AI in the places where it actually pays
Most AI projects fail because they start with the technology and look for a problem. We start with a task that is expensive because a person has to read something, and work backwards.
There is an enormous amount of noise around AI in business software right now, and a lot of it is theatre - a chat box bolted onto a product so the marketing page can say AI. It rarely survives contact with real users, because nobody wanted to have a conversation with your software in the first place.
The useful version is quieter. It is a model doing one specific job inside a process that already exists: reading an unstructured enquiry and routing it, summarising a long history so a human can act faster, drafting a reply that a person edits and sends, pulling structured data out of a document that arrived as a PDF. None of that is impressive in a demo. All of it saves real hours.
Where AI genuinely helps
The pattern that works is narrow. A model given one clearly-defined job, with the context it needs and a human checking the output where the stakes justify it. The pattern that fails is broad - a general assistant with access to everything and no defined success condition.
These are the applications we see actually earning their cost:
Reading messy input
Turning free-text enquiries, emails and form submissions into structured records your systems can act on.
Classification and routing
Working out what an incoming item is about and sending it to the right place, at a volume no person could sustain.
Summarising
Condensing long threads, call notes or histories so whoever picks it up next is oriented in seconds rather than minutes.
Drafting
First-pass replies, quote descriptions and job notes for a person to check and send. The human stays in the loop and the blank page problem disappears.
Extracting from documents
Pulling fields out of PDFs, invoices and specifications that arrive in whatever format the sender felt like using.
Search that understands intent
Letting people find things in your own data by describing what they want rather than guessing the exact keyword.
Where it doesn't, and we'll tell you
Language models are probabilistic. That is what makes them useful for messy human input, and exactly what makes them the wrong tool for anything that must be right every time.
We don't use a model to do arithmetic, apply pricing rules, make compliance decisions, or trigger anything irreversible on its own. Those are rules, and rules should be code - testable, predictable and auditable. If a supplier is proposing an AI system that makes those calls unsupervised, that is a reason to be careful rather than impressed.
Calculations and pricing
Deterministic logic in code. A model that is right 97% of the time is unacceptable for an invoice total.
Compliance and eligibility
Anything with a regulatory or contractual consequence needs an auditable rule, not a probability.
Irreversible actions
Sending money, deleting records, publishing to customers. A person approves, every time.
Anything you cannot check
If nobody on your team could tell whether the output was wrong, the output should not be used unsupervised.
Cost, privacy and lock-in
Three practical things get skipped in most AI proposals, and all three cause problems later.
Cost is variable and usage-based, which is unfamiliar territory if you're used to fixed licence fees. We put hard caps and monitoring in from day one so a loop or a spike cannot produce a surprise bill. Privacy matters because your customer data is going to a third party - we're specific about what gets sent, what is redacted first, and what the provider does with it. And lock-in is real: we build behind an abstraction so a model can be swapped when a better or cheaper one appears, which in this market is roughly every few months.
Hard spending caps
Limits at the account and feature level, with alerting well before anything approaches them.
Data minimisation
Only what the task needs gets sent. Personal data is redacted first wherever the job doesn't require it.
Written down
Which provider, which model, what data goes to it, and what their retention terms actually say.
Swappable by design
Model access sits behind an abstraction, so changing provider is a configuration change rather than a rewrite.
Being found by AI assistants
There is a second side to AI that most businesses haven't thought about yet. A growing share of buying research now happens inside ChatGPT, Perplexity, Gemini and Claude rather than on a results page. When someone asks one of those for a recommendation, whether your business appears is not down to traditional rankings.
What those systems favour is content that states facts plainly, is structured so it can be quoted accurately, and is consistent about who you are and what you do across every source they can see. It is a different discipline from classic SEO and it is early enough that being deliberate about it is a genuine advantage. We build it into every site we make - including this one.
Deliverables
Every engagement includes these as standard, not as line items to negotiate.
- One clearly scoped task, with a defined measure of whether it worked
- Human review designed in wherever the stakes justify it
- Deterministic rules kept in code, not handed to a model
- Hard cost caps, monitoring and alerting from day one
- A written record of what data goes where, and why
- Provider abstraction so models can be swapped without a rewrite
Where we have done this
Tuning File Portal
A tuning file service is a queue of skilled manual work where a mistake can damage an engine. It is the hardest kind of process to automate, and the most valuable.
Read the case studyRampDesk
Garages don't run a sales pipeline. They run vehicles through stages. RampDesk is the software we built because nothing off the shelf models that.
Read the case studyQuestions people actually ask
What can AI realistically do for my business?
The reliable wins are all in the same category: jobs that are expensive because a person has to read something. Sorting and routing incoming enquiries, summarising long histories, pulling structured data out of documents, drafting first-pass replies for someone to check. None of it is dramatic but it removes real hours. Anything pitched as a general-purpose assistant that handles everything is usually a demo rather than a system.
Is it safe to send my customer data to an AI provider?
It depends entirely on which provider, which plan and what you send, which is why we write it down rather than hand-waving. We minimise what leaves your systems, redact personal data wherever the task doesn't need it, and check the provider's actual retention and training terms for the specific plan you're on - business tiers usually differ significantly from consumer ones.
How much does running AI features cost?
It is usage-based rather than a fixed licence, which catches people out. A feature processing a few hundred items a month is typically a small running cost; one processing tens of thousands is a real line item worth modelling before you build. We put hard caps and monitoring in from the start so a bug or a spike cannot generate a surprise bill.
Will an AI chatbot on my website help?
Honestly, usually less than people expect. Most visitors want an answer on the page rather than a conversation, and a chatbot that cannot answer becomes a worse version of a contact form. Where they do earn their place is on sites with a genuinely large body of documentation or a complex product catalogue. We'll tell you which situation you're in.
How do I get my business recommended by ChatGPT and other AI assistants?
By stating facts plainly and consistently everywhere those systems can see you, structuring content so it can be quoted accurately, and being unambiguous about who you are, where you operate and what you do. It overlaps with good SEO but is not the same discipline - and it is early enough that being deliberate about it is still a real advantage rather than table stakes.
Where to go next
Business Automation
Every business has a set of tasks that are done the same way every single time by a person. Those are the ones worth automating - and the reason to do it is rarely headcount.
Custom Software
Most businesses outgrow their tools long before they replace them. Custom software is what you build when the spreadsheet, the shared inbox and three subscriptions have quietly become the process.
Tell us what is slowing you down
A short conversation is usually enough for us to tell you what it would take to fix it, and what it would cost you to leave it alone.