Custom AI assistants that read, classify, answer, or summarise — so your team focuses on the decisions, not the triage.
The bottleneck is rarely the decision itself. It is the reading. Your team spends hours every day reading emails, figuring out what each one is, and deciding where it goes. The decision takes seconds, but the reading and routing take hours.
An AI agent handles the reading. It classifies the email, extracts the key data, and either drafts a response or routes it to the right person. Your team reviews the edge cases and makes the final call, so the agent handles the volume while the humans handle the judgement.
This never gets used for creative work or for decisions with legal or financial liability. There is always a human in the loop on anything that matters.
Pick the one that ate the most time last month. That is the candidate.
Partners and executives receive hundreds of emails per day. Someone reads each one to decide what's urgent, what to delegate, what to reply to, and what to archive. Important emails occasionally sit unread for days.
New customers email their requests as prose — sometimes with PDF specs attached, sometimes describing the problem in plain text. Someone reads each one, figures out what's being asked, and routes it to the right salesperson, engineer, or project lead.
20-30% of incoming questions are the same handful of things: opening hours, pricing, whether a service is available, document requirements. The front office answers each one by hand. When a key person is off, answers get inconsistent.
A support ticket started by one person gets picked up by another — at shift change, during holiday cover, or after escalation. The new owner re-reads the thread, calls the client back, and often asks the same questions again.
When a role is open, CVs arrive at a generic HR inbox or the hiring manager's inbox. Someone opens each one, reads the summary, decides whether it fits, and files the candidate into a spreadsheet. Most CVs are never answered.
Clients ask questions through WhatsApp, email, the website form, phone calls, and sometimes in person. Each channel is watched by a different person (or not watched at all during busy periods). The same question gets answered several times a week, inconsistently.
You tell me where the volume is: the inbox, the document type, the question category. I map what the agent has to read, what it has to decide, and where the human reviews.
I build the agent and run it on a sample of real cases. We compare its output against what the team would have done. I adjust the rules and the wording until the disagreements drop to noise.
It goes live with the human-review checkpoint switched on. I monitor the agent for 30 days and tighten it where it slips. After that it runs on its own, or you can add the annual support contract.
Discovery meeting
Agent design (role, boundaries, fallback rules, human-in-the-loop checkpoints)
Model configuration
Initial tuning
Testing and launch
30-day support after go-live
AI API costs (OpenAI and equivalents)
Creative content generation
Decisions with legal or financial liability
Scope changes after delivery
Delivery notes go in — scanned paper or PDF. The AI extracts every field, edge cases get flagged for human review, and the result is fully digitised.
Pre-production. Real numbers.
RFQs arrived as free-text emails with attached specs. The ops team spent ~90 min every morning classifying them by discipline.
~1.5 h/day freed, faster first response to clients
Tenant questions on WhatsApp — rent, inspections, small repairs — were handled one by one by the front office, interrupting the rest of the day.
~3 h/day freed across the front office, faster tenant responses
If you recognise the patterns above, thirty minutes is enough to know whether I can build it for you.
Let's talk