The part of the company that isn't the product
Every founder building a subscription business ends up doing three jobs that aren't the product: answering support, chasing churn, and re-deriving the same numbers. Brohns triages and drafts the support replies, segments your customers from their real history and drafts a win-back for the valuable ones going quiet, and answers questions about your own data with arithmetic done by code — the model shapes the question, code computes the answer. Everything outward waits on you.
Sound familiar?
Before the team
The work that keeps not happening.
Support eats the morning that was for building
Twenty messages, seventeen of them the same three questions. Each one is two minutes, and together they're the reason the thing you meant to ship didn't ship.
You find out about churn after it happened
The renewal doesn't come through and you look back and see it was obvious for six weeks. The signal was in the data; nobody was looking at it, because looking means building the thing that looks.
Every number gets re-derived
Someone asks why revenue moved and it's another export, another spreadsheet, another hour arriving at a number you calculated last month too — slightly differently.
The dashboard you'll build "later"
It's been on the list for a year. It's not hard, it's just never more urgent than shipping — so the company runs on questions nobody has time to answer.
A week with Brohns
With the team
What actually changes, day by day.
[01] of 05
Monday — the questions worth asking
Before any data arrives, the first thing a fresh workspace makes is the measurement plan: the questions this business should answer every month, what data each one needs, and what a good and a bad answer look like. No numbers in it — the numbers come from your own import. It's the artefact that stops the endless re-deriving.
[02] of 05
Tuesday — import, then just ask
Bring in a CSV, a spreadsheet, or connect a source. Ask in plain language and a constrained query spec is produced, then plain code computes the answer with the exact rows behind it. The model shapes the question; it never does the arithmetic. There's no generated SQL, so there's nothing to inject into.
[03] of 05
Wednesday — support, triaged in one pass
An incoming message is classified and answered in a single call: category, sentiment and priority land as real columns, and the draft is grounded in the actual order or account context and your own policies. Where a policy doesn't exist, it says it'll check rather than inventing a refund window. First-response deadlines are derived from priority, so what's about to breach is visible.
Outreacher drafted an email
To: Sparkle Clean · sparkle-clean.nl
Hi — I noticed your booking page takes ~6s to load on mobile, which is where most of your customers are. I put together a couple of quick wins…
Why: opens on the site's weakest point, in your tone.
[04] of 05
Thursday — churn, before it's a renewal notice
Customers are segmented from their real history — new, loyal, champion, at risk, lapsed — by code, and the most valuable one that's gone quiet gets a win-back drafted from what they actually bought. Contracts approaching their notice deadline get a short keep, renegotiate or cancel brief with the real dates and cost.
[05] of 05
Friday — a dashboard you asked for out loud
Ask Bro for a dashboard and it proposes one before building it: which metrics, from which source, and whether the data lives somewhere it needs to connect first. You say yes, and it builds it from a validated spec — a metric it can't point at is dropped rather than displayed. Put a watch on any number and you're told when it crosses.
The payoff
Numbers from the pricing page
Concrete, not hand-wavy.
14
days of free trial — no credit card
500
credits included in the trial
€39
per month for Brohns Pro
4000
credits included every month on Pro
Deterministic where it counts
Balances, buckets, segments and metrics are computed by code. A model can shape a question, propose a category, or draft a message; it never produces a number you'd act on. That's the difference between a dashboard you check and one you trust.
Autonomy you grant per kind of work
Support replies can send themselves only in categories that are safe by design — where-is-my-order and product questions — never money, complaints, cancellations or address changes, never on negative sentiment, and never on a message showing signs of prompt injection. Your settings can only make that stricter, never looser.
Predictable cost, honest metering
€40 a month for 4,000 credits, and credits move only when agents work. The monthly meter is computed server-side from real rows against the same public price list you see on the pricing page — not estimated in the browser. Run out and the engine pauses with a visible reason rather than quietly overspending.
Every outward move — each email, reply or post — waits for your approval first. That's not a setting; it's how Brohns is built. The full credit table lives on the pricing page.
Dashboards & data
Ask Bro for a dashboard in plain language: import a spreadsheet, connect an API, and get live KPIs, trends and alerts — every number computed by code, never generated by AI.
Support
Brohns reads incoming questions, triages them, and drafts helpful replies from your knowledge — sensitive ones always wait for you.
E-commerce growth
Agents spot abandoned carts and win-back moments, segment your shoppers, and draft on-brand campaigns — you approve before anything sends.
Good to know.
Does the AI calculate my metrics?
No, and that's deliberate. A model turns your question into a constrained specification, and plain code computes the result from your rows — with the rows themselves available behind the number. No SQL is generated, so there's no injection surface, and no figure is a model's arithmetic.
Can support replies send themselves?
Only if you turn it on, and only for categories that are safe by construction: where-is-my-order and product questions. Money, complaints, cancellations and address changes always wait for a person, as does anything with negative sentiment, high urgency, or signs of prompt injection in the incoming message. There's a daily cap, and every send goes through the same guarded path as a manual one.
Can I connect our own API as a data source?
Yes. Point it at a JSON endpoint and Bro derives the schema and the field mapping from a sample, creates the collection and syncs it. The URL is validated against internal and metadata addresses on both saving and syncing, responses are size- and row-capped, and an auth header's secret lives encrypted in the vault — never in the connector row.
How is this different from a BI tool?
A BI tool shows you the number. This also does the work that follows from it: the win-back for the customer the number is about, the support reply, the renewal brief — all as drafts that wait for you. And when you correct one, the correction is kept as a lesson that shapes the next.
Can other assistants reach my data?
Only if you issue a key, and only within a deliberately narrow surface: four read tools and exactly one write, which puts a request in the same queue as your own "ask the agents" button. There is no tool that sends, publishes, pays, posts to the ledger or approves — what isn't in the list can't be done. Keys are stored as hashes, comparison is constant-time, revoking keeps the record, and every proposal is written to the audit log.
Put a team on it — built for saas & subscription founders.
Free 7-day trial with 500 credits. No credit card. Nothing goes out without your approval.