What a Growth OS Costs to Run: The Bill After the Build

The build gets quoted. The run gets discovered in month four, when finance asks what the eleven new line items on the corporate card are and nobody can name them all.

The pricing article covers what an engagement costs. This covers the year afterwards: what a working Growth OS bills monthly, which lines grow with what, and where the arithmetic stops working.

Infrastructure is cheap until an automation gets clumsy

Hosting a first-party capture endpoint is genuinely small. A container or serverless function handling a few million events a month sits in the tens of euros, and object storage for raw events is cheaper than people expect. Warehouse storage is minor too.

Query compute is where the surprise lives, and it almost never comes from analysis. It comes from a dashboard refreshing every fifteen minutes against a full events table for a screen three people open, or a reverse-ETL sync re-reading a whole table hourly because nobody configured incremental mode. Partition by date, cluster on what you filter by, and set per-user byte quotas before handing out access. The budgeting point, which the warehouse piece covers further, is that a well-behaved stack runs a few hundred euros a month and a badly behaved one several thousand for the same answers.

Data is the line that scales with ambition, not revenue

Usually the largest external cost, and the least examined.

Enrichment. A few cents per record resolved in volume, with the bill set by how promiscuously you enrich. Every inbound form fill is reasonable. Every company that visits the site, monthly, whether or not anything changed, is how a 500 euro budget becomes 2,500.

Intent providers. The step change: third-party intent sits in the low thousands per month and is normally sold on annual commitment, so it is not a variable cost at all but a fixed one you agreed to for a year.

Verification. Fractions of a cent per address, and not optional for outbound, because the alternative is bounce damage that costs far more.

A working range for a thirty-person B2B company running both motions: 800 to 3,000 euros a month across all data lines, with intent deciding the end you sit at.

Sending is cheap per unit and expensive per mistake

Domains cost about fifteen euros a year each. Inboxes are five to eight euros a month per seat, so twenty sending inboxes is under two hundred a month, a rounding error against the reputation risk they contain. A sequencer runs one to five hundred.

The real cost here is not the invoice. Deliverability failures waste everything upstream: the credits, the research, the copy. A burned domain never appears as a line item. It appears as a quarter of flat outbound.

Model spend is smaller than feared and harder to cap than expected

Teams brace for a large AI bill and get a modest one. Drafting, classification and enrichment agents at realistic B2B volumes, thousands of records rather than millions, land between one and eight hundred euros a month.

What catches people is shape, not size: a retry loop with no ceiling, an agent reprocessing the same records nightly because idempotency was never implemented, a context window carrying a whole account history into every call. Set hard monthly caps per workload, alert at seventy per cent, and log token usage per workflow so the bill attributes to a job, not to "AI".

The observability nobody puts in the budget

Every system must say when it stops working, and this category gets skipped at build time and bought after the first silent failure: error tracking, uptime monitoring on the capture endpoint, log retention long enough to debug last week, alerting that reaches a human. Fifty to three hundred euros a month, against the question of how long anyone would take to notice an enrichment provider quietly returning empty results. If the answer is weeks, you already pay this in bad data.

The human line, stated honestly

A Growth OS does not run itself, and a provider implying otherwise is describing a demo. Somebody approves agent output, reviews the suppression list, investigates the weekly reconciliation, updates routing when territories change, and fixes whatever broke when a vendor changed an API. That is eight to fifteen hours a week, and it needs one named person with the access and context, because split across three it is nobody's.

The honest claim is narrow. A system like this reduces the rate at which you add headcount; it does not reduce headcount. The team stops growing linearly with volume because the mechanical work is automated. If someone promises it replaces two FTEs, ask which two, and who does their approvals afterwards.

Which lines scale with what

With contacts and accounts. Enrichment, verification, some CRM seat models, marketing automation tiers. These grow with database size, which grows whether or not the database improves, so the category quietly compounds.

With volume. Sending, model spend, query compute, event ingestion. These track activity: halve outbound, they halve.

Flat. Intent commitments, most subscriptions, observability, hosting baseline. They ignore what you do, making them the first place to look when the total must come down.

Added up, a mid-market Growth OS commonly runs three to eight thousand euros a month in tooling and data, plus the operator's time, and one or two lines usually exceed half.

What to do when a line item runs away

Do not renegotiate first, find the driver. Query spend is almost always one scheduled job, named by the query history in minutes. Enrichment is almost always a trigger firing more often than intended, visible by hour in the usage log. Model spend is almost always a retry loop. Then ask whether the spend ever bought a decision: an intent subscription that never changed who got contacted should be cancelled, not renegotiated.

A ten-minute check on your current stack

Export three months of invoices into one sheet, then make two passes. Sort by annual cost and write, beside the top five, what decision each changed last quarter; a blank line is a candidate.

Then list every per-seat tool and count how many people actually logged in last month. Fewer than three real users means one person's habit, priced as such, or shelfware nobody cancelled. Teams doing this first time typically find fifteen to thirty per cent of software spend with no owner.

Where the model breaks

Take a run cost of five thousand a month. To be worth paying, it must save more than five thousand of loaded labour or generate more pipeline than it costs. Eight to fifteen hours a week of operator time is about a third of an FTE, near two to three thousand a month fully loaded in Germany. So a system whose only benefit is saved time does not pay for itself at that rate. It has to produce pipeline.

For a small team with one channel and a few dozen deals a year, it will not. Volumes are too low for automation to compound, and the operator hours come out of the person who should be talking to customers. The right shape is one motion run manually and documented, with an audit when growth breaks it. Infrastructure built before it pays is a cost centre with good branding.

Frequently asked questions

What does it cost to run a Growth OS each month?

For a mid-market B2B company, tooling and data commonly run three to eight thousand euros a month: hosting and warehouse compute in the low hundreds, enrichment, intent and verification between eight hundred and three thousand, sending and model spend a few hundred each, observability under three hundred. Intent data decides which end you land on, and on top sits eight to fifteen hours a week of an operator's time.

Does marketing automation reduce headcount?

It reduces the rate at which headcount grows, a different claim. The mechanical work, list building, enrichment, first drafts, audience syncs and report assembly, stops scaling with volume, so the team handles more without adding people. Approvals, judgement and customer conversations stay human. A provider promising replaced FTEs should be asked which roles, and who does their approvals.

Which running costs grow fastest?

Those priced per contact or per record, because a database grows whether or not it improves: enrichment, verification and per-contact tiers in marketing automation. Volume-priced lines such as sending, query compute and model spend shrink when activity does. Flat commitments, especially annual intent subscriptions, respond to nothing, making them the first candidates when a budget must fall.

How do I find out if we are overpaying for our stack?

Put three months of invoices in one sheet, sort by annual cost, and write beside each of the top five the decision it changed last quarter; a blank line is a cancellation candidate. Then check every per-seat tool for how many people logged in last month, since under three real users means a habit or shelfware.