What does a real AI system in a business actually consist of?

Five layers: a data layer that connects the systems where work already lives, a model layer that reads and drafts, a workflow layer that carries the output into your real software, a control layer that keeps humans approving anything consequential, and a deployment layer that monitors what is running. The model is the smallest of the five.

Do I need to move my data somewhere new before using AI?

No. In most builds the systems you already pay for stay exactly where they are and the AI system reads from and writes back into them. What does have to happen is record matching — making sure one customer, job or invoice is recognisable across those systems — otherwise the AI produces confident but wrong answers.

Which AI models are used, and does it matter?

Different jobs use different models: a fast, cheap one for sorting and tagging, a stronger one for drafting and reasoning over documents. The choice matters less than most people expect, and it can be changed later — the value sits in the connections, the workflow and the approval design around it.

How is an AI system deployed without disrupting the business?

In stages. It runs in shadow mode first, producing output nobody acts on while the team checks it. Then it becomes assisted — the system prepares, a person approves. Only once the routine cases are consistently right does it handle them on its own, with exceptions still routed to a human.

What does it cost to run an AI system once it is live?

There are four ongoing lines: model usage (usually the smallest, often tens of pounds a month for a few thousand documents), hosting the workflows, monitoring and audit storage, and optional support. The build is the significant spend; the running cost is normally a modest monthly figure.

How do I know an AI system is safe to let loose on customers?

It should not be let loose. Anything priced, contractual or customer-facing keeps a named human approver, low-confidence cases escalate instead of guessing, every action is logged with the evidence it used, and each workflow has its own off switch.

What a real AI system looks like inside a business

Not a chat window. A working system has five parts: the data it reads, the models that read it, the workflow that carries the result into your real software, the approvals that keep a human in charge, and the monitoring that proves it is still working. This page walks through all five, in the order they get built.

Take the free 2-minute scan Book the £499 AI Ops Audit

The short answer

A real AI system reads your own data, uses a model to turn messy input into structured output, writes that output back into the software you already use, pauses for a human wherever the decision carries consequence, and is monitored once live. The model is perhaps a tenth of the work and usually the smallest line on the bill — the connections, the workflow and the approval design are the system.

The five layers

1. Data layer

Where the facts come from

Nothing works until the system can read the same information your team reads. That means connecting the places the work already lives: the shared inbox, the accounting system, the CRM or job system, the document folders, the spreadsheets people actually maintain. Records get matched to one customer, one job, one invoice, so the system is not guessing which 'Smith' it is looking at.

Read access to the systems you already pay for — no migration
One identity per customer, job and supplier so records line up
A store of past jobs, quotes and documents the AI can look things up in
A record of where every fact came from, so any answer can be traced

2. Model layer

The part that reads, drafts and classifies

The model is the smallest part of the system and rarely the interesting one. It reads unstructured text — emails, PDFs, forms, notes — and turns it into something structured: this is an enquiry, it is about this job, it needs a quote, here is the draft. Different jobs get different models: a fast cheap one for sorting and tagging, a stronger one for drafting and reasoning over documents.

Extraction: pull the fields out of an email, invoice or form
Classification: what kind of enquiry is this, and who owns it
Drafting: a quote, a reply, a job pack, a weekly summary
Retrieval: answer from your own documents rather than from the internet

3. Workflow layer

Where the work actually happens

This is the bulk of any real build. The model's output has to land somewhere useful: a CRM record updated, a quote created in your quoting tool, a task assigned, a chase email queued, a spreadsheet replaced by a live dashboard. Each step has a rule for what happens when something is missing, ambiguous or unusual.

Triggers — a new email, a new job, a date passing, a status change
Steps in order, each one writing back into your real systems
Explicit handling for the awkward cases, not just the happy path
Retries and alerts when an upstream system is down

4. Control layer

What stays human

Anything that carries money, safety, legal or reputational consequence keeps a named human approver. The system prepares the work and shows its evidence; a person presses send. Where the model is unsure, it escalates rather than guessing — and it never quietly invents a fact it could not find.

Approval steps on anything priced, contractual or customer-facing
Confidence thresholds — low confidence routes to a person
A full audit trail: what it did, when, on what evidence
An off switch per workflow, not an all-or-nothing shutdown

5. Deployment and run

Live, monitored, maintained

Going live is a pilot, not a launch. The system runs alongside the current process while the team checks its output, then takes over the routine cases with the exceptions still flowing to humans. Once live it needs monitoring — failures, queue backlogs, model costs, and the hours actually saved against the hours forecast.

Staged rollout: shadow mode, then assisted, then routine cases automated
Monitoring and alerts when a workflow fails or falls behind
Usage and cost tracking so the running bill holds no surprises
A quarterly review against hours saved, not against activity

How one enquiry moves through it

The same five layers, followed end to end for a single job.

01An enquiry lands in the shared inbox — the trigger.
02The system reads it and pulls out the fields: who, what, where, when, budget signals.
03It matches the sender against existing customers and past jobs so context comes with it.
04It classifies the enquiry and routes it to the right person or queue.
05It drafts the reply or quote from your own pricing and previous jobs.
06A human reviews the draft, adjusts the price and approves it.
07On approval, the CRM record, the job system and the follow-up schedule all update.
08If the customer goes quiet, the chase sequence runs — and stops the moment they reply.
09Everything above is logged, and the weekly numbers assemble themselves from it.

Only steps 2–5 and 7–9 are automated. The price still gets approved by a person, and anything the system is unsure about goes to a human instead of being guessed.

Where the money actually goes

Part of the system	Share of cost	What it covers
Data connections	Often the largest slice of the build	Connecting and matching records across the systems you already use.
Workflow build and testing	The bulk of the remaining build	Steps, edge cases, approvals, retries — the part that makes it survive real use.
Model usage	Usually the smallest line	Reading a few thousand emails or documents a month typically costs tens of pounds.
Hosting and monitoring	A modest monthly figure	Running the workflows, storing the audit trail, alerting when something breaks.
Support and improvement	Ongoing, optional	Tuning as volumes change and extending to the next process.

Figures for your own case come out of the £499 AI Ops Audit — it prices the first build against your real volumes rather than a generic range. See the pricing page and the project roadmap for the wider shape.

Four things people get wrong

"We just need ChatGPT for the team."

A chat window is a tool for a person. A system does the work whether or not anyone opens a tab — it triggers on real events and writes back into your real software.

"The AI is the expensive bit."

Model usage is normally the smallest line on the bill. Integration, testing and approval design are where the cost and the value both sit.

"We'd have to replace our systems first."

The opposite. A system that reads and writes into the software you already use is faster to build, cheaper to run and far easier for the team to accept.

"It either works or it doesn't."

Real systems run on a spectrum: fully automated for routine cases, assisted for the ones needing judgement, fully human for the rare and risky.

See one that is running

The clearest way to understand the five layers is to look at a system in use. The Kerbco permit platform is ours, live, and built exactly this way — connected data, models doing the reading, humans approving what matters.

The Kerbco build, step by step

Phases, what is automated, what stays human, and the results.

All Forth projects

The platforms and systems we have designed, built and run.

Questions people ask

What does a real AI system in a business actually consist of?: Five layers: a data layer that connects the systems where work already lives, a model layer that reads and drafts, a workflow layer that carries the output into your real software, a control layer that keeps humans approving anything consequential, and a deployment layer that monitors what is running. The model is the smallest of the five.
Do I need to move my data somewhere new before using AI?: No. In most builds the systems you already pay for stay exactly where they are and the AI system reads from and writes back into them. What does have to happen is record matching — making sure one customer, job or invoice is recognisable across those systems — otherwise the AI produces confident but wrong answers.
Which AI models are used, and does it matter?: Different jobs use different models: a fast, cheap one for sorting and tagging, a stronger one for drafting and reasoning over documents. The choice matters less than most people expect, and it can be changed later — the value sits in the connections, the workflow and the approval design around it.
How is an AI system deployed without disrupting the business?: In stages. It runs in shadow mode first, producing output nobody acts on while the team checks it. Then it becomes assisted — the system prepares, a person approves. Only once the routine cases are consistently right does it handle them on its own, with exceptions still routed to a human.
What does it cost to run an AI system once it is live?: There are four ongoing lines: model usage (usually the smallest, often tens of pounds a month for a few thousand documents), hosting the workflows, monitoring and audit storage, and optional support. The build is the significant spend; the running cost is normally a modest monthly figure.
How do I know an AI system is safe to let loose on customers?: It should not be let loose. Anything priced, contractual or customer-facing keeps a named human approver, low-confidence cases escalate instead of guessing, every action is logged with the evidence it used, and each workflow has its own off switch.

Want to know which layer your business is missing?

The free two-minute scan tells you where you'd actually start. The £499 AI Ops Audit maps one process properly and prices the first build — and it's credited against any work that follows.

Take the free 2-minute scan Back to the AI adoption hub