Example builds

Working systems we have built, written up without names. A person approves before anything is saved.

support email triage · sales-order intake · quality diagnostics · document / compliance · voice intake · knowledge-in-chat · lettings · meeting notes · property searching · appointment reminders · invoice extraction · call helper

Support inbox

Support email triage

Support desk, shared mailbox in Microsoft 365.

Challenge
Customer questions and junk landed in the same inbox. Staff wrote every reply by hand, with no label/category for what was urgent or how the sender sounded.
Approach
New mail is labelled Urgent or Routine. The agent determines the tone, for example upset or angry. Urgent mail also sends a separate Teams message so the support lead sees it straight away. When the support FAQ covers the question, a draft reply to the sender is saved for the support lead to send. Suspected spam, a question the FAQ does not answer, or a draft that conflicts with earlier mail from the same sender waits for the support lead. Mail is not deleted.
Outcome
The support lead sends from a saved draft. FAQ gaps, and conflicting email thread wait for a person to review.
Sales order intake

Purchase orders into the ERP

Order desk, shared mailbox, ERP writeback.

Challenge
Staff re-typed inbound purchase orders into the ERP. There was no reliable review gate before rows hit the ledger.
Approach
Intake from the order mailbox; extract the PO; match customer, SKU, and price to live catalogs; flag gaps and mismatches; side-by-side approver console; writeback only after Approve.
Outcome
The desk resolves exceptions instead of re-keying. Unapproved rows do not reach the ledger.
Quality diagnostics

Quality diagnostics and reports

Quality / compliance at a pharmaceutical manufacturer or GDP site, diagnostic dashboard, not a public website.

Challenge
Staff assembled the quality picture from workbooks and wrote the site narrative by hand. Scores, exceptions, and the written report drifted from each other.
Approach
Structured intake into a scored set of quality signals; AI interprets the already-computed results into a draft executive report; a reviewer checks the dashboard and the narrative before it is treated as the quality pack.
Outcome
The team reviews a scored picture and a draft report instead of assembling the story from scratch. Unreviewed narrative does not become the pack.
Document / compliance

Supplier licence verification

Quality / compliance at an EU manufacturer, their supplier register is the list to check. Licence facts come from the official EEA register, not staff searching the internet or company website.

Challenge
Staff checked supplier licences by hunting official registers and re-typing addresses and scope into a spreadsheet. Drift after the first check was easy to miss.
Approach
Pull MIA/WDA records from the official EEA register; present address and scope for a human confirm; alert when the register later drifts from the confirmed baseline. Nothing becomes the compliance record without that confirm.
Outcome
The team confirms a structured record instead of re-typing the register. Later changes surface as alerts, not silent drift.
Voice intake

Customer voice intake

A customer describing a problem or issue, AI phone agent, not a web form.

Challenge
Callers waited for a human agent or left a voicemail. Staff reconstructed the issue from memory or a recording. There was no reviewed summary before someone picked up the work.
Approach
The AI phone agent verifies date of birth, email, and phone number, then the customer describes the problem. The agent grills the customer about the problem from a predefined FAQ list. A written summary is drafted from the call and sent to support staff. A person reviews it before it is treated as the record.
Outcome
Support starts from a reviewed summary instead of replaying the call. Unreviewed intake does not become the ticket.
Knowledge-in-chat

Workplace knowledge in chat

Employees asking HR, policy, holidays, and compliance in one place.

Challenge
Policy, leave, customer records, and compliance lived in different places. Staff hunted the library, guessed, or had to know which system a question belonged to.
Approach
Staff ask in one place. Policy and procedure answers come only from the organisation’s library, with a link to the source document. Holiday balances and booking go to Workday. Customer questions go to the CRM. Compliance questions go to the compliance source. If it is not in a document the employee can open, the agent refuses rather than inventing an answer.
Outcome
Staff ask once. Leave is booked in Workday, and other questions reach the document or system that holds the answer. Off-library policy is not invented.
Lettings

Property enquiry desk

A property desk, applicant chat and an ops board, not a listings search.

Challenge
Viewing requests and repair messages landed in the same inbox. Staff qualified buyers by hand, and a reply could go out before a viewing was actually booked.
Approach
The applicant chats about the property. The agent asks for budget, how soon they want to move, and whether they have approval in principle or proof of funds. A qualified applicant is offered viewing times. A repair message becomes a maintenance ticket instead of a viewing. The reply stays a draft on the ops board until someone on the desk sends it.
Outcome
The desk sends a draft and books a viewing for a qualified applicant. A repair is a ticket. An unqualified enquiry does not become a booking.
Meeting notes

Meeting notes in Teams

A Teams organizer after a call, channel and notes file, not a recording dump.

Challenge
After the call, the record lived in the recording or in someone’s notes. There was no written summary and action list in the channel before people treated it as the record.
Approach
The meeting transcript is turned into a written summary and action items in the channel, with a link to the notes. A person checks them before they are treated as the record. The recording itself is not sent for analysis.
Outcome
The team works from a checked summary in Teams. Unreviewed notes are not the record.
Appointment reminders

Clinic appointment reminders

A clinic or practice, calendar and patient phone.

Challenge
Upcoming appointments needed a reminder, and inbound replies mixed confirmations, cancellations, and people who still needed the desk. Booking changes were easy to treat as done when they were not.
Approach
Upcoming appointments trigger a reminder SMS. Inbound replies are classified so the desk sees who confirmed, cancelled, or still needs a person. Nothing is treated as a booking change without a person checking it.
Outcome
The desk sees who still needs a person. Unchecked replies are not treated as booking changes.
Invoice extraction

Invoice extraction

Finance or ops receiving invoice PDFs.

Challenge
Staff re-typed vendor, dates, amounts, and lines from invoice PDFs. A file that was not an invoice could still be treated as the record.
Approach
The invoice file is read once into a structured record and a list row. A person checks the fields before they are treated as posted. If the file is not an invoice or cannot be read, the row is marked as a failure.
Outcome
Finance reviews a structured record instead of re-typing the PDF. A failed read does not become a posted invoice.
Call helper

Call Helper

A person on a live call, with an overlay on their laptop.

Challenge
On a live call, the next question, a summary, or the actions lived in someone’s head. Speech leaving the laptop, or a bot joining the call, was not acceptable.
Approach
An overlay sits on the laptop during the call. Speech stays on the device. A cue appears only when the person clicks: questions to ask, answer the last question, summarise, or actions. The overlay does not join the call.
Outcome
The person asks for a cue when they need it. Speech stays on the laptop, and the call is not joined by a bot.
Book a Discovery Call