Back to our workCase studies

Conversational AI, and what it actually fixed.

Six problems businesses brought to us on the voice and chat side, and what we built for each one. No jargon: what was going wrong, and what we did about it.

01Voice calls

The AI took too long to answer, so calls felt robotic

The problem

On a phone call, a pause longer than about a second feels wrong. People start talking over the assistant, repeat themselves, or hang up. Replies were taking 1.7 seconds on average and nobody could say which part of the call was actually slow.

How we solved it

We broke 200 real calls down step by step and timed every stage. Turning speech into text was eating two thirds of the wait, and the timer we had been trusting was reporting numbers that were physically impossible. So we fixed the measurement first, then tested six speech providers side by side on real audio, moved to the fastest one, and put every service in the same part of the world as the phone line.

The provider we moved to measures 0.7 seconds against 2.0 and 3.7 for the two we retired.

02Outbound calling

A lead list that had to be dialled by hand

The problem

A few thousand leads is a week of somebody sitting on the phone, and most numbers will not pick up. Whatever gets said on the calls that do connect ends up in a notebook, so at the end of it nobody can tell which leads are worth chasing.

How we solved it

Upload the list and the system works through it. It holds each business to its own cap on simultaneous calls, so nobody floods their own phone lines or gets throttled by the carrier. Busy and unanswered numbers go back in the queue and are tried again later, at a sensible hour in the contact’s own time zone. Every call is recorded, written up, and given an outcome.

The list comes back sorted by who is worth calling, not just used up.

03Website chat

The same questions, all day, answered a day late

The problem

Pricing, opening hours, whether you cover their area. It is the same handful of questions, and the reply lands the next working day, by which point the person has already bought from somebody else.

How we solved it

A chat assistant that sits on the website behind one line of code and answers from the business’s own material. Upload PDFs, Word files, spreadsheets and price lists, or simply point it at the website. It reads them, answers from them, and shows which document the answer came from. Website pages are re-read on a schedule, so when a price changes on the site the answer changes with it.

No answer is invented. If it is not in your material, the assistant says so.

04Lead capture

Plenty of conversations, almost no leads

The problem

The usual fix is a form. A form demanding name, phone and email before the conversation starts converts at around 2 to 3 people in 100, and making the phone number compulsory pushes more than a third of visitors to abandon it on the spot.

How we solved it

We ask one question at a time, inside the conversation, and only once there is a reason to ask. The business picks the questions and the moment they appear: when the chat opens, on the first message, or only when the assistant judges there is something worth following up. Anyone who has already given their details is recognised and never asked twice. The moment a lead is submitted it is pushed straight into the business’s CRM.

Asking this way converts at 15 to 25 in 100, against 2 to 3 for a single upfront form.

05Reliability and cost

Calls that got stuck, and a bill nobody could explain

The problem

Every so often a call would hang instead of ending. It sat there marked as in progress, the recording and the summary never arrived, and it was still counted. The invoice at the end of the month was one number with nothing behind it.

How we solved it

A watchdog runs continuously, spots any call open longer than a real call could be, closes it properly, and still processes whatever did happen so the record is not lost. Anything that fails after the call is retried rather than quietly dropped. Every call is priced on its own, with the listening, thinking, speaking and phone line costs itemised.

Every line on the invoice traces back to a specific call.

06Compliance

Telling people they are speaking to a machine

The problem

In both the US and India, an outbound call made by an AI has to say that it is an AI, and if the call is being recorded that has to be said too. It has to be said before the conversation starts, not buried later on. Getting it wrong is a regulatory problem, not a design one.

How we solved it

Every outbound call opens with a spoken introduction and, whenever recording is on, a recording notice. Those opening lines cannot be talked over: until they have finished, the system ignores incoming audio, so the disclosure is heard in full instead of being cut short by someone saying hello. Businesses can supply their own wording, and inbound calls, where the rules differ, are left alone.

The disclosure is part of the call, not something staff have to remember.

If one of these sounds like your business, the first conversation is about the problem, not the technology.

Have a problem worth solving?

Tell us what is difficult, repetitive, inefficient, or technically challenging. We will help determine what should be built.

Or write to contact@lojits.com with a short description of the problem.