Every system you run has an owner. The one spending fastest doesn’t.
Governance and enablement for enterprise Claude deployments. Your databases have an owner. Your metadata has a governance committee. Your infrastructure has IT. Your AI platform has enthusiastic users and a budget line — and nobody accountable for it as a system.
Uber found out in four months
Uber exhausted its entire 2026 AI budget by April. Roughly 5,000 engineers, adoption climbing from 32% to 84% in a single month, per-engineer cost running well past the few hundred dollars finance had modeled — power users at $500 to $2,000 a month, and one two-hour session that billed $1,200.
The cause was not overuse. It was this:
Nobody owned the platform as a platform.
Uber capped usage at $1,500 per tool per month. Walmart, Amazon and Cisco followed with controls of their own. They are not unusual — 79% of enterprises overran their AI budgets in 2026, and only about a quarter have real-time visibility into what their AI actually costs to operate.
Sources: Fortune and Forbes, May 2026 · DoiT survey of 500 finance leaders at organizations with 1,000+ employees, February 2026 · KPMG AI Pulse, Q2 2026. These are third-party research findings, not Data Meaning results.
Five questions you will be asked
Most departments at this stage cannot answer any of the five. That is not a failure of judgment. It is what happens when a platform reaches people before anyone is made responsible for it — which is the normal order of events almost everywhere.
What it looks like from the inside
- Nobody can name the owner. Adoption is championed by people with no budget accountability, and the budget sits with people who have no visibility into usage.
- Spend is a surprise, not a forecast. The first real signal is an invoice, because no cap and no alert was ever set.
- A small number of people drive most of the cost — and nobody knows which people.
- The answer to “how many agents do we have?” is a range, and the range is wide.
- The system of record is a spreadsheet, or an intranet form, thirty rows deep and out of date.
- Someone is doing real work on a personal account, because a company one was not available — so that usage is outside your audit trail and outside your spend report.
- A tool that started out reading things now writes to something.
An organization that cannot list what it has cannot control what it spends, and cannot answer a regulator either. Ownership is the missing piece under all three — and it is the cheapest of the three to fix.
What we do about it — starting with your people
Your builders — the analysts and managers actually making things. Twelve modules, about ten and a half hours across four sessions, in cohorts of 10 to 25 so it stays hands-on. At 100 seats that is four to six cohorts. They learn what changes when an agent can write rather than only read; how to test that it says “I don’t know” instead of inventing a plausible number; how to write down what a good output looks like so it can be checked; and how to recognize when what they built has quietly become something riskier.
Your administrators — the people who will run the platform after we leave. Eight modules covering every control, what each one does and does not do, and the procedures they will execute. They run each one themselves during handover rather than watching us. The test we hold ourselves to: can they change a connector policy, remove a leaver’s access, produce an audit export and run a quarterly review without calling us?
What your team can build without asking, what needs a second pair of eyes, and what needs a real conversation first. Three factual questions decide it — not a committee, and not the builder’s own opinion of their work.
One list of every agent your team has built: what it does, who owns it by name, what it can reach, and how to stop it. If someone leaves, you still know what they left behind.
The platform settings switched on behind your people, so the dangerous mistakes get hard to make. People sign in as themselves. Leavers lose access automatically. Only approved systems can be connected. You can see the spend before the invoice arrives.
Someone keeps that list true after the project ends: the re-confirmations, the reviews, the retirements, and turning your governance decisions into actual configuration.
And someone watches the money. Seats bought against seats actually used, who your power users are, spend against a cap with alerts well before it is reached, and a defensible position at renewal. Sector-average waste on software licenses runs around a third — on AI seats, bought in bulk on optimism, it is plausibly worse. We measure yours in the first 30 days rather than quoting you a number we have not seen.
No server, no software on your network, no firewall change. Almost every control is a setting in an admin screen that already exists.
How you buy it
Ready to Build
$45,000 · 6 weeks
- Both training tracks delivered — builders and administrators
- Every builder leaves having registered a real agent
- Your administrators able to run the platform unaided
- An assessment of what is actually switched on today, tested not inspected
- A full inventory of what your team has already built
- The rulebook: what needs review, and what does not
- The register design, and who owns what
Who signs: you, on your own signature
Controls and Build
$110,000 · 13 weeks
- The guardrails configured and verified by test
- The register live and populated — every agent found, every one owned by name
- Audit records exported and retained beyond the platform’s own limit
- Five working automations, built against your documented processes
- Each one accepted by your reviewer, with a tested stop procedure
- Your processes documented — which survives the automation
Who signs: escalates — but now with a delivered engagement behind it
Operate and Optimize
$3,600/month · annual
- Support for your people during business hours, in your time zone
- The technical work your governance generates — access changes, new connectors, policy changes
- The register kept true: re-confirmations, reviews, retirements
- Re-testing when the model changes underneath you
- Your license spend actively managed — utilization, right-sizing, renewal position
- Covers up to 100 seats and 50 registered agents
- A quarterly pack for your governance forum
Who signs: operating budget
At the end of six weeks your people know how to build well, your administrators know how to run it, and you have a rulebook and a populated register — with nothing further needed from us. It is deliberately sized to be a decision you can make yourself.
What arrives in the first two weeks
A control baseline assessment: 23 administrative controls, each tested rather than inspected — we attempt the thing the control is supposed to prevent, and record what happened. Alongside it, a discovery sweep that reconciles what is actually running against what is registered.
That reconciliation produces one number. It is usually uncomfortable, and it is the most useful number you will see all year.
Two names.
One individual who owns governance for your department, and one individual with authority to accept what we build. Not teams — people.
Everything else is access and calendar time.
A governance practice, applied to agents
Registers, stewardship models, policy frameworks and governance councils are what we have built for enterprise clients for years. This is the same discipline applied to agents instead of data assets.
Your governance council holds every decision. We carry them out and evidence them — and every quarter you get the complete change log to confirm nothing happened outside your authority.
Every automation we build is accepted by your named reviewer, never by us. Your internal audit function will ask about that, and the answer is in the contract rather than in a conversation.
Your administrators are trained mid-engagement, not at the end, and they run each procedure themselves rather than watching us do it. The ongoing contract is meant to be a choice.
Start with the number you don’t have.
How many agents does your team have running, and who owns each one? If the honest answer is a range, that is the conversation.
Data Meaning is an independent consultancy. Claude is a product of Anthropic, PBC.