Essays from Moweb partners on the enduring questions in enterprise AI: the topics that shape the decisions our clients are making today.
Most retrieval-augmented generation systems are built to demo well and fail under model-risk review. The architecture that survives an audit is decided by four choices made before the first prompt: citation lineage, the evaluation harness, access control at retrieval time, and provenance.
Read the essay
Most enterprise AI pilots never reach production, and the reasons are organisational, not technical. The stall happens at ownership, evaluation, change management, and run-cost, in that order.
Read the essay
The build-versus-buy question for AI is not a technology decision, it is a decision about where your durable advantage actually sits. Buy the boring, build the differentiating, and account honestly for the total cost of ownership that vendors and internal teams both understate.
Read the essay
Selling AI to UK public bodies runs through frameworks, not pitches: G-Cloud 14, the Digital Marketplace, and a growing set of transparency obligations. Suppliers who treat the Algorithmic Transparency Recording Standard as a deliverable rather than an afterthought win the work others cannot.
Read the essay
A pilot's headline number and a P&L line are different objects. This is a finance-grade method to baseline, attribute, and net out run-cost so that an AI claim survives an audit committee.
Read the essay
The EU AI Act is not a paperwork exercise. The classification rules force operating-model choices that take months, not weeks, to land. Here is the 14-week working plan we use with mid-sized corporates.
Read the essay
Multi-step agents succeed in the back office when the work is bounded, reversible, and instrumented, and fail when it is open-ended, irreversible, and opaque. The distinction is not about model capability. It is about workflow shape and where you place the human.
Read the essay
US and EU AI obligations look like two programmes but share most of their underlying controls. A single internal control catalogue, mapped to both the NIST AI Risk Management Framework and the EU AI Act, lets a transatlantic enterprise run one governance function rather than two.
Read the essay
Keeping data and inference in-region is now an architecture problem, not a contract clause. The real decision is what you trade away when you choose an EU-hosted model over a frontier US one.
Read the essay
Ambient documentation and prior-authorisation AI are the two US healthcare use cases with real measured savings. The governance bar that separates the working deployments from the recalled ones is specific and knowable.
Read the essay
The unit cost of inference will fall, then plateau, then rise again as the workloads enterprises actually deploy get harder. We explain why, and what it means for procurement and architecture decisions in 2026 and 2027.
Read the essay
Most enterprise AI does not stall on models. It stalls on data access through systems built before anyone planned to read them programmatically. The modernisation work is the project.
Read the essay
Model risk leaders are not blocking generative AI; they are blocking generative AI with no documentation. A short note on the documentation that actually passes review at SR 11-7 and PRA SS1/23 institutions.
Read the essay
Pipelines win on Day 1. Products win on Day 365. The shift from pipeline thinking to product thinking is the single highest-return architectural decision the average enterprise data team can make.
Read the essay
CPS 230 came into force in July 2025. It is the cleanest operational-risk frame we have seen for AI workloads in any market, and we recommend it as a reference even to non-APRA-regulated firms.
Read the essay
Pan-African enterprise AI has its own physics: hybrid power, telco-grade infrastructure expectations, mobile-first user behaviour and cross-border regulatory texture. A field report on what works.
Read the essay