The routine work, done before anyone sits down to it

Often the second. If the work is a fixed rule (an order over a limit needs approval, a part below its reorder level gets ordered), plain software does it exactly, and AI would only add uncertainty. AI earns its place when the work is reading, sorting, judging or drafting things that arrive in messy human form: a WhatsApp order in three languages, a scanned invoice, a complaint. The first call sorts your problem into one pile or the other, and we say which even when the answer is not what you hoped.

Reading what arrives and turning it into something a system can act on: orders from messages, line items from supplier invoices, the intent behind a customer question. Sorting: which claim needs a person today, which can wait. Suggesting: the reorder list, the reply to a routine query, the likely reason a delivery was returned. Drafting: the quotation, the follow-up, the summary of a long thread. In every case a person approves, and the system shows why it suggested what it did.

When a flow has to know something only your business knows: whether this customer is on credit hold, which godown has the part, what the reorder level is this month. Those tools move data between apps well, and for a handful of simple flows we will tell you to use them and go home. They become hard to trust when a flow has to make decisions, when it fails silently at three in the morning, or when the twentieth flow depends on the nineteenth. That is the point at which the connections need to be your own system, watched, tested and documented, rather than a subscription nobody dares to touch. When do you outgrow Zapier, Make or n8n? works through the three tests in full.

The test it must pass, agreed in writing and expressed as a number on your own examples: a few hundred real orders or claims, labelled by the people who handle them today, with the ones they disagreed about marked as such. That set defines the job, defines done, and later tells you when the live work has drifted away from it. A vendor who will not show you the test is selling you the demo. Why the test is published before the model is chosen sets out the rest.

Into your own accounts and nowhere else. We build on paid business plans where what you send is not used to train anyone’s model, or on models running in your own cloud account, and every key and credential sits with you. An NDA is signed before you share anything sensitive. If a document must never leave the building, we say whether the job can still be done, and how.

They go to a person, with the reason shown. A system that is unsure and says so is worth more than one that guesses confidently, so the threshold for asking is agreed with you and tuned on real traffic. Every correction a person makes becomes a new example in the test, so the system improves on your work rather than on someone else’s. Accuracy is reported monthly against the test, in plain numbers.

Because the counts were agreed before the build and measured after it, on your data. For most businesses they are plain: orders that reached dispatch without a human retyping them, items sold that were already out of stock, the time between the last bill of the day and the books being closed. If a number has not moved, the review says so in writing, and why. Nothing else is promised, which is why it can be a contract term. The furniture maker is the shape of it: quote to dispatch, every stage confirmed on a phone.

Yes, and it is usually the best starting point, because it tells us what your people actually want the machine to do. What it lacks is what makes it safe to rely on: a test it must pass, a person in the loop where it matters, a record of every decision, and a running cost you have seen in writing. We keep the part that works and build the rest around it, in your own accounts, so it stops being one person’s browser tab and becomes part of how the business runs.