Notes on AI, automation and building systems that last.

Case study · AI Agents

How we built the AI Voice Agent

A phone line that answers itself, understands the caller and keeps a record of the conversation — built as one automation rather than a stack of products.

The problem

A business with a published phone number gets calls at inconvenient times: during a meeting, after closing, or three at once. Every unanswered call is a potential customer with no way in, and the answered ones leave no searchable trace of what was discussed.

How the system works

The call reaches a webhook, which passes the audio to Whisper for transcription. That transcript goes to an OpenAI chat model with instructions to return a structured result: the caller's intent, a confidence score, a written response, the next action and the language. Because the output is structured rather than free text, the rest of the workflow can branch on it reliably.

A condition checks the parsed result and decides what happens next. The written response is converted to speech using a text-to-speech voice and played back to the caller, so the conversation continues in audio. In parallel, the call is appended to a Google Sheet with the timestamp, caller number, intent, duration, status and next action.

What it delivers

• Calls answered at any hour, and more than one at a time
• Natural spoken replies instead of a menu tree
• Intent captured on every call, with a confidence score
• A complete, searchable call log without manual note-taking
• A base that extends into booking, qualification and follow-up

Built with

AI Automation

How AI automation saves businesses time and money

Automation pays off when it removes work that repeats. The skill is in spotting which work that is.

Look at a normal week and most of it splits into two kinds of work. There is the work that needs a person — deciding, negotiating, designing, understanding a customer's situation. And there is the work that follows a pattern: reading an enquiry and deciding where it goes, copying a form response into a spreadsheet, sending the same three-line reply, checking whether something arrived.

The second kind is where automation belongs. What has changed recently is that automation is no longer limited to rigid rules. A language model can read an unstructured email or form answer, work out what it means, and return a clean result the rest of the workflow can act on. That is the difference between "if the subject contains the word invoice" and "this message is a billing question with high urgency".

What to automate first

• Anything that happens more than a few times a day
• Anything where information is retyped from one tool into another
• Anything with a predictable response, even if the input is messy
• Anything that stalls when one specific person is busy

How the saving actually appears

Time comes back in two forms. The obvious one is minutes per task. The bigger one is attention: a person who is not monitoring an inbox all day can concentrate on work that needs concentration. There is a cost saving too, but it is usually about capacity — handling more volume without adding headcount — rather than replacing anyone.

A fair test before building: write down how often the task happens, how long it takes, and what goes wrong when it is missed. If the answer to the last question is "we lose a customer", it is worth automating first.

AI Agents

What AI agents are, and what they can do for your business

An agent is software that understands a request, decides what to do about it, and does it.

A traditional chatbot follows a script. It matches keywords, offers buttons, and fails as soon as the person phrases things unexpectedly. An AI agent works differently: it interprets the message, forms an understanding of the intent, looks up whatever information it needs, and produces a response. When it has permission, it can also take an action — create a record, send an email, schedule something, notify a colleague.

What a useful agent needs

• Grounding: answers drawn from your own documentation, not general knowledge
• Structure: results returned in a fixed format the workflow can act on
• Confidence: a score that lets weak answers escalate instead of guessing
• Memory of the exchange: a stored record of what was asked and answered
• A handoff path: a clear route to a human when the agent shouldn't decide

Where agents earn their place

Support is the common starting point, because the same questions repeat. Sales is next: an agent can qualify an enquiry, summarise it and route it to the right person while the lead is still warm. Voice is the newest area — an agent that answers the phone, transcribes the caller and replies in speech.

The boundary matters as much as the capability. Decide up front what the agent may answer on its own, what it must escalate, and what it is never allowed to do. That boundary is what makes an agent safe to put in front of customers.

Workflow Automation

How workflow automation improves business efficiency

The slowest part of most processes is the gap between two tools.

A lead arrives in a form. Someone opens it, copies the details into a spreadsheet, messages a colleague, and adds a reminder to follow up. None of those steps is difficult, but together they add hours across a week and every one is a chance to forget something.

Workflow automation replaces the gap with a defined path. A trigger starts the flow — a form submission, a new email, a scheduled time. The data is validated and formatted. Rules decide which branch it takes. The relevant systems are updated. The right person is notified. Every run is recorded.

The parts that make it reliable

• A single trigger, so it is obvious what starts the process
• Validation before anything is written to a live system
• Explicit branches instead of one flow that tries to cover everything
• Error handling and retries for the steps that touch other services
• Logs that show what ran, what it decided and what it produced

Efficiency is the visible outcome, but consistency is the deeper one. A workflow does the same thing on a Monday morning and at midnight on a Friday, whoever is on holiday.

Ecommerce

Ecommerce automation: how to scale an online store

Growth in ecommerce usually means more of the same work, not different work.

Every extra order brings the same follow-on tasks: a confirmation, a shipping update, a "where is my order" message, a stock number to adjust, a review request. At low volume these are fine to do by hand. At higher volume they become the whole job, and they are the first things to slip during a busy period.

What to automate in a store

• Order confirmations and status updates to the customer
• Answers to the repeat questions: delivery, returns, sizing, availability
• Low-stock alerts and inventory synchronisation
• Product content workflows: descriptions, listings, images
• Post-purchase follow-ups, review requests and abandoned cart messages

Keeping it human

Automation should not make the store feel like a machine. Two rules help: automate the message that is genuinely the same every time, and always leave a visible route to a person. An automated reply that ends with a real way to reach support is a better experience than a slow human reply three days later.

Start with the single question your inbox receives most. Automating that one is usually enough to notice the difference in a week.

Customer Support

AI customer support automation: what you need to know

Support is the easiest place to start with AI, and the easiest place to get it wrong.

It is the easiest place to start because the questions repeat and the answers already exist somewhere — in a help doc, an old email thread, a guide nobody has updated in a year. It is easy to get wrong because a confident, incorrect answer damages trust faster than a slow one.

How a support automation is put together

A question arrives from a chat widget, website form or inbox and is cleaned into a consistent structure. A model classifies it — intent, category, urgency — and drafts an answer from your documentation, returning a confidence score and the sources it used. Above a confidence threshold the answer is sent automatically. Below it, the question goes to a person with the draft attached. Either way, the exchange is written to a record and the team is notified.

What makes it work in practice

• Answers restricted to your own material, with sources shown
• A confidence threshold agreed before launch, not after a complaint
• An escalation path that a customer can reach in one step
• A shared log so the team sees what the agent has been saying
• A review loop: the questions it handles badly become new documentation

Done properly, support automation clears the repetitive layer and leaves your team the conversations that actually need them.