Most AI pilots fail in a recognisable way. Something impressive is demonstrated, everyone agrees it is the future, and six months later nobody uses it because it did not remove a cost anyone was actually paying. The tools that survive are narrower and less exciting: they take one repetitive task that currently consumes a person's day, and they do it inside the systems the business already runs.
We are building a set of these now. They fall into three patterns, and the patterns are more useful than the product names, because they tell you where to look in your own operation.
Pattern one: turning messy inbound into structured records
Enquiries arrive as voice notes, photographs, forwarded emails and WhatsApp messages. Someone reads them, works out what is missing, asks for it, and retypes the result into a system. That person is doing transcription, not judgement, and it is the single most common AI-shaped hole in a small business.
- QuotePilot takes an incoming enquiry, asks for the dates, headcount or specification that are missing, applies your saved rate sheet, and drafts a complete quotation for a human to approve. Built first for travel and events agencies.
- OrderCapture reads a purchase request sent over WhatsApp, matches it against your product catalogue, clarifies quantity or pack size, and creates a draft order in Odoo. Built for B2B distributors.
- ServiceDesk AI turns a customer's photo and voice note into a job card with symptoms, location and preferred time, ready to dispatch. Built for air-conditioning and appliance repair firms.
- DocReady collects a complete application without staff chasing clients: a branded upload link, a checklist, expiry checks and automatic reminders for what is still missing.
- TenderStart reads an RFQ and tells a small supplier whether it is worth pursuing, what the deadline is, what is excluded, and what a response has to contain.
- ParentFlow answers routine school enquiries from approved responses, handles absence reports and admissions checklists, and escalates anything it should not answer.
The common shape is worth noticing: each one ends with a draft that a person approves. None of them send anything on their own, and that is deliberate.
Pattern two: watching the numbers nobody has time to watch
Every operation leaks in places that are individually too small to investigate and collectively significant. A margin that drifts on one product line. A variation on site that was instructed but never priced. A return pattern on one customer that nobody has aggregated. These are not hard questions; they are questions nobody gets to.
- MarginShield AI watches margin at transaction level and raises the cases where it has moved, before the quarter closes rather than after.
- VariationGuard AI tracks instructed variations on a project against what has been priced and claimed, which is where specialist contractors most often lose money they have already earned.
- ReturnSense AI finds the pattern behind returns for retailers and distributors: which product, which customer, which reason, repeatedly.
Pattern three: an assistant that works across the whole system
The first two patterns automate a task. The third changes how people reach information at all. These are heavier tools, intended for organisations with enough volume to justify permissions, audit history and ongoing monitoring rather than a one-off report.
- Tarynix ERP Analyst investigates operational questions across Odoo or another ERP, ties every conclusion back to the actual records behind it, flags anomalies, and prepares actions for approval.
- Procurement Control AI compares supplier quotations against the RFQ, the contract, previous prices, delivery terms and approval limits, then produces a decision pack and a draft purchase order.
- Project Evidence AI organises project correspondence, RFIs, site reports, variations and approvals into a searchable timeline, and flags the items that have gone unanswered long enough to become a cost.
What makes these different from a chatbot
Three things. They are connected to the systems that hold your records, so an answer can be traced to the row it came from. They end in an action that follows your existing approval rules rather than bypassing them. And each one is aimed at a specific, measurable cost, so there is a number that tells you whether it worked.
Where they are today
In active development, at different stages. We are deliberately building each one for a narrow first audience rather than launching a general tool, because a product that does one industry's workflow properly is worth more than one that half-fits six. If any of these describes a cost you are currently paying a person to absorb, that is the conversation we want to have, and it is also how we decide what to build next.