AI in your systems: what it’s actually good for
AI is a component. The system around it is the product. Most of the value in what people call an “AI project” comes from work that has nothing to do with the model — defining the process, connecting the systems, and deciding who approves what.
What it is genuinely good at
- ✓Reading messy, unstructured text and understanding what it says
- ✓Sorting things into categories — this is a quote request, this is a complaint
- ✓Drafting a first version of a reply, a quote, or a summary
- ✓Pulling specific fields out of a document, an email, or a form
- ✓Summarizing a long thread into what actually needs a decision
Those are real capabilities and they save real hours. Notice that every one of them is a step inside a process — not a process on its own.
What it should not be trusted with
- ✕Arithmetic you need to be right — pricing, totals, tax
- ✕Rules that must never vary, like who is allowed to approve a discount
- ✕Anything where being confidently wrong is expensive
- ✕Deciding what a customer is owed
- ✕Being the only thing standing between a mistake and a customer
We write ordinary software for those. It is cheaper, it is predictable, and it does the same thing every time — which is the whole point.
Plenty of the systems we build contain no AI at all
If the process is not written down anywhere, AI will make it faster to produce a mess. Map it first — and quite often the mapping alone fixes the problem for free.
If there is no history to check the output against, there is no way to know whether it is right, and “it looks reasonable” is not a standard anyone should run a business on.
If a plain rule would do the job, use a plain rule. A lookup table that is always right beats a model that is usually right, every time.
Telling you that costs us the bigger project. It is still the right answer, and it is a large part of why people call back.
The work we actually build with it
Each of these is a system with an AI step inside it, not an AI product. The surrounding work — capture, routing, records, approval, measurement — is where most of the value is.
Lead capture and follow-up
Enquiries classified and answered in minutes, routed to the right person, followed up on a schedule.
Customer reply triage
Inbox sorted, replies drafted in your voice, complaints and refund requests escalated immediately.
Quote and estimate drafting
Draft quotes assembled from your own pricing rules for staff to review and send.
Order processing
Routine orders flow through; payment, stock, and address exceptions are flagged for a person.
Reporting
Operational digests pulled from your existing tools, reconciled to source, with anomalies flagged.
Procedures and internal knowledge
Real documentation captured from how the work is done, plus an assistant that answers only from it.
Why most AI rollouts quietly stop
Not because the technology failed. Because nobody could say what it was supposed to do, and so nobody could tell whether it was doing it.
- ✕A licence gets bought before anyone defines what work it is doing
- ✕Output is accepted because it looks reasonable, not because it was checked
- ✕There is no number, so a month later nobody can say if it helped
- ✕Nobody owns the approval step, so mistakes reach customers unnoticed
- ✕It is switched on for everyone before it has worked for anyone
The questions people actually ask
Find out what the manual work is costing you.
The assessment gives you the map, the number, and a ranked list of what to fix — priced before you commit to any of it. If it doesn’t find opportunities worth more than it costs, you don’t pay for it.