An AI is reading the reasons businesses give for raising prices

Imagine two restaurants with the same increase in electricity costs. One raises menu prices. The other holds prices because customers are already spending less, accepting a smaller margin. A search for mentions of energy costs would find both. To understand the different outcomes, a reader needs the rest of the sentence.

In ongoing research published on 15 October, Bank of England staff used GPT-5.4 to read pricing narratives. Their corpus contains 66,820 company-visit reports from January 2005 to September 2026. They analysed a random third through 2025 and all 2026 reports, excluding texts without enough pricing information.[1]

Behind those documents is an unusually human information network. The Bank’s staff operate through 12 regional agencies, holding confidential conversations with businesses and community organisations. They already score companies’ performance and economic conditions, and their intelligence informs policy committees. AI enters after people have asked questions, listened and written down what they heard.[2]

The distinction hiding inside a price rise

The researchers separate six explanations: direct energy costs, other input costs transmitted through supply chains, labour costs, demand, changing profit margins and responses to competitors’ prices. Factors count only when explicitly linked to pricing. Unexplained pressure has its own category. Scores cover reported and expected changes over three to 12 months.[1]

Return to the hypothetical restaurants. If the second owner describes expensive electricity but unchanged menu prices, the economically interesting action is absorbing the shock. If the first says a rival raised prices and it followed, competition enters the explanation. An identical bill can produce different decisions because each business faces different customers and alternatives.

Reading such accounts systematically could help distinguish a cost shock that businesses are absorbing from one spreading through their prices. That is a richer question than whether a particular word appears more often. It asks how a business connects what happened to it with what it decided to do.

One reported pattern: broad cost pressures in 2021 gave way to greater labour and price-setting contributions in 2022–23, especially in services. Labour dominated upward pressure by 2024–25. The pricing scores are rescaled to CPI for comparison; they are not independently measured inflation rates or a tested forecast.[1]

What counts as progress

A related Australian project shows how text can become a useful input to a forecast. In an August 2025 paper, Reserve Bank of Australia researchers described a tool for searching liaison notes, measuring topics and extracting figures such as reported wage growth. Its methods include dictionary-based text analysis: conversational evidence becomes searchable and measurable.[3]

A nowcast estimates current conditions before all the official figures arrive. In a retrospective test covering 39 quarters, from March 2015 to September 2024, a wage-growth model combining conventional variables and liaison indicators cut root mean squared error by about 20% against the baseline statistical model. That measure gives extra weight to larger mistakes. The gain included changes to the estimation method, rather than a standalone AI contribution. The RBA’s published nowcasts incorporating expert judgement still performed better over the full period, particularly because of their handling of early 2020.[3]

That is a useful standard of progress: does the additional information help when the answer is still unknown? It also leaves room for different strengths. A tool that makes a pattern visible can improve an economist’s questions before it improves a numerical forecast. A forecasting test measures a further contribution, with its own comparator and period.

Every measurement begins with someone’s account

Researchers at the Bank for International Settlements have approached the same problem through journalism. Their October 2025 working paper used US coverage from The Wall Street Journal, initially about 200,000 articles spanning January 2000 to April 2025. Models filtered relevant coverage and classified growth and inflation narratives into demand, supply and finer explanations. Their manual review of 300 classified articles estimated 92% accuracy, with most errors involving finer categories or failed API calls.[4]

The authors also identify a reason news sentiment can track economic statistics: reporting responds to published economic data. A close fit can partly reflect that information returning through another channel. Business interviews offer a different vantage point, closer to individual pricing decisions. Neither source should be mistaken for an unmediated view of the economy.[4]

The choice of source therefore shapes the question a model can answer. Newspaper analysis can map the explanations circulating in public. Company conversations can reveal how managers describe their own decisions. Comparing those accounts could be valuable precisely when they disagree: perhaps the public debate is focused on a pressure that some firms are managing to absorb.

The Bank’s regional network already exists to connect economic analysis with those experiences. Giving its written accounts a second, systematic reading could make more of that listening available across time and sectors. The promising output is an explanation an economist can investigate: this cost rose, this business responded, and here is the reason it gave.

Report an error or suggest a correction