What tasks can be automated with AI?

Four questions for any task

Question	A good sign	A warning sign
Is the input clear?	Emails, forms, invoices, records	Hallway conversations, unwritten rules
Does it repeat?	Every day or every week, the same shape	Once a quarter, different each time
Can someone check the result?	A person can see at a glance if it is right	Nobody would notice a mistake for months
What does a mistake cost?	A draft to correct	Money sent, a contract signed, a person refused

A task that passes all four is a good first automation. A task that fails the last one can still be automated, with a person approving each result.

Tasks that usually pass

• Enquiry handling. Messages from email, WhatsApp and web forms read, sorted, matched to the customer, and answered in a draft your team approves.
• Document reading. Invoices, forms and reports turned into clean records your systems can use.
• Moving data between systems. The same details no longer typed twice.
• Routine reports. A daily or weekly summary written from your own records, in the format you already read.

Tasks that should keep a person

• Anything that sends money, signs, refuses or promises.
• Decisions about people: hiring, credit, discipline.
• Anything a regulator would expect a named person to answer for.

An automation can prepare all of these. A person should decide them. We do not offer emotion recognition or gender classification as a feature of anything we build.

The risk in reading other people’s text

An automation that reads email reads text written by strangers, and some of that text will try to give it instructions. OWASP lists prompt injection first among the risks of applications built on language models: it happens when what the system reads alters the model’s behaviour or output in ways you did not intend.

OWASP also names excessive agency, where a system can take damaging actions because it has more functions, permissions or autonomy than it needs. Its advice for both is the same as ours: give each automation only what it needs, and require a person to approve high-impact actions before they are taken. In practice that means drafts, not sends, and narrow access rather than a master key.

How to test one

1. Collect real examples: a few weeks of your own messages or documents, including the awkward ones.
2. Decide in advance what right looks like for each.
3. Run the automation on all of them, and count what it gets right and wrong.
4. Read every mistake. The pattern in the failures tells you whether to fix it, narrow it or leave that task to people.
5. Keep a log of every run once it is live, so anyone can check what it did and why.

Where we stand

SceneScribe is the nearest work we have published: a chain of language model agents that turns a spoken question and a camera image into a short spoken answer. The same engineering can handle an enquiry, an invoice or a daily report. We have no business automation to publish yet, and one will appear here only once a client agrees.

We do not ship a model that has only seen its training set: it is tested on data from where it will run. Our AI automation practice builds this work, joined to the tools you already use.