Why two quotes for the same brief differ tenfold
Because the brief usually describes the outcome and not the plumbing. “An AI that answers customer emails” is one sentence and at least four different projects, depending on whether it drafts or sends, whether it can see order data, how many systems hold that data, and what happens when it is not sure.
The AI part — choosing a model, writing prompts, handling responses — is a small and fairly predictable slice of the work. The rest is ordinary software engineering: authentication, data shape, error handling, logging, and the tedious business of two systems disagreeing about the same customer.
Which means the right first question to any quote is not “which model do you use”. It is “what exactly does it connect to, and what happens when one of those is down”.
The five things that actually drive the hours
| Factor | Cheap end | Expensive end |
|---|---|---|
| Systems it must reach | One, with a documented API | Several, one of which has no API at all |
| What happens on failure | It asks a human | It must detect, recover and retry correctly on its own |
| How wrong it may be | A draft somebody approves | It acts unsupervised on live records |
| Where the data lives | Already clean and in one place | Three systems that disagree about the same customer |
| Testing material | Real cases you already have | Cases that have to be invented first |
Source: Cybrial’s own scoping factors, written from the builds. No build-cost figures are published here for us or for anybody else — see the sources at the foot of this page.
The third row is the one that most often doubles a project without anybody noticing at the quoting stage. Something that drafts and waits for approval needs a fraction of the testing, logging and safeguarding that something acting alone does — and the version that acts alone is usually what gets described in the meeting and the version that drafts is usually what the business actually wants.
The running cost, which surprises people downwards
Model usage for a small business workflow is normally tens of pounds a month. Reading a few hundred emails, classifying them and drafting replies is not an expensive thing to do at current prices, and prices have moved in one direction for two years.
That changes how you should read a large monthly platform fee. If a provider charges several hundred a month for a wrapper around usage that costs a few pounds, the fee is for the wrapper — and sometimes that is entirely fair, because somebody maintains it, supports it and takes the call when it breaks. Often it is not, and the way to tell is to ask what the underlying usage costs.
The model used here removes the question: the build is quoted from hours, and it then runs on your own account with the provider at the provider’s published price. You see the bill, and nothing is resold at a markup.
What AI integration actually looks like in a small business
The useful examples are unglamorous, which is why they rarely appear in marketing material.
- Reading inbound enquiries and routing them — new job, existing customer, supplier, complaint — with a draft reply attached for the ones that are routine.
- Turning a photograph of a delivery note or an invoice into a row in the system, instead of somebody typing it at eleven at night.
- Answering “where is my order” from the actual order data, rather than from a page about delivery.
- Writing the first draft of a product description from the supplier’s specification sheet, for a person to correct.
- Reading a phone call transcript and writing the follow-up email that nobody ever gets round to sending.
Notice what those have in common. Each is a specific, repetitive job with a known input and a known output, and in most of them a person still approves the result. That is where the value currently is — and it is a long way from the autonomous-agent framing the category is usually sold with.
When the answer is not AI
When the task is deterministic. If the rule is “if the order is over £100, apply free delivery”, that is a rule. Handing it to a model that might decide differently on a Tuesday is paying more for less reliability.
When the data is not there. An assistant cannot answer questions about stock if stock levels live in a spreadsheet on somebody’s desktop. Quite often the honest first project is not an AI project at all — it is getting the data into one place, which then makes several later things possible.
And when the underlying problem is a missing page. If people email to ask the price, the lead time or the service area, the cheapest integration in the world is writing that on the website.
