Three weeks, read-only, fixed fee. On our last engagement that was 7.7% of a $263k account — six configuration changes and one bug report — with a further 11% sitting behind two decisions the client already controlled.
A corporate Snowflake estate running at $262,887/year. Read-only access, no changes to their data model.
| Tier | $/year | % of spend |
|---|---|---|
| Committed — config only, no change window | $20,318 | 7.7% |
| Probable — adds two decisions | $48,552 | 18.5% |
| Probable + contract renegotiation | $74,841–101,129 | 28–38% |
Read the full findings register
One engagement is one data point. We quote the committed tier because it's the number we'd defend without further work — not because it's the biggest.
A read-only role on SNOWFLAKE.ACCOUNT_USAGE — the metadata
views Snowflake maintains about your account. Warehouse sizes, query durations,
storage totals, retention settings, credit consumption.
Your tables. Your data. Production. We install no agent, require no network access, and hold no credentials beyond that one role. Nothing we do can change state in your account — the access is read-only by construction, not by promise.
We send you annotated SQL, you run it, you return CSVs. Several clients prefer this and it costs nothing in quality. Your security review gets much shorter.
QUERY_HISTORY stores query text, which can contain literal values
from your queries. If that matters in your environment, run the queries yourself
and redact before sending. We'd rather flag this than have your security team
find it.
Three weeks from access to final register.
Your side costs roughly half a day: one kickoff call, one access grant, and occasional questions when a finding needs context only your team has.
Fixed fee, set by the size of your estate. Not a percentage of savings.
| Your annual Snowflake spend | Fee |
|---|---|
| $200k – $500k | €15,000 |
| $500k – $1.5M | €25,000 |
| Above $1.5M | From €40,000 |
Savings recur; the fee doesn't. On the engagement above, the committed tier alone returned the fee inside the first year and kept returning it after that.
If we don't think the committed tier will cover the fee, we'll say so on the first call and decline the work. Below roughly $200k of annual spend it usually doesn't.
Every consultancy page lists benefits. Here is the other column — the things we'd raise ourselves if we were reviewing this proposal.
Possible, and it's the main risk you carry. Mitigated by qualifying honestly on the first call rather than after the invoice — a well-run estate is a real outcome and we'll tell you before you commit.
A team changing their code, a licence budget. That's precisely why the register is tiered — the committed tier deliberately excludes anything needing someone else's agreement.
Cutting time travel on staging data saves real money and genuinely reduces what you can roll back. We'll tell you which data we think is reloadable; confirming it is your call, not ours.
30–60 seconds on reopen. Fine for internal and sandbox workloads, not fine for anything customer-facing — which is why that finding ships with a scope check attached.
Where that's true the register says so on the row. You'll always be able to see which figures we measured and which we estimated.
We hand you findings, not changes. Most are configuration your team can apply in an afternoon — but somebody on your side has to actually do it, or nothing is saved.
One row per finding: annual dollar value, the assumption it rests on, the change required, and who has to approve it. Sorted by expected value, not by how interesting it was to find.
Committed, probable, and contract. You choose which number goes to your CFO, and you can defend whichever one you pick.
Annotated SQL you keep and re-run. The audit is a one-off; the ability to repeat it isn't.
What we checked and found healthy, stated as explicitly as the problems. Knowing your clustering and pipes are fine stops you spending next quarter there.
Credits multiplied by an assumed rate is how estimates end up 40% wrong. We reconcile to billed dollars before quoting anything, and reconcile storage against account totals before touching a single table. Both checks exist because skipping them has burned us.
On the engagement above we called storage clean, then reversed it twice as the reconciliation closed. The final answer was a $5,800/yr finding we'd have missed by stopping early. A review that never retracts anything isn't being careful.
The register is the deliverable. An audit that exists to generate an implementation quote isn't an audit — and you'd be right to discount every number in it.