Automation in a business you own is only worth having if you know exactly where it stops.
So here is the line, drawn in the software rather than in a policy document.
Reversible things happen. Irreversible things wait.
The gate is by class of action, not by named automation — so a feature added next
year cannot slip past a rule by being new. Filing a CV against a role happens on its own.
Replying to that candidate waits for you.
Nothing reaches a client unapproved
The first message of any kind stops for a person. Approve it and that kind is
yours; it does not quietly widen into something else you never agreed to.
A retainer bills once, provably
Each period is claimed by a row with a unique key, so a job that runs twice — or
retries after a crash — cannot invoice your client twice. That is a database guarantee,
not a careful code path.
Backdated agreements do not fire back-invoices
Record a retainer that began three months ago and it starts from the next billing
day. Under-billing you can fix; a client charged for three months they already paid for
by hand is a refund and a phone call.
Payments are reconciled, not assumed
A webhook is a promise, not a guarantee. The system reads the provider's own list of
payments back and finds what never arrived — and refuses to guess which invoice an
unmatched payment belongs to, because a debt wrongly marked settled is never chased again.
Peer reviews cannot be swung by headcount
Peers share a fixed slice rather than each holding a vote, so a rating does not move
because somebody was on leave. And nobody learns who wrote what — below three replies the
comments are withheld entirely, because two are guessable.
Utilisation says what it counted
Measured against contracted hours, from approved time only, and the screen says so.
Two figures that appear to disagree because one counts submitted hours and the other
counts approved ones is how people stop believing a dashboard.
Your data is yours, structurally
No other studio can reach your records. Enforced three times over, the last by the
database itself. Where a model drafts text, GSTINs and PANs are stripped before the
request and never restored.
There is no claim here about a proprietary algorithm, because the useful part is not the
cleverness — it is that the arithmetic is written down, tested, and does the same thing on
the last day of the quarter as it does today.