a usage report to the administrators of an enterprise account, and Oabo reads that report on a nightly job — which is how usage from web chat is counted and not only calls made through an API.

Each run re-reads a trailing window rather than yesterday alone, because vendors revise cost after the fact.

Each way states what it can and cannot see before you connect it. Registering the agent is a separate step and comes first: an agent enrols itself by sending one telemetry event. Oabo never writes back to your source systems.

The QuickBooks bill-side comparison is built, but it is not enabled in this deployment.

When Oabo enables its Intuit application credentials, an administrator can connect QuickBooks Online. Oabo then pulls bills and card purchases for known AI vendors nightly, read-only, and compares what you were billed against what it measured — per provider, per month, and never merged into one figure, because an invoice names a vendor and never an agent.

Four outcomes come back:

• Billed with nothing observed — usually the tool one team bought that nobody else knew about.
• Observed with nothing billed.
• A variance wider than tolerance.
• Agreed.

The rows that agreed are shown rather than hidden, because a reader who can see what matched will believe what did not.

Oabo never posts to your books, and there is no copy of your ledger inside Oabo: bill lines are fetched, compared and dropped. Wider ingestion from finance and people systems is on the roadmap.

Usage, cost, and the records your team writes. Not the content of your prompts, and not the model’s replies.

That is structural rather than a promise. The telemetry records have no field able to hold free text.

The receiver that managed machines report to accepts metrics only and refuses log and trace payloads outright, because those are the payloads that can carry prompt and response content.

On the gateway path, where traces are read, the two attributes carrying prompt and completion text are refused at the single place attributes are turned into values — and the event they would land on has no field able to hold them either.

Core workspace workflows make no model inference calls, and Oabo does not send workspace data to a model provider. Support is the narrow exception: when an authorized person asks for a draft, OpenAI receives the disclosed support subject and recent text messages. A person reviews every draft, and Oabo never sends one automatically.

Identities are handled by where they arrive from. Usage read from a vendor’s admin API has the person’s identifier hashed against your organization before it is stored, so the identifier itself is not kept. Usage pushed from a managed machine carries the machine name and the signed-in account, so your own administrators can tell whose usage they are reading.

Free-text fields — notes, memos, descriptions — hold whatever a person types into them, and Oabo does not scan them. Do not put regulated data there.

To build your records, and for nothing else.

It is not used to train a model, it is not sold, and it is not shared for advertising.

Your workspace data stays yours and Oabo processes it on your instructions. Two things are ours rather than yours: the email address on an access request, and anything you send us through the contact form.

On the way out, the honest position. Oabo does not yet delete data on a schedule, and there is no button in the product that erases an organization. When a trial ends the workspace turns read-only and access is later switched off, and the records are not erased at that point. If you need something removed, ask a person and it is done by hand.

In the United States.

The application runs in Google Cloud’s us-east1 region and the database is Neon’s aws-us-east-1. Both are in the United States, and that is where your records sit.

In transit is a different question, and the answer there is no. The site is fronted by Cloudflare’s global edge network, so a request can cross a point of presence outside the United States on its way to us. Stored in the US is true. Never leaves the US is not, and we will not say it.

Backups and point-in-time recovery run at the database provider, with a six-hour recovery window. That is shorter than the phrase "point-in-time recovery" usually leads people to assume, which is exactly why the number is written here.

Your workspace is scoped to your organization and the application binds every query to it, so one organization’s records do not appear in another’s.

• Access runs on nine roles with segregation of duties and business-unit scoping, so a finance approver, an AI lead and an auditor each see only their slice.
• Each application write adds a linked event to its organization’s hash chain. It is application-maintained history, not an immutable database ledger.
• Provider credentials are sealed to the organization and vendor that own them, so a copied row does not decrypt somewhere else.
• Sign-in today is email and password; single sign-on with your identity provider is on the roadmap.

What we will not claim: Oabo has not completed a SOC 2 examination and holds no SOC 2 report. If your review needs one, tell us before you start rather than after. The rest of the posture, including what we collect and who processes it, is on the privacy page.

Oabo ships the controls diligence asks about.

Segregation of duties, finance-signed rates, a hash-chained audit log behind every change, and signed JSON audit packs an auditor can verify independently.

Whether that is sufficient for your regulated data still depends on your deployment and your own regulatory review — that call stays yours, and we will not make it for you.

Two limits are ours to state rather than yours to discover. Protected health information is prohibited in Oabo and Oabo does not sign Business Associate Agreements. Payment card data never enters the product. Both are written into the terms.