Less repetitive work.
The right tools for it.

What should happen without someone doing it manually?

If staff repeat the same steps every day, there may be an opportunity to automate part of the task. That could mean transferring data, generating a report, sending a reminder or making information easier to find. Understanding the process and its exceptions comes before deciding what to automate.

AI can extend these possibilities, particularly when a task involves interpreting text or drafting a response. It also introduces uncertainty. I look at whether that additional capability is useful enough to justify the cost, review process and maintenance it needs.

Conventional automation is often enough

Clear rules usually belong in conventional software and API integrations. A scheduled database query can prepare a report. An integration can transfer an approved order. A short script can extract fields from a structured export, check required information and flag duplicates. None of these inherently needs an LLM.

Where ordinary code is cheaper, faster, safer or more reliable, I am comfortable recommending it. The aim is less work for your business, not a particular technology on the invoice.

Where AI may add something useful

• Sorting varied written enquiries into categories for a person to review.
• Summarising lengthy material so staff can locate relevant information faster.
• Drafting answers from an approved source, with human checking where needed.
• Helping people search internal information using everyday language.
• A chatbot for suitable questions, with a clear route to a person.

For example, a support workflow could use an API to retrieve an exact order status and AI to interpret the customer's wording. The factual lookup stays deterministic. If customers simply need a status page, that simpler option may be better than a chatbot.

A scheduled workflow could collect figures overnight, prepare a report and notify a person if data is missing. AI-assisted classification could help triage varied enquiries; summarisation could help staff review lengthy information. In each case, use the simplest reliable solution that solves the problem, with measurement of the time and review effort involved.

Make room for exceptions and human judgement

I would first understand the process, the information involved and what happens when an answer is wrong. A useful trial needs representative examples, including awkward cases, rather than only a demonstration that works once.

We should decide what can happen automatically, what needs approval and how uncertain or failed actions are escalated. For AI, we also need to consider what information may be sent to a provider, whether the source material is reliable and how outputs will be checked. For any automation, monitoring and an owner for failures matter.

Measure the whole task: time spent reviewing and correcting output counts too. An automation that shifts work elsewhere without reducing the overall burden is not much of an improvement. Start with a contained opportunity and expand when the evidence supports it.

Modern tools, established development experience

I was building software long before modern generative AI. Professional development since the 1990s, database and systems work, and around 15 years in experimentation inform how I use today's tools. My existing work areas include AI-assisted development, data processing, information analysis, chatbots and workflows. See the broader experience behind my approach.

Common questions

Do we need AI to automate our business?

No. Many useful automations involve ordinary code, integrations or scheduled tasks. AI is one option when the task calls for its capabilities.

Can everything run without human involvement?

That depends on the consequences and variability of the task. Clear rules may allow full automation of a small step. Uncertain answers or important decisions may need review and an easy way to stop the process.

Can we try a small part first?

Yes. A limited trial can reveal data problems, review effort and likely running costs before you commit to a wider implementation.