AI automation for businesses

Repetitive work handled inside the workflow it belongs to: documents turned into records, drafting and research, with a person still deciding the exceptions.

AI is most useful inside the workflow it is meant to help: reading documents, drafting, researching and deciding what happens next. Not a chatbot bolted onto the homepage.

What can be automated

Work that happens often, looks roughly the same each time and today is done by someone moving information between two places. Four places where there is almost always work to remove:

• Enquiries. An enquiry by form, email or phone should become a case, a row in a list and a reply. What gets automated is the logging, the confirmation to the customer and the reminder when nobody has replied. You still write the reply yourself.
• Reports. Someone exports from a system, pastes into a spreadsheet, formats it and sends it, the same steps every week. The report becomes a by-product of the numbers already being in one place.
• Documents. Invoices, delivery notes, contracts and forms that arrive as PDFs or images and need to become fields in a system. The machine reads and suggests, and a person approves what is uncertain.
• Systems that don't know about each other. Booking system, point of sale, bookkeeping, email and spreadsheets that each work on their own. What gets built is a link that moves the right information in the right direction, with a log of what went wrong.

Where there is usually most to gain is in What can a small business automate?.

Rules, AI and a person

Rules fit when the answer can be written down in advance. They are cheap to build, easy to test and never do anything unexpected. AI fits when the material varies: free text, documents that look different, summaries. It costs more to verify, because the answer is not always the same. Most solutions end up as a mix: rules for what is predictable, AI for what is messy, and a person for what is expensive to get wrong.

AI sometimes gets things wrong, so the systems are built around that. They score their own confidence, escalate uncertain cases to a person and log what they did. A model that guesses quietly is a liability.

What is included

An automation sprint takes one process end to end: scoped, built, documented and handed over.

• One workflow, one integration or one AI automation.
• A fixed scope and a fixed price.
• Typically one to two weeks.
• Documentation and handover included.

If you keep finding things worth automating, there is the ongoing partner: an agreed amount of capacity each month, priority response and continuous improvements, with 30 days' notice.

How it works

First a free call about what takes time, and whether it is worth automating at all. Then you get the scope and a fixed price in writing before anything is built. The whole way of working is on the home page.

What you need to contribute is access to the systems to be connected, real examples including the awkward ones, a decision about what should happen when the automation is unsure, and one person who owns the process and can answer questions during the build. Twenty real cases say more about what is needed than a description of how it usually looks.

Price

Common questions

When should a person still look?

For decisions that cost money or affect a customer directly, when the material is unclear or contradicts itself, when personal data appears in sensitive contexts, and for the first weeks after deployment, however good it looked in testing.

When is it not worth it?

When the work happens a few times a year, when the process keeps changing, when nobody owns the process, or when the material is too poor. Rubbish in gives rubbish out, faster than before.

How long does an automation take?

A single automation or integration is usually one to two weeks. You get a fixed scope and a date before anything starts.

Who owns the automation?

You do. The code is delivered into your repository with documentation and runs on your accounts. I work under an NDA by default when data is involved.