We ask every client for honest feedback once a project reaches the three-month mark. Below is a selection of responses, lightly edited for length.
"We handed them 18 months of messy sales data and got back a churn prediction model that identified at-risk accounts two weeks earlier than our old spreadsheet method. The accuracy was 89 %, which is exactly what they promised during scoping. No surprises on the invoice either."
"Their document extraction pipeline processes roughly 1,200 invoices a day for us. Before this, three full-time staff spent most of their week on manual data entry. Two of those people now work on vendor negotiations instead, which has saved us more money than the AI project cost."
"The chatbot they built handles about 60 % of our inbound support queries without any human involvement. It occasionally misinterprets questions with heavy slang, but the team has been quick to retrain the model whenever we flag issues. Solid work."
"I was sceptical about AI hype, honestly. But the demand forecasting model they delivered cut our overstock by 22 % in the first quarter. The dashboard is simple enough that our warehouse manager checks it every morning without asking IT for help."
"We needed sentiment analysis on product reviews across four languages. They delivered a fine-tuned multilingual model in six weeks. F1 score was 0.87 on our test set, which beat the off-the-shelf solution we had been trialling by a clear margin."
Northline receives around 1,200 supplier invoices every day in PDF format. Each invoice has a different layout. Three accounts-payable clerks spent roughly five hours a day copying line items into the ERP system. Error rates hovered around 4 %, which triggered payment disputes and delayed shipments.
An OCR and NLP pipeline that reads each PDF, identifies key fields (supplier name, invoice number, line items, totals, VAT), validates them against the supplier master list, and pushes clean records into Northline's SAP instance via API. Invoices that fall below a confidence threshold are routed to a human review queue.
74 % less manual entry Error rate down to 0.6 % £42k annual saving
Two of the three clerks moved into vendor negotiation roles. The remaining clerk handles the review queue and edge cases, typically spending less than an hour a day on it.
Greenfield's support team was handling 900 live-chat conversations a day, mostly about billing queries, speed tests and router setup. Average response time had crept up to 14 minutes, and CSAT scores were falling.
A retrieval-augmented chatbot grounded in Greenfield's knowledge base (about 320 articles). The bot answers questions directly, runs diagnostic lookups against the network API, and hands off to a human agent when it detects frustration or cannot find a confident answer. We trained the intent classifier on 8,000 labelled conversation snippets from Greenfield's chat logs.
61 % queries resolved by bot Response time under 20 seconds CSAT up 11 points
The human agents now focus on complex technical issues and escalations. Greenfield hired two fewer seasonal support staff than planned for the winter peak, saving roughly £28k in temporary labour costs.