Services

Five kinds of AI system, and the month after.

Most of what eats a week falls into one of these. We build automations across all five, then keep them running, which is the part that decides whether any of it was worth doing.

contact@zhorx.com See a build in detail

What we build

AI agents

For the work that cannot be written down as rules, and only for that.

Some jobs need a decision, not a step. Working out what an enquiry is actually asking for. Pulling figures out of a supplier document that never looks the same twice.

That is where a model earns its keep. Everywhere else a plain rule is cheaper and easier to trust. When we do put an agent in, it has a confidence threshold, a human check on anything it is unsure about, and a log of what it decided and why.

Typical work: Enquiry triage and routing, document and invoice extraction, inbox summaries, classification, first-draft replies
Built with: Current frontier models, with the fallback path always defined
Usually live in: Two to five weeks

AI agents

For the work that cannot be written down as rules, and only for that.

Some jobs need a decision, not a step. Working out what an enquiry is actually asking for. Pulling figures out of a supplier document that never looks the same twice.

Typical work: Enquiry triage and routing, document and invoice extraction, inbox summaries, classification, first-draft replies
Built with: Current frontier models, with the fallback path always defined
Usually live in: Two to five weeks

Custom workflow automation

If you could write the steps on a card, it does not need a person doing it every week.

Almost every business runs four or five tools that do not know about each other, and the gaps between them are filled with people copying and pasting.

We close the gaps and hand the repeatable steps to a workflow you can look at. Every run is logged, every step is visible, and the whole thing exports. Automation you cannot inspect is a different kind of dependency.

Typical work: CRM to accounts, forms to CRM, invoice and quote chasing, client onboarding, weekly reporting, document filing
Built with: Your own infrastructure or ours, your choice
Usually live in: Two to four weeks

Chatbots

The same eight questions, answered instantly, in your wording.

Most of a support inbox is repetition: opening hours, order status, what is included, where the invoice went. Making a customer wait until Monday for one of those costs more than it looks.

The bot is trained on your own material rather than the open internet, so it answers the way you would. Anything outside what it knows goes to a person with the conversation attached, and it says plainly that it is a bot.

Typical work: Website and WhatsApp support, order and booking status, qualification before a human picks up, internal help desks
Built with: Your documents and past replies as the source, retrieval over a model, escalation into your inbox or CRM
Usually live in: Two to four weeks

Voice agents

Every call picked up, including the ones at ten at night.

A missed call is usually a lost job, and most businesses miss plenty: on site, out of hours, or because two rang at once. Voicemail does not fix it, because nobody leaves one any more.

A voice agent answers, works out what the caller wants, books it or takes the details, and writes it into your systems before the call ends. Anything it cannot handle is transferred or logged with a transcript.

Typical work: Out-of-hours answering, booking and rescheduling, order status, qualification and routing, callback capture
Built with: Speech models over your booking and CRM systems, with transcripts kept against every call
Usually live in: Three to six weeks

Marketing agents

The channel that goes quiet first when a week gets busy.

Marketing is the thing that slips, and the gap shows up in the pipeline two months later. Writing, scheduling, chasing and reporting all land on somebody who already has a day job.

The agents draft, schedule and follow up on a rhythm that does not depend on anyone having a spare hour. You approve, you do not produce, and you see everything before it goes out.

Typical work: Post and newsletter drafting, scheduling, lead nurture sequences, review requests, monthly performance reporting
Built with: Your CMS, email platform and analytics, with a human approval step before anything publishes
Usually live in: Two to four weeks

The difference

AI and automation are not the same thing.

Most of what a business needs is the second one. The interesting part is knowing which is which.

Automation
Follows rules.
When this happens, do that. It does not think, and that is the point: it is exact, it is cheap, and it does the same thing at 3am as it does at noon.
For example
1. Invoice paid
2. Mark the job closed
3. Tell the office
AI
Understands.
It reads something written by a person and works out what it means. Useful exactly where a rule falls over, which is anything phrased a hundred different ways.
For example
1. An email arrives
2. Work out what they want
3. Sort it
Both together
Understands, then acts.
The model reads and decides. The workflow does the ten boring things that follow, in the right systems, without waiting for anyone.
For example
1. Reads the enquiry
2. CRM, reply, diary

AI thinks. Automation acts.

Part of the job is telling you which one your problem actually needs. Sometimes the honest answer is a rule, and it costs less to run.

One path in full

One enquiry arrives. Nine things happen.

01
TriggerEnquiry form submitted
02
02AI reads what they asked for
03
03Intent identified
04
04Lead created in the CRM
05
05Qualified against your rules
06
06Reply written and sent
07
07Salesperson notified
08
08Follow-up sequence starts
09
ResultAppointment booked

One trigger. Nine actions. Zero copy and paste.

The monthly service

The build is the easy half.

An automation is not finished when it works on a Tuesday. It is finished when it has survived a month of your business being itself: the odd file, the supplier who changed something without telling anyone, the week where volume tripled.

We price for that on purpose. There is a fixed fee to build, and then a flat monthly fee to run it. Both are agreed before we start. If we quote wrong, that is our problem, not an invoice you weren’t expecting.

Daily monitoring: Every workflow is watched. Failures raise an alert to us, not to you.
Fixes: When something breaks, say an API changes or a file format moves, fixing it is part of the fee.
Changes: Your process moves. New step, new tool, new person. We adjust it. Not a new quote.
A monthly report: What ran, what broke, what we changed, and what we would do next.
A person who knows your system: Whoever built it picks up. You don’t explain it again to a support queue.
Your data and your workflows: Documented and exportable. Leaving is a decision you can make in a month, with no exit fee.

What it costs

Two numbers, both fixed before anything starts.

The build

A one-off fee, quoted from the plan we write after looking at how you work. It does not move once agreed. If the job turns out bigger than we thought, that is a lesson for us.

The month

A flat fee covering monitoring, fixes, changes and the report. Not per task, not per run, not per seat. It scales with the size of the system, not with how much you use it.

The ranges each shape of project lands in are published, and the fixed number comes after we have seen the process.

See the ranges Ask about your own

By function

By sector

Five kinds of AI system, and the month after.

What we build

AI agents

AI and automation are not the same thing.

Follows rules.

Understands.

Understands, then acts.

One enquiry arrives. Nine things happen.

The build is the easy half.

Two numbers, both fixed before anything starts.

The build

The month