The brief was simple in one paragraph and brutal in execution. A specialist commercial insurer had spent eighteen months and seven figures building an AI underwriting model that scored beautifully in offline backtests. The problem was that it had never made a real underwriting decision. Three months after engaging us, it was making twelve thousand decisions per week across three lines of business — and the underwriters had stopped complaining.
The client situation
The client was a specialist commercial lines insurer headquartered in the City of London, with annual gross written premium in the mid-hundreds of millions of pounds. Their underwriting AI initiative had launched in late 2024 with a clear hypothesis: machine learning could augment human underwriters in property and casualty pricing, reduce quote turnaround from days to hours, and improve loss ratios through better risk selection.
The hypothesis was correct. The execution had stalled. By February 2026, the project had a working model in a Jupyter notebook, a data lake containing most of the inputs, and a deployment pipeline that nobody quite trusted enough to flip to production. The underwriting team had grown sceptical. The board had grown impatient.
We had a model. We had data. We had a roadmap. What we didn’t have was anyone who could see the whole picture and own getting it across the line.
What we did differently
The mistake in the original programme had not been technical. The model was sound. The mistake was structural. The data science team owned the model. A separate platform team owned the deployment pipeline. A third team owned the underwriting workflow. No single team owned the outcome. Each had done their job; none had shipped value.
Week 1–2: Discovery workshop
We started with a structured discovery workshop running in parallel from Casablanca and London. Hassan Bahej and the senior delivery lead spent four days in the City, embedded with the underwriting and data science teams. By the end of week two, we had identified the binding constraints — three of them, all organisational, not technical.
Week 3–4: Team formation
We formed a dedicated 8-person team: 2 senior ML engineers, 2 data engineers, 1 MLOps specialist, 2 backend engineers, and 1 delivery lead reporting directly to the client’s Chief Underwriting Officer. The team was based in Casablanca, with one senior engineer spending two days a week in London. The one-hour time-zone overlap meant the underwriting team could ask questions and get answers the same day.
Month 2: First production decision
By the end of week six, the model was making its first real underwriting decisions on a constrained subset of property risks — deliberately scoped so the model could be wrong in cheap ways before it could be wrong in expensive ways. Underwriters reviewed every decision for the first 400 cases. By case 200, agreement rates had crossed 80%.
Month 3: Scale across three lines
By day 90, the model was deployed across property, casualty, and a specialty marine line — covering approximately 70% of the insurer’s commercial book by premium volume. Quote turnaround had fallen from an average of 38 hours to 14 hours. The underwriting team had moved from sceptical to demanding more.
Reduction in delivery time
Lines of business in production
Weekly decisions automated
Avg quote turnaround (from 38h)
What made the difference
Three structural choices separated this engagement from the one that had stalled. First, a single accountability line: the delivery lead reported directly to the Chief Underwriting Officer, not to a steering committee — decisions that previously took three weeks of alignment were made in a 30-minute call. Second, same-day collaboration: the one-hour time-zone gap meant the model was iterated in real conversation with the people who would use it. Third, bilingual technical communication: native French and Spanish speakers on the team meant the client’s French and Spanish commercial portfolios were understood in context, not in translation.
A note on team continuity
The same 8-person team that delivered the first 90 days is still in place ten months later. The model has now been extended to four additional lines of business, two new geographies, and is being adapted for the client’s broker partners. Continuity of team matters at least as much as quality of team.
Lessons for AI buyers
Three lessons generalise from this engagement. Stalled AI projects are almost never stalled for technical reasons — they stall because no single team owns the outcome, or because that team cannot have a real-time conversation with the business it serves. The right delivery model is the one that compresses the iteration cycle: time-zone alignment, language fluency, and a single accountability line do more for delivery speed than any architectural decision. And the path to production runs through a constrained scope where being wrong is cheap.
A final word
The most striking moment came in the month-three readout. The Chief Underwriting Officer, asked what had changed, said: “You gave us a team we could actually talk to.” That sentence captures why the traditional offshore model fails for AI work, and why nearshore — done properly — succeeds. The technology was never the problem. The conversation was.
Client name withheld for confidentiality. Outcome data validated by the client’s internal audit function. Atlantia AI’s underwriting AI delivery practice supports insurers across the UK, France, and the GCC.