We ask every client for written feedback once a project reaches production. Here are some of those responses, shared with permission.
"We had tried two other consultancies before Mettle AI Drive. The difference was immediate: they spent the first week understanding our claims data instead of pitching a generic platform. The fraud-detection model they built catches roughly £120,000 a month in suspicious claims that our old rules missed."
"Our warehouse team used to spend four hours a day manually checking delivery paperwork against purchase orders. The document extraction pipeline Mettle built reduced that to about twenty minutes of exception handling. The ROI was obvious within the first month."
"They told us upfront that one of our three proposed use cases would not benefit from machine learning and suggested a simpler approach instead. That honesty saved us about £30,000 and meant we could focus the budget on the two projects that genuinely needed AI."
"I was sceptical about chatbots after a poor experience with a previous vendor. The knowledge assistant Mettle deployed for our customer service team actually works: it answers 73% of tier-one queries without human intervention, and the answers are accurate because they pull from our own documentation."
"What I appreciated most was the handover. They did not just deliver a model and disappear. We got a full documentation pack, three training sessions for our data team and ninety days of support. Our engineers can now retrain the model themselves when new product lines launch."
"The demand forecasting model improved our stock accuracy by 18% in the first quarter. We still have some edge cases in seasonal product lines that need tuning, but the team has been responsive and is working through them with us during the support window."
Pennine processes around 14,000 motor insurance claims per month. Their existing rules engine flagged about 5% of claims for manual review, but missed a significant volume of coordinated fraud rings operating across multiple policies.
We trained a graph neural network on three years of claims data, linking claimants, repair shops, solicitors and witnesses. The model identifies clusters of related entities that share suspicious patterns, such as the same witness appearing in unrelated claims filed weeks apart.
Hartwell receives delivery notes, invoices and packing lists in dozens of formats from over 200 suppliers. Their warehouse staff were manually keying data from PDFs and scanned images into their ERP system, a process that consumed roughly 160 person-hours per week.
We built a pipeline combining optical character recognition with a fine-tuned layout-aware language model. The system extracts supplier name, PO number, line items, quantities and totals, then pushes structured JSON into Hartwell's SAP instance via API.
If you are a current or past client and would like your feedback featured here, drop us a line at [email protected]. We are also happy to co-author a detailed case study if your compliance team approves it.
Thinking about starting a project? Get in touch and we will set up a call to discuss your needs.