AI Analysis of 66,820 Business Reports Tracks the Drivers of UK Price Changes

Firms’ accounts of how they set prices can reveal pressures that inflation data alone don’t explain. An analysis using an artificial intelligence model to read 66,820 company visit reports finds that recent cost increases linked to conflict in the Middle East have so far been more concentrated than the broad pressures seen in 2021–23. That distinction matters: the reports suggest weaker demand and wage pressures may limit how much of today’s costs feed through to prices.

Marcus Buckmann, Galina Potjagailo and Shobhit Rawat

Reading firms’ accounts at scale

Inflation measures show how prices change across the economy, but offer less detail about the decisions behind those movements. Bank of England staff meet companies around the United Kingdom and record discussions about demand, costs, wages, margins, investment and prices. These reports combine company details, such as location, employment and turnover, with qualitative descriptions of what businesses are experiencing and how they respond.

The research applies GPT‑5.4 to those narrative sections, turning them into consistent scores for past and expected price changes and the factors firms connect to them. The data set covers January 2005 to September 2026 and contains 66,820 reports, averaging around 260 per month. The monthly count had fallen to 237 in 2026. To limit processing costs, the researchers use a random one-third sample through 2025 and all available reports for 2026.

There are limits to the coverage. Reports span 2,000–3,000 firms a year, and manufacturers and medium-sized businesses appear somewhat more often than in the wider UK business population. The reports have also grown longer: median text length rose from 585 words in 2005 to 1,254 words recently. That gives the later observations more detail, but means the material available for analysis varies over time.

Separating price pressures by cause

For each report, the model assesses output-price changes over the previous three to 12 months and expectations for the next three to 12 months, relative to normal conditions. The researchers ask it to identify six possible contributors: direct energy costs; indirect input costs transmitted through supply chains; labour costs; demand; profit margins, including whether firms absorb or pass on costs; and strategic complementarities, when businesses adjust prices in response to competitors facing similar pressures. Any contribution that cannot be assigned to those categories is recorded as unattributed.

The method depends on an explicit connection between a reported factor and a pricing decision. For example, a reference to wages counts as a labour-cost pressure only if the report says wages influence the firm’s prices. Reports without enough pricing information are left out. The researchers then average the report-level scores and contributions to create monthly measures for the economy and individual sectors.

The model’s realised-price score moves closely with UK CPI inflation, though it varies more from month to month. It is also strongly correlated with scores assigned manually by Bank staff. At some turning points, the model’s measure shifts earlier, including in 2016 and 2021, and retreats more quickly afterward, particularly in 2023. Its fit is less close during the period after the global financial crisis and in the recent inflation episode, when the unexplained component is larger.

Chart 1: LLM score of business contacts’ output price changes over the past three to 12 months, against manual Bank staff (Agents’) scores and CPI inflation

Notes: LLM output-price scores (black line), demeaned and rescaled to CPI inflation (dotted line, mean of 2.9%, max: 11.1% in October 2022). Aggregated across all reports used. Original scale for LLM score: [-100;100], Bank staff (Agents’) scores: [-5;5].

Different inflation episodes, different drivers

The decomposition gives the price score an economic story. Around the global financial crisis, firms’ accounts point to weak demand as the main source of downward pressure, followed by strategic pricing and factors the model could not attribute. That softness was less apparent in CPI inflation. Labour costs also restrained prices during the recovery that followed.

In 2014–15 and during the Covid-19 pandemic, lower global oil and commodity prices showed up as weaker indirect cost pressures, although the movement was smaller than in CPI inflation. After the Brexit referendum, reports instead described rising non-labour costs and a larger unattributed component. The researchers suggest that greater uncertainty may help explain the latter, while the measure itself cannot identify the cause.

Chart 2: LLM decomposition of business contacts’ output price changes over past three to 12 months

The 2021–23 surge had a broader sequence. In 2021, businesses reported direct and indirect cost pressures alongside strong demand; margins and competitor responses played smaller roles. During 2022–23, labour costs and price-setting channels gained importance, especially among service providers, even as demand weakened. By 2024–25, labour costs had become the dominant upward pressure, alongside announced increases in the National Living Wage and employers’ National Insurance contributions.

Splitting the data between goods producers and service providers helps show that these pressures do not affect all firms in the same way. The respective measures are rescaled against goods and services inflation, whose reported means are 2.3% and 3.6%. The researchers’ account suggests services firms continued to report pressure from margins and strategic pricing, indicating some capacity to pass costs on. Yet weaker wage pressure than in 2022–23 may leave more room for firms to absorb some increases.

Chart 3: LLM decomposition of output price changes, by goods and services producers, from 2021

Notes: Demeaned and rescaled LLM output-price scores (black line) and decompositions (bars), aggregated separately by goods and services producers. Demeaned and rescaled based on goods inflation (mean: 2.3%) and services inflation (mean: 3.6%), respectively.

Expectations are running ahead on direct costs

Recent reports show a temporary rise in direct cost pressures after conflict in the Middle East, followed by some increase in indirect costs. So far, both look more muted than during 2021–23. Weak demand and subdued labour-cost pressure also distinguish this period from the earlier surge. For service firms, persistent contributions from margins and strategic pricing suggest that some cost pass-through is continuing, although less wage pressure could make cost absorption more feasible.

Past and expected price narratives usually move together. That may reflect how difficult it is to separate recent developments from forward-looking comments in the reports, as well as firms’ tendency to extend current conditions into their expectations. Around major shocks, however, the measures diverge. In early 2021, expected pressures strengthened before realised pressures, particularly for direct and indirect costs and margins.

Since the Middle East conflict began, direct cost pressures have risen more sharply in expectations than in accounts of realised pricing and remain elevated. A temporary increase in expected indirect costs appears to be feeding into realised pressures, but to a much smaller degree than in 2022–23. Whether these expectation measures can provide an early signal for inflation is a question for further research.

Chart 4: Comparing contributions to past and expected output-price changes

Notes: Selected LLM decomposition contributions for past three to 12 months (solid) and next three to 12 months (dashed) narratives, 2020–26, aggregated across all reports, no rescaling, 12-month one-sided moving average.

What the comparisons can and can’t show

The researchers compare the model’s direct-energy contribution with UK CPI energy inflation and its demand contribution with a Bank of England staff estimate of the UK output gap. Both pairs move in similar ways. The output-gap estimate comes from a state-space model covering 1988 Q1 to 2026 Q3 and is interpolated from quarterly to monthly frequency for the chart. Since real-time output-gap estimates are uncertain, the firms’ demand measure may offer a useful additional view.

These comparisons support the view that the model extracts meaningful signals from company narratives. They do not establish that its categories identify the underlying economy-wide causes of inflation. The practical value is more specific: a consistent way to compare what businesses say about prices, costs and demand across sectors and over time. In the latest reports, the clearest distinction from 2021–23 is the limited evidence so far of broad propagation from direct costs into wages, demand and other pricing channels.

Chart 5: Selected LLM decomposition contributions compared against external series

Notes: LLM direct-energy (left) and demand (right) contributions from decomposition of realised output-price changes over past three to 12 months. The LLM contributions are rescaled to the corresponding dashed-line external series. The output gap is a Bank of England staff estimate from a state-space model estimated over 1988 Q1 to 2026 Q3 and is linearly interpolated from quarterly to monthly frequency for plotting.

Sources: Bank of England staff calculations, ONS and authors’ calculations.

Marcus Buckmann works in the Bank’s Advanced Analytics Division, Galina Potjagailo works in the Bank’s Monetary Policy Innovation Lab and Shobhit Rawat works in the Bank’s Agency Network.