AI automation that pays for itself in month one

Most AI projects fail the same way: someone picks the technology first, then goes looking for a problem worthy of it. The automations that actually pay for themselves are chosen in the opposite order, and they are almost always unglamorous.

Start with the hours, not the model

Before writing any code we ask one question: which task, done by a person, eats the most hours for the least judgement?

Then we measure it. Not estimate — measure. How many of these arrive per week, how long does each take, who does it, and what happens when they are on holiday? That baseline is the whole business case, and it takes a day to collect.

If the answer comes back at two hours a month, we say so and stop. An automation that saves two hours a month will never earn back the cost of building and maintaining it, and hearing that from us early is cheaper than hearing it from your accountant later.

Three shapes that reliably work

Document intake. Invoices, applications, delivery notes, CVs — anything that arrives as a PDF or a photo and gets retyped into a system. Extraction plus validation plus routing removes the retyping, and the validation step is where the real value sits.

Inbox triage. A shared mailbox where a person reads each message and decides where it goes. Classification and routing handle the obvious majority; the ambiguous minority still reaches a human, but a human who is now looking at ten messages instead of two hundred.

Answers grounded in your own documents. An internal assistant that can answer “what is our policy on X” by quoting the actual policy, with a link to the source. The grounding is the point: an assistant that cannot cite where an answer came from is a liability, not a tool.

Notice what is missing from that list — nothing here writes your marketing copy or replaces a member of staff. The wins are in the handling, not the thinking.

Guardrails that keep it trustworthy

Every automation we ship has four things wired in from the start:

• A confidence threshold. Below it, the item goes to a person rather than through the pipeline. Tuning this number is most of the work in the first fortnight.
• A visible audit trail. Every automated decision is logged with its inputs, so a disputed outcome can be reconstructed rather than argued about.
• An escalation path. A named human who receives the things the system will not touch.
• Cost and latency monitoring. Token spend on a dashboard from day one. Bills that surprise people are how AI projects get cancelled.

We also put the data handling in the contract. Enterprise API tiers exclude your content from model training, and that should be written down rather than assumed.

What to skip

Skip anything where being wrong is expensive and being right is unremarkable — final legal wording, medical advice, anything that goes out under someone’s signature without review. Skip processes that change every quarter, because you will spend the savings on maintenance. And skip the automation entirely if a form field, a dropdown, or deleting a step would have solved it. Quite a lot of “AI problems” are really process problems wearing a costume.

The test is simple: if you cannot describe the hours saved in a sentence, it is not ready to build yet.