Automating a newsletter is a weekend. Automating invoice approval, claims triage, KYC review, order exceptions or payout reconciliation is a different discipline, because when those go wrong they go wrong silently, at volume, into your ledger, and you find out in an audit.
So every action is scoped to what it is allowed to touch, anything irreversible waits for a person, and there is a tested way back before go-live rather than a plan for one.
Exception handling is most of the work and it is the part the demo never shows. If a no-code tool can do it safely, use the no-code tool.
What we build is a different purchase from the one most AI automation agencies sell: two SaaS apps wired together in a visual builder, a real service and a crowded one.
We get called when the process moves money, changes a record somebody else audits, or contacts a customer, because the failure mode there is not that the automation stopped. It is that it kept running, wrongly, and nobody noticed until the month closed.
This fits a business where one person can point at the process and say what it is supposed to do — usually an operations or finance lead who knows exactly which spreadsheet is load-bearing.
Scope
We follow the real process rather than the flowchart of it, including the exceptions your team handles by hand and never wrote down. Those exceptions are usually why the last automation attempt failed. Sometimes they move the question upstream, to what is worth automating.
Scoped credentials, idempotent actions, and the deterministic steps left deterministic. A model asked to total a column will eventually total it wrong and sound certain doing it, so the model reads, classifies and drafts, and your database stays the thing that is true.
The gate list is written before the build and kept deliberately short: money leaves, a record is destroyed, a customer hears from you. Everything else runs unattended.
A system that asks permission for all of it gets approved blind by the third week, which is worse than having no gate at all.
The log records what the system saw before it acted: the input, the version of the rule in force, the confidence, the decision it took. When somebody queries a March payment six months later, that is the difference between an answer and a shrug.
A runbook written for the person who runs the process, not for an engineer. Alerts that fire when volumes or rejection rates drift.
A pause switch your team can use on a Friday evening without us, and nothing in the delivery that depends on a licence or a login we control.
Two very different businesses trade under that name. One builds integrations: connect these apps, move this data on this trigger, usually in a visual tool, usually sold as a small fixed project.
The other rebuilds a process that has money or compliance attached and treats it as production software, with permissions, approval gates, an audit record and a way back. Both are legitimate work.
The expensive mistake is buying the first when the process needed the second, and discovering it during a reconciliation. Ask any AI automation agency two questions: what happens when this is wrong, and how would we find out.
Those tools are good and that class of tooling is often the right answer. The difference shows up off the happy path.
A visual builder gives you a run history; it does not give you a permission boundary per action, an approval step tied to your own roles, a record a reviewer can read, or a tested way to reverse three hundred wrong actions at once.
If your process does not need those things, the tool is the correct purchase and you do not need a firm for it. If it does, you are no longer buying an integration, you are buying software that happens to have a workflow shape.
Rank them on two numbers you already have: how often the process runs, and what one wrong instance costs. High frequency and low cost of error is where an off-the-shelf tool wins on price and you should let it.
Low frequency and high cost of error usually deserves a checklist and a human, not software. The work worth commissioning as business process automation services sits where both are high: invoice approval, order exceptions, refunds and chargebacks, onboarding and KYC checks.
That is also the corner where your team is quietly losing its afternoons.
It will get something wrong. The only design question is whether being wrong is loud or quiet.
So every action is reversible, gated, or logged in enough detail to be unwound by hand, and the system is built to stop rather than guess when it meets an input it does not recognise.
What it stops on lands in a queue a person works.
That queue is the real measure of the build: empty for a month means the exceptions were scoped properly, filling up means a path was missed and it gets fixed while your ledger stays clean.
Scope sets it, and the scope is driven by things you can count before you call. Count the exception paths the process really has. Count how many of the systems it touches have a usable API, and how many need a person in the middle.
Then ask whether the output has to survive an audit. A notification flow and a payout reconciliation can be the same diagram and are never the same job.
We quote a fixed fee against a written scope and bill on milestones as verified work lands; how we work has the mechanics.
If you have an engineer who already owns the systems the process touches, they should probably build it.
AI automation consulting is then worth buying for a day rather than a quarter: someone to argue with about the gate list and the failure paths before any code exists.
Bring the build to a firm when nobody internal owns those systems, when the process spans finance and operations so no single team can change it, or when the result has to be explainable to an auditor.
The hire versus engage arithmetic is on its own page.
You do, outright. IP assignment is signed before repository access, by everyone who touches the code, and it covers AI-assisted output. Handover includes the runbook and the rollback path. Nothing depends on a licence we control.
Nobody before an NDA. Access is read-only until scope is signed, and you can revoke it at any time. The AI tools used on your code, and what each one retains, are listed on the governance page.
Retainers end on 30 days written notice from either side, with no minimum term. Milestone billing means you never pay ahead of delivered work.
Founder, Provingmark
Thirty minutes, no charge.
Related: AI consulting & implementation · LLM & RAG integration · Hire vs engage a firm: a cost model · Staff augmentation · How we work