Kenya Tax Model (RIAPA-AI) — KIPPRA's AI-Chat-Driven Computable General Equilibrium Platform for Tax Policy Simulation

In 2024 the Kenya Institute for Public Policy Research and Analysis (KIPPRA), a state corporation under Kenya's National Treasury and Planning, co-created the Kenya Tax Model with the Kenya Revenue Authority (KRA) and the International Food Policy Research Institute (IFPRI/CGIAR). It is built on IFPRI's RIAPA computable general equilibrium (CGE) platform, with a chat-driven natural-language interface (RIAPA-AI), and now sits within KIPPRA's Economic Modelling Hub.

To let policymakers run tax and trade policy simulations without operating specialised CGE modelling software, so that proposed measures can be assessed for their poverty, nutrition and growth effects before enactment.

The model was applied to a real proposed policy: a 5% withholding tax on agricultural produce sold through cooperatives. The simulation results were presented on 5 April 2024 to a stakeholder workshop including KRA, county governments, the National Cereals and Produce Board and farmer cooperatives.

Without revenue recycling, the simulation projected a national poverty increase of 106,313 people, a rural poverty increase of 109,805 people, and 19,832 more undernourished people nationally; overall GDP would rise a marginal 0.02% while agricultural GDP would fall 0.05%. Agricultural GDP was projected to rise 1.07% instead if the new tax revenue were recycled into agricultural investment. KRA was reportedly open to refining the proposal in light of the evidence.

The quantitative results are developer-published by KIPPRA/IFPRI/CGIAR rather than externally audited, and the underlying analysis itself flagged data gaps around cooperative sales volumes requiring incidence assumptions. No source confirms whether the withholding tax was ultimately dropped, modified or enacted as a direct result of the modelling — only that the findings reached policymakers. The 'AI' element is a natural-language interface on top of a conventional CGE model rather than a machine-learning system in itself.