Bank of England uses an LLM to read firms’ inflation stories

The Bank of England has used GPT-5.4 to analyse company-visit reports and trace what businesses say is driving price changes. Its research covers 66,820 reports and finds that the resulting indicators track familiar inflation measures while offering a view into the pressures firms describe.

Watch Desk analysis

What happened

The Bank’s researchers used a large language model to score reports from company visits between January 2005 and September 2026. They asked it to assess past and expected output-price changes, and to separate six possible influences: direct energy costs, indirect input costs, labour, demand, profit margins and competitors’ pricing. Reports without enough pricing detail were excluded.

The dataset covers 2,000 to 3,000 firms a year. The researchers analysed a random third of reports up to 2025 and the full sample for 2026. They say the model’s aggregate price scores closely track CPI inflation and Bank staff assessments, while the breakdown offers a picture of the factors firms link to their pricing decisions. Their full account of the research also notes that manufacturers and medium-sized firms are somewhat over-represented.

Why it matters

Inflation data show how prices move, but not always what businesses say is behind those moves. This approach turns detailed conversations into a consistent series that can be compared across sectors and over time. The researchers’ analysis suggests recent cost pressures linked to conflict in the Middle East have so far been narrower and less persistent than the broad pressures seen in 2021–23.

That makes the work potentially useful as another lens on inflation, not a replacement for established measures. The scores reflect what firms describe in reports, and the sample has limits; correlation with CPI does not by itself prove that the model has identified the causes of inflation.

Our read

This is a substantial use of an LLM: not to produce another fluent summary, but to sort a large archive of business conversations into distinct economic signals. The interesting test is whether those signals help policymakers spot changes earlier or understand them more clearly. A model can make the filing cabinet speak, but the labels still deserve scrutiny.

What to watch

• Whether future work tests if firms’ expectations provide an early signal of inflation.
• How the indicators perform across sectors and against other economic measures.
• Whether the researchers publish more detail on validation and the sample’s limits.

Discussion spark: Should policymakers treat LLM analysis of firms’ conversations as a useful early-warning signal, or is the risk of mistaking a model’s categorisation for economic evidence too high?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.