The AI line item, finally defensible.
Depthdata reads the consoles your vendors already give you and returns a single figure a finance team will sign: what went out, what came back, and how much of that was counted rather than guessed.
Four consoles. Four definitions of “active”. No answer.
A company buys ChatGPT, Claude, Copilot and Gemini. Four admin consoles, four billing pages, four ideas of what a user doing something means. The CFO asks one question, and answering it takes a week of exports.
The tools report usage. Nobody reports whether it was worth it.
Every vendor gives you a console. None of them gives you an answer.
Each one is good at exactly one tool, and each one defines its own terms. Put four side by side and you get four numbers that cannot be added up.
| Where you look today | What it covers | What it cannot tell you |
|---|---|---|
| ChatGPT admin | Seats, spend and audit events, through the official API | Usage only through exports and logs. Nothing about your other five tools. |
| Claude console | Seats, usage, spend and audit, all through the API | Its own definition of an active user. Not comparable to the console next to it. |
| Lovable | Seats through SCIM 2.0, on Enterprise with SSO | No usage or spend API. Credits live in the app, and audit is UI-only with no export. |
| Gemini admin | Workspace seats and usage | Spend lives in Cloud Billing, a different system with a different owner. |
| Vercel | Team members and roles by API, spend through the FOCUS billing API | Usage needs AI Gateway tagging on Pro and up. Audit only streams to a SIEM, with no pull. |
Nobody needs a fifth dashboard. They need one definition.
Every AI vendor measures something different, and every one of them is measuring their own product's success, not yours. The gap is not more charts. It is a single definition of active, spend and value that holds across all of them, and that a finance team will sign off on.
The timing is not a hunch. It is in the matrix: vendors have started shipping read-only admin APIs, and a normalizing layer is only possible once they do. Where they still expose nothing, we say so.
One schema. One baseline. Every number labelled.
It attaches to each vendor console with read-only scopes, translates their units into a single set of terms, and grades progress against what this workspace was doing in its first month. Nobody else's average is involved, so nobody can dismiss the result as someone else's company.
A vendor's own API said so. Seat counts, timestamps, invoices.
Arithmetic on the above. The formula is published, so you can redo it.
An estimate. The method is written down and the range travels with it. Never dressed up as a count.
Other tools show you a number. This one shows you how much to trust it.
That is the whole product in one sentence. A dashboard that cannot tell you which of its figures were counted and which were estimated is asking you to take all of them on faith, and no CFO signs that.
From four consoles to one defensible number.
Five steps. Each one answers a question the last one raised, and each one is published, so a sceptical analyst can follow the arithmetic end to end.
Seats, sessions, invoices. Nothing anybody typed.
What a person wrote into a chat box is not an input to any figure on this page, and cannot be. Where a vendor's feed carries that text anyway, it is dropped before it is written down. There is no setting to turn that off, which is the point.
Four questions. Four instruments.
Executive dashboard
Four figures, each labelled. A board member reads it in thirty seconds and knows which numbers were counted and which were estimated, without asking anyone.
Adoption by team
A company average is an average of a healthy population and a stalled one. Split by department, six teams sit at or above the line. Two do not, and the average was hiding both.
Waste, priced line by line
Every item traces to named seats and their activity, not to a percentage someone assumed. You can hand the list to procurement and defend each row.
A full downgrade saves 31%, roughly four times the number above. It also breaks 109 power seats whose premium usage is justified by workload. The tool reports the bigger number and then argues against it.
Goals and root cause
Two goals were failing for what looked like unrelated reasons: adoption was short, spend was climbing. Both traced to the same 22 seats. Two dashboards would have shown two problems.
What we left out is the product.
Every one of these cost us a feature a competitor lists on their pricing page. Each was worth it, and here is what it bought.
Measured, derived or modelled, on the figure itself rather than in a footnote.
The dashboard leads with four numbers and a health score. Everything else is one level down.
A full downgrade saves 31%, four times what we recommend. We report it and then argue against it.
Progress is graded against what this workspace itself did in month one, team by team.
The connector matrix names the vendors that expose nothing, next to the ones that do.
Metadata only. Where a vendor feed can carry conversation content, it is discarded at ingestion.
Connect once. Then it keeps answering.
The friction in this category is the first hour. Every step below is designed to remove a reason to stop.
$4.04k a month back, and one goal that stops missing.
On the modelled workspace, priced line by line. These are expected impacts on a workspace shaped like this one, not results from a deployment.
The difference is not features. It is whether you can defend the number.
| Vendor consoles | BI dashboard | Depthdata | |
|---|---|---|---|
| Tools covered | One each | Whatever you build an ETL for | Eight, one schema |
| Definition of “active” | One per vendor, incomparable | Yours, if someone maintains it | One, published |
| Baseline | None | Whatever you pick | Your month one, per team |
| Spend and value together | No | Only if you model it yourself | Yes, both labelled |
| Confidence on each figure | No | No | Measured, derived or modelled |
| Estimates ship with a band | No | Rarely | Always. Never a point estimate |
| Effort to keep current | A week of exports | An engineer, permanently | A daily read-only sync |
| Reads prompts | n/a | n/a | Never. Metadata only |
Where it goes next. Direction, not dates.
No timelines on this page. Coverage depends on what each vendor exposes, and they ship on their schedule rather than ours. The matrix already says so, and a roadmap that pretends otherwise would contradict it.
Every figure above
was one workspace. Try yours.
Twenty minutes of read-only grants and you will know which of your seats have never been opened, and which of your numbers you are allowed to take to the board.
✦ FAQ
The questions we get asked first.
✦ CONTACT
Know what your AI is actually worth.
Connect your tools once. Get answers you can take to the board.
READ-ONLY ADMIN APIS · NO AGENTS ON LAPTOPS · PROMPTS NEVER READ
