Services / AI & chatbots
AI & chatbots
The useful version of this is not a robot that chats. It is a member of staff who works nights, never forgets the price list, takes the booking while you are on a ladder, and puts their hand up the moment they are out of their depth. Built on your documents, on your hosting, with every conversation logged so you can read exactly what it told your customers.
PILOT
3–4 weeks
RUNS ON
Your data, your hosting
TAKE A FEEL
The businesses winning on service are not answering faster. They are answering at all.
The gap between a good experience and a bad one is rarely the quality of the answer. It is whether anyone answered — at eight in the evening, on the Sunday somebody was comparing three suppliers, in the ninety seconds before they gave up and tried the next result. That is what organisations are actually using this for, and it is why the numbers move.
Nobody waits until Monday
An answer at 9pm on a Saturday is not a better answer than one on Tuesday morning. It is the difference between a customer and a competitor’s customer.
The repetitive half stops reaching people
Hours, coverage, lead time, parking, do you take card. Handing those over is not about replacing anyone — it is about giving your team back the conversations that actually need them.
Nobody has to repeat themselves
The assistant hands over the whole conversation when it steps aside. Being asked to explain it all again is the single most reliable way to annoy somebody who was already annoyed.
So have a conversation with one
This is not a video or a scripted tour. It is a working assistant built the way we build them for clients: a fixed knowledge base about this studio, a written set of rules, and a visible way to reach a person on every single turn.
Try asking it something it should not answer — the weather, a maths question, what a competitor charges. Watching it decline politely and offer you a human is the most useful thing on this page.
Talk to a person insteadHello. I am a demo of the assistants we build — ask me what Mavs does, or describe a problem and I will tell you how we would approach it.
HOW ONE WORKS
Three steps, and no magic in any of them
Almost every worry people arrive with — it will make things up, it will say something mad to a customer, it will go rogue — is answered by how the thing is actually wired together.
WHERE IT EARNS ITS MONEY
The jobs worth handing over
Not "an AI for your business". A specific job, with a clear definition of done, that is currently costing you either somebody’s evening or a customer you never heard from again. These are the ones we build most, and the ones that pay for themselves fastest.
THE DISTINCTION THAT MATTERS
A chatbot talks. An agent acts.
These two words get used interchangeably by people selling both, and the difference decides your budget, your timeline and your risk. One reads. The other reaches into your systems and changes something.
AI assistants
Two very different things with one name
It answers
Chatbot
Reads your material and replies. If it is wrong, somebody is misinformed — bad, but recoverable.
- Answers from your documentsPolicies, pricing, lead times, coverage. The same forty questions your inbox gets every week.
- Captures the enquiryName, need, contact, and a summary that lands in the inbox or the CRM.
- Deflects the repetitive loadThe measurable win: the share of conversations a person never has to see.
- Live in weeksA pilot in 3–4 weeks, because the hard part is curating what it may read.
- Fails safelyThe worst realistic outcome is an unhelpful answer and an escalation.
It does things
Agent
Connected to the calendar, the payment system, the order database. If it is wrong, something in the real world is now wrong too.
- Books, reschedules, cancelsWriting to the actual calendar, with the double-booking rules enforced in code and not in the prompt.
- Takes payment and depositsThrough your existing checkout. The agent never touches a card number.
- Reads and writes your systemsOrders, CRM records, stock. Every action is a defined tool with a defined limit, never free rein.
- Has a written boundary per actionWhat it may do unattended, what needs a human to approve, and what it must never attempt.
- Is auditable by designEvery action logged with what it did and why, because "the AI did it" is not an answer you can give a customer.
An agent is a chatbot that has been given hands. Everything on the left is a reading problem; everything on the right can create a booking, move money or annoy a real customer, which is why it takes longer, costs more, and needs a boundary written down for every single action.
WHY MOST OF THEM ARE BAD
What separates one people use from one people swear at
Nearly everyone has now had a miserable conversation with a support bot. The failures are consistent, they are all design decisions rather than model limitations, and they are the whole job.
A visible way out, on every turn
The single most infuriating thing a bot does is trap you. Ours offers a human from the first message and never hides the exit. Counter-intuitively this raises deflection: people stop fighting it once they know they can leave.
It says "I do not know"
A model with nothing relevant retrieved will happily invent something confident. Refusing is a behaviour you have to build and then test, against real questions from the real inbox, before launch.
Narrow beats clever
An assistant that does four jobs properly beats one that attempts forty. We scope it to what it can be trusted with, and let it decline the rest gracefully.
It writes like your business
Not chirpy, not apologetic, no exclamation marks unless that is genuinely you. Tone is not decoration — it is what makes the difference between representing you and embarrassing you.
Somebody reads the transcripts
Every week for the first month, then monthly. The conversations tell you what customers actually ask, which fixes the assistant and usually the website too.
It is measured on escalations, not chats
Chat volume is a vanity number. Deflection rate, escalation rate and how often a human had to correct it are the three that decide whether it is working.
SOUND FAMILIAR
What we would actually build
Six real situations, and the specific thing we would put in place. None of them is "an AI for your business".
A dental practice. Half the calls come after six, and reception is already flat out during the day.
WHAT WE WOULD DO
An after-hours assistant on the site and WhatsApp: answers treatment and pricing questions from the published list, offers real appointment slots, takes the deposit, and escalates anything clinical to the morning list with the transcript.
A trades business. The phone rings while you are under a sink.
WHAT WE WOULD DO
A missed-call responder that texts back within seconds, asks what the job is and where, and offers the next three slots. Most of the value here is speed, not intelligence — the first to reply usually wins the job.
An online shop drowning in "where is my order".
WHAT WE WOULD DO
An assistant wired to the order system, answering status questions in seconds and only involving a person when the news is bad. Typically the largest single deflection win we see, and the quickest to build.
A B2B firm where every enquiry arrives as three words and an email address.
WHAT WE WOULD DO
A qualifying assistant that asks the five questions your salesperson always asks, scores the enquiry, routes it, and writes the summary straight into the CRM. Sales open a qualified brief instead of a guessing game.
A professional services firm sitting on fifteen years of documents.
WHAT WE WOULD DO
Retrieval over your own files, for staff rather than customers. Usually not a chat window at all — a search box that finally returns the right document with the paragraph highlighted.
You want it to handle complaints and refunds unattended.
WHAT WE WOULD DO
We would say no, and mean it. Complaints are exactly where a wrong answer is expensive and a human is cheap. The assistant should recognise one instantly, stop, and hand it over with everything already written down.
What building one involves
Week one: deciding what it may know
The longest part of the job is not the model. It is agreeing which documents are in scope, finding the three that contradict each other, and getting a straight answer on the pricing that has never been written down anywhere. Clients are always surprised that this is the work.
Weeks two and three: building the boundaries
What it answers, what it escalates, what it is allowed to do unattended, and what it must never attempt. Then testing it against a hundred real questions pulled from your actual inbox, including the rude ones and the trick ones.
After launch: reading what it said
Transcripts reviewed weekly for the first month. Every bad answer is either a missing document, a boundary in the wrong place or a genuine gap in what your business has ever written down — and all three are worth knowing.
We will talk you out of it
A fair number of the AI projects that reach us are not AI projects. They are a search problem, a form, a notification, or a report nobody had time to build. Those are cheaper, faster and far more reliable to solve directly, and we will say so before you spend anything. The test we apply is simple: if you cannot describe the job in one sentence, with a clear definition of done and an obvious thing to do when it goes wrong, it is not ready to be handed to a model. When one genuinely is the right tool, most of the design work is boundaries — what it may answer, what it must escalate, and how you find out it got something wrong before your customer tells you.
Questions we get asked
Where does our data go?
Into infrastructure you control. We scope which documents are in play before anything is indexed, and by default nothing you send is used to train a third-party model. If you have a data-residency requirement, say so early — it changes the architecture, not the price.
What stops it making things up?
It answers from retrieved passages rather than memory, and cites which document each answer came from. When it finds nothing relevant it says so and offers a human instead. That behaviour is tested before launch against a set of real questions from your inbox.
How do we know whether it is working?
Every conversation is logged and reviewable. The numbers that matter are deflection rate, escalation rate and how often a human had to correct it — all three land in your monthly report. Chat volume is not one of them.
Will it annoy our customers?
It will if it is built badly, and most are. The fixes are all design rather than technology: offer a human on every single turn and never hide that exit, keep the scope narrow enough that it is genuinely good inside it, and let it say plainly that it does not know. People forgive a bot that cannot help. They do not forgive one that will not let them leave.
Can it actually book something, or just talk about booking?
It can book — into your real calendar, with your real rules about lead time, double-booking and deposits enforced in code rather than in the prompt. That is an agent rather than a chatbot, and it is a bigger piece of work because every action it can take needs a written boundary and an audit trail.
What happens when it gets something wrong?
You will see it, because you can read every conversation. Answers are corrected by fixing the source document rather than by arguing with a prompt, which means the fix also helps the humans reading the same document. Anything it can do rather than say is reversible or requires approval, by design.
Tell us what you're building.
Send a paragraph. You'll get a written point of view back within one working day — no deck, no discovery fee.
Start the conversation