The Limits of Aggregate Data
Headline inflation numbers are sterile. They tell us the temperature of the economy, but they reveal nothing about the weather systems creating that heat. Aggregate consumer price indices show us that prices are moving, yet they remain entirely silent on the actual mechanisms driving those corporate pricing decisions. To understand the true anatomy of inflation, economists must look beneath the surface. For years, Bank of England staff have conducted regular visits to companies across the United Kingdom. They sit down with business leaders. They document these conversations in detailed company visit reports.
 |
| AI Quantifies Hidden Inflation Drivers in Bank of England Reports |
These documents are goldmines of qualitative insight. A typical report contains rich, localized data on employment, turnover, and margins. More importantly, it captures the nuanced, qualitative notes summarizing discussions on demand, costs, wages, and investment. This business intelligence provides a direct, unfiltered view of how firms perceive economic conditions and explain their own strategic choices. Yet, for a long time, this vast archive of text-based reports remained largely underexplored. Manually scoring thousands of pages is prohibitively slow. The narratives were trapped in plain text, waiting for a tool capable of reading them at scale.
Unlocking the Narrative Archive
That tool has arrived. Ongoing research now utilizes an advanced large language model to systematically analyze these text sections. The dataset is massive. It encompasses 66,820 reports spanning from January 2005 to September 2026. That averages around 260 reports per month, though volumes have naturally consolidated to about 237 per month in recent years. The texture of the data itself has evolved. The median length of these reports has more than doubled, growing from 585 words in 2005 to 1,254 words recently.
This expansion is not mere verbosity. It allows for a far more nuanced quantification of pricing narratives in the later part of the sample. The reports cover between 2,000 and 3,000 firms annually, with manufacturers and medium-sized enterprises slightly over-represented relative to the broader UK business population. By processing a random one-third sample up to 2025 and the full sample for 2026, researchers can manage computational costs while maintaining rigorous statistical validity. The model does not just scan for keywords. It reads context. It understands the difference between a passing mention of wages and a deliberate statement linking wage growth to a specific pricing decision.
The Anatomy of a Pricing Decision
The methodology represents a significant leap forward in economic analysis. For each report, the language model quantifies both reported and expected output-price changes over the past and next three to twelve months, relative to normal conditions. This mirrors the Bank’s existing practice of manually scoring firm assessments, but applies it with flawless, systematic consistency across the entire dataset.
Crucially, the model goes beyond simply tracking whether prices are rising or falling. It decomposes the narrative. It maps the factors cited by corporate contacts to a specific set of economic determinants. The preferred specification isolates six distinct drivers. Direct energy-related costs form the first layer. Indirect input costs follow, capturing the ripple effects through supply chains, such as energy used to produce raw materials. Labor costs and demand fluctuations make up the core operational drivers.
The model also isolates profit-margin adjustments, reflecting the extent to which firms choose to absorb higher costs rather than pass them on to consumers. Finally, it measures strategic complementarities. This is a vital, often overlooked metric. It captures how firms adjust their own prices in direct response to competitors facing similar market pressures. The model operates under a strict constraint: it assigns contributions only when a factor is explicitly linked to the firm’s pricing decisions. If wage developments are mentioned but not described as influencing price-setting, the labor cost component receives a zero contribution. Reports lacking sufficient pricing information are cleanly excluded.
Tracing the Economic Timeline
When aggregated, these report-level scores produce a remarkably intuitive historical record of inflation narratives. The model’s output-price score tracks actual CPI inflation closely, despite exhibiting greater month-to-month volatility. It also correlates highly with the manual scores assigned by Bank staff, yet it occasionally turns sooner. The model detected shifts in 2016 and 2021 before human analysts finalized their assessments, and it unwound those signals more quickly in 2023.
The historical decomposition is strikingly accurate. During the Global Financial Crisis, the model correctly identifies price pressures dominated by weak demand, followed by downward pressure from strategic pricing. In the 2014–2015 period and during the Covid-19 pandemic, the narrative accurately reflected weaker indirect cost pressures driven by lower global oil and commodity prices. Following the Brexit referendum, the text data pointed to rising non-labor costs and a larger unattributed component, perfectly capturing the heightened economic uncertainty of that era.
The 2021–2023 inflation surge tells a different story. Firms reported broad-based direct and indirect cost pressures, supported by strong demand. By 2024 and 2025, the narrative shifted again. Labor costs emerged as the dominant upward driver. This coincided precisely with announced increases in the National Living Wage and employers’ National Insurance contributions, proving the model’s ability to link policy changes to corporate behavior.
The Muted Echo of Recent Shocks
Current narratives offer a reassuring, albeit complex, picture. The effects of the Middle East conflict on realized pricing have remained relatively subdued. A temporary rise in direct cost pressures was followed by some indirect cost pressures in recent months. However, these effects are far more muted than the broad shocks seen in 2021 and 2022.
Weak labor cost pressures and subdued demand are acting as a buffer. For services firms, the decomposition suggests sustained price pressures from strategic pricing and profit margins. This points to a continued, albeit limited, ability to pass higher costs into consumer prices. Yet, compared to the peak inflation years, weaker wage pressures provide firms with greater scope to absorb part of these shocks internally. The economy is showing resilience, and the text data proves it.
The Forward-Looking Lens
Around economic turning points, what businesses expect is often more informative than what they report about the recent past. The research separately quantifies expected output-price changes and their determinants over the next three to twelve months. Past and expected contributions tend to move closely together. This reflects both the difficulty of fully separating past and future developments in natural language and the human tendency to extrapolate current conditions into the future.
The largest divergences emerge around major economic shocks. In early 2021, expected price pressures strengthened ahead of realized pressures, particularly through direct and indirect costs and profit margins. Since the start of the Middle East conflict, direct cost pressures have risen more sharply in expectations than in realized narratives. A temporary rise in expected indirect costs appears to be feeding into realized pressures, albeit at a fraction of the intensity seen previously. This forward-looking decomposition contains vital leading information for inflation forecasting.
Validating the Machine
Skepticism of algorithmic economic analysis is healthy, but the validation here is robust. The language model-derived pricing determinants closely resemble established macroeconomic indicators. The direct cost component tracks UK CPI energy inflation with high fidelity. The demand component aligns seamlessly with a Bank of England staff estimate of the UK output gap, utilizing a complex state-space model.
Given the inherent uncertainty surrounding real-time output-gap estimates, this language model-derived measure of firm-level demand pressures provides a highly useful complementary indicator. It proves that the decomposition recovers economically meaningful signals from corporate narratives, even though it was not explicitly designed to identify aggregate structural inflation drivers.
From Qualitative Intelligence to Quantitative Evidence
We are witnessing a fundamental transformation in economic monitoring. Large language models can now convert narrative business intelligence into rigorous quantitative indicators. This approach preserves meaningful distinctions between pricing determinants while complementing the qualitative assessments of human analysts. It provides a consistent, scalable framework for comparing narratives across firms, sectors, and time.
Beyond merely tracking whether firms report rising prices, this methodology uncovers the hidden drivers behind those changes. It allows economists to distinguish inflation episodes by their unique narrative anatomy. In the current economic conjuncture, it demonstrates that cost pressures remain concentrated in direct costs, with limited evidence of the broader propagation mechanisms that defined the 2021–2023 surge. This opens entirely new possibilities for turning qualitative business intelligence into timely, actionable quantitative evidence for modern economic policy.
 |
| AI Quantifies Hidden Inflation Drivers in Bank of England Reports |
Advanced language models are transforming thousands of qualitative Bank of England company visit reports into rigorous quantitative indicators, revealing the hidden economic determinants driving corporate pricing decisions and future inflation trends with unprecedented clarity.
#ArtificialIntelligence #Inflation #Macroeconomics #DataScience #BankOfEngland #LLM #EconomicPolicy #PricingStrategy #QuantitativeAnalysis #BusinessIntelligence