Enterprise AI programmes: more data, stronger models or better evidence selection?

THE 8TH GENERATION OF THE SENSIA ENGINE · FOR AI, DATA AND TRANSFORMATION LEADERS

When organisations deploy AI to analyse text, three investment choices often emerge: provide more data, use a more powerful model or improve how the information is understood and prepared. The right choice depends on what each contributes to the business outcome.

From the outset, Sensia has developed specialised models to understand consumer feedback and prepare relevant context for each question. With the 8th generation of our engine, we put this architecture to the test: what does substantially deeper reading add? What changes when a more powerful reasoning configuration is used?

The tests examine production market analyses. They address a broader question for enterprise AI programmes: where should organisations invest to turn large volumes of text into knowledge their teams can use?

What the tests show

• Substantially increasing the evidence supplied to AI, up to its maximum capacity, or using a high-end reasoning model produced no observed improvement in the main strategic findings or their order of priority in the comparisons studied.

• Resource use nevertheless increased sharply: 13.4 times the LLM usage cost at the maximum window, and more than 40 times the cost in some comparisons with a high-end model. Including preparation, observed processing times can range from a few minutes to more than an hour, depending on the configuration and preparation pipeline—a substantial constraint on on-demand conversational exploration.

• These results highlight the value of Sensia’s preparation: selecting the information that already supports the identification and prioritisation of useful findings, then enabling teams to investigate and reuse them.

Three levers, different investments

Increasing volume allows a system to examine more source material. A more powerful model may improve reasoning or synthesis. Better understanding and preparation help establish which information is relevant to the question.

Each lever draws on different resources: usage cost, processing time, specialist skills and development effort. The investment decision begins with identifying which lever improves the result you need.

Our benchmark examines two of these levers: reading depth and a more powerful reasoning configuration. Sensia’s specialised understanding and evidence preparation provide the common foundation across the configurations compared.

This allows us to observe what additional resources contribute when the information has already been analysed and prepared for the task.

How the Sensia engine organises the work

Sensia turns large volumes of heterogeneous text into findings teams can examine and use. The examples here concern consumer voices: product reviews, experiences, needs, barriers and drivers of choice.

The technical challenge extends beyond consumer insights: understanding fragmented text, identifying what answers a question, measuring qualified data and preparing useful context at scale.

This eighth generation advances the whole engine. It combines around twenty proprietary specialised models with developments in measurement, evidence selection, context preparation and the use of expanded context windows.

The architecture distributes the work across five complementary functions. The general-purpose model works with context already prepared by Sensia.

The architecture of the 8th generation of the Sensia engine. Evaluation is a requirement applied to the output. The methods used to qualify data, select evidence and prepare context are proprietary.

Specialised understanding is central to our investment. Across the three reference portfolios, between 29% and 47% of collected reviews contained usable information for the analyses studied. Other reviews could be relevant to different questions; they did not contribute to those examined here.

This distinction matters. A large collection of text does not automatically form a useful context. The engine must understand what each review contributes before preparing the analysis.

General-purpose models provide reasoning, synthesis and writing capabilities. A complete portfolio, however, can exceed their largest context window several times over. Research by Chroma also documented uneven use of long contexts and declining reliability as inputs grew across the models and tasks tested.

Anthropic recommends finding a compact set of high-signal tokens for the task. OpenAI describes the value of fast, economical models for well-defined tasks. These observations reinforce the importance of preparing context around the work expected of the model.

Two experiments on production analyses

We selected five production analyses from three of our largest and most heterogeneous portfolios, in skincare and pet food. Their size, fragmentation and diversity provide demanding conditions for testing the engine.

We ran 122 generations across two experiments:

1. Vary reading depth. Twenty sections were regenerated with the same model at six volumes of evidence.

2. Compare reasoning configurations. On two sections, we compared the production configuration with a high-end reasoning model using high reasoning effort and a larger context.

Sensia’s understanding and preparation of the data provided the common foundation.

Benchmark scope. The first experiment compares reading depths using the same model. The second compares configurations that differ on several parameters and therefore does not isolate the effect of the model alone. Each condition was generated once: this is a production benchmark, not a statistical repeatability study. Sensia’s preparation layer was held constant rather than evaluated independently

We examined four complementary dimensions:

• Main findings: the conclusions that could inform a team’s choices, assessed by comparing the analyses side by side.
• Selection reach: the share of total relevant review volume associated with products represented in the sample. This measures the weight of represented products, not the percentage of insights recovered.
• Supporting detail: the product references, competitive detail and consumer quotations supporting the findings.
• Resources used: input tokens, LLM usage cost calculated from published prices, and measured preparation and generation times.

Experiment 1: what does substantially deeper reading add?

Using the same general-purpose model, we regenerated twenty sections at six reading depths: the samples used previously, comprising around 160–280 reviews per section; around 500, 1,000 and 1,500 reviews; a standard-window configuration receiving around 250,000 tokens, or 1,800–2,400 reviews; and a maximum-window configuration receiving around 900,000 tokens, reaching up to around 9,000 reviews per section.

The window names describe configured capacity; the figures above indicate what was actually supplied. Even the maximum window remains smaller than a complete portfolio.

Main findings. Sensia’s focused selections already supported the main strategic findings and their order of priority. In the analyses compared, extending reading to the model’s maximum capacity produced no observed improvement in those findings or their prioritisation.

The core opportunity areas, leading choice driver and central trade-offs remained consistent in the analyses where they were assessed. This strategic interpretation is the basis of the test’s conclusion.

Selection reach. For four questions, we reconstructed the full pool of relevant reviews and measured the share associated with products represented at each reading depth. The question was whether the samples left out products carrying a substantial share of relevant feedback.

At around 1,000 reviews, represented products accounted for 94.4%–98.8% of relevant review volume across the four questions. At around 1,500 reviews, the range reached 98.1%–99.5%.

The gain varies by question. These results guide the choice of depth rather than establish a universal sample size. A particular product or topic may also justify a targeted examination.

Supporting evidence. Deeper reading can provide additional product references, examples or quotations. This enrichment should be distinguished from an improvement in the strategic findings and their order of priority.

Sensia lets users explore this evidence on demand, according to the brand, segment or question they want to investigate. Detailed observations are available in the benchmark report.

Resources used. Across the same twenty sections, LLM usage cost doubled at the standard window compared with the configuration of around 1,000 reviews. At the maximum window, it increased approximately thirteenfold, partly because the calls crossed the provider’s long-context pricing threshold.

Our production configuration uses around 1,000 reviews per section. This choice takes account of the findings obtained, the supporting detail required, usage cost and the need for fluid conversational exploration.

Depth remains adjustable. Some questions justify a larger context, and additional capacity is used according to what it contributes to the analysis.

Experiment 2: what does a more powerful reasoning configuration add?

We compared the production configuration with a high-end reasoning model on two sections: opportunity analysis and choice drivers. Both configurations used evidence prepared by Sensia.

The reference used high reasoning effort and received around 230,000 input tokens per section, approximately twice the production input. This compares two complete configurations; it does not isolate the effect of the model alone.

The high-end reasoning model produced no observed improvement in the main strategic findings or their order of priority in the sections compared. Its LLM usage cost was approximately 35–41 times higher, depending on the section.

Nuances in wording or additional examples do not, by themselves, constitute an improvement in the priorities informing a decision.

Across the two sections, the production configuration used 51.8% of the reference’s input tokens. Depending on the section, its generation time was 31%–45% of the reference’s. Input volume, generation time and usage cost are distinct measurements.

The engine separates quantitative measurement from narrative synthesis. Counts and weights are calculated from review data qualified by the specialised models. The general-purpose model uses the prepared context to develop the analysis.

In the choice-driver comparison, the same factor ranked first in all seven versions, while the percentages written in the narrative varied. Those estimates are distinct from calculated indicators: a figure intended for a dashboard or quantitative use should come from the underlying calculation on qualified data.

Evolving the model while retaining domain understanding

A complementary pricing comparison informs another decision: when should a programme adopt a new general-purpose model?

Newer models do not necessarily cost less per token. GPT-5.5 doubled GPT-5.4’s list prices, from $2.50 to $5 per million input tokens and from $15 to $30 per million output tokens (OpenAI, The Decoder).

Task cost also depends on token consumption and model behaviour. On its own evaluations, Artificial Analysis reports an average task cost around 64% lower for GPT-6.1 Sol than GPT-5.6 Sol, alongside a higher Intelligence Index score.

Source: Artificial Analysis, accessed 24 October 2026 (GPT-5.6 Sol, model leaderboard). Average task cost is weighted across its evaluations and includes token consumption, reasoning and caching. It is not a cost per successful task or a Sensia workload cost.

These results describe Artificial Analysis evaluations, not Sensia’s workloads. They illustrate why models should be assessed on the outcome and resources required for a task, alongside their price per token.

For an enterprise programme, adoption depends on whether the new model improves quality, usage cost or response time on its own tasks—and on verifying that improvement before deployment.

GPT-5.4 was the model used in Sensia production before Luna. Because understanding, measurement and context preparation are provided by our own components, the change did not require rebuilding them.

At standard prices, excluding caching, Luna represents a 12.5-fold reduction in both input and output token prices compared with GPT-5.4. Applying GPT-5.4’s prices to the initial benchmark workload gives an estimated LLM usage cost of $2.08 across twenty sections, compared with $0.166 using Luna.

With Luna and larger samples of around 1,000 reviews per section, the measured cost was $0.527—around a quarter of that initial-workload pricing reference.

This is a pricing simulation, not a measurement of historical production savings. The reports were not regenerated with GPT-5.4. The findings comparisons in this article come from the reading-depth and reasoning experiments described above.

Preparing context at the pace of conversation

In conversational use, the time that matters to the user includes preparing the information needed for the question and generating the analysis. Model generation time alone does not describe the whole experience.

The tests show substantial differences: including preparation, some processes take a few minutes, while others exceed an hour. These timings depend on the volume involved and the preparation pipeline. That difference changes the ability to ask another question, adjust its scope and keep exploring without losing the thread.

Evidence selection is therefore also a usability consideration: use the information needed to answer the question, then investigate further where it contributes meaningful value.

The benchmark distinguishes preparation from generation. Maximum-window tests used an adapted preparation tool, so their timings do not form a homogeneous production series. They indicate the scale of the processes measured, not end-to-end conversational latency.

Where to invest, and where to partner

The objective is to identify useful findings and their order of priority, then make them usable across the business. In the comparisons studied, the additional resources committed to maximum-depth reading and the high-end model did not improve these central outcomes, despite higher costs and longer processing times.

Sensia’s value extends beyond efficient evidence selection. The engine enables teams to consolidate findings, retrieve and reuse them, and generate analyses on demand from different perspectives, scopes and segments. Users can investigate the sources and examples relevant to their question while building on a structured understanding of their market.

These observations reinforce our investment in specialised understanding, measurement and evidence preparation. That foundation is shared across the configurations compared; its contribution was not measured independently.

The levers can also be combined. A lower-cost general-purpose model can create room to examine more evidence within a comparable budget. In our pricing simulation, Luna at around 1,000 reviews per section costs approximately 3.9 times less than the GPT-5.4 pricing reference applied to the smaller initial workload.

These are different configurations compared at standard prices. The relevant decision considers the quality and resources of the overall system, including both the context supplied and the model used.

For an enterprise programme, the decision to develop internally or work with a partner concerns the capabilities that must be built and maintained over time.

General-purpose models provide capabilities their vendors continuously improve. Domain understanding requires qualified data, specialised models, business evaluation and sustained maintenance. This is the foundation Sensia brings to AI programmes that need to understand consumer voices.

A specialist partner can provide a foundation on which internal teams integrate AI into business processes, while retaining the flexibility to adopt advances in general-purpose models.

MIT’s 2025 study of enterprise generative AI reported higher success rates for tools acquired from specialised vendors or developed through partnerships than for the internal builds studied. It highlighted integration with real workflows as a central factor (Computing). The decision should therefore consider available expertise, development time and the deployment requirements of the programme.

The organisation remains responsible for defining the business outcomes and acceptance criteria. Testing, interpretation and ongoing quality monitoring can then be developed with the partner

Five questions to guide your AI programme

About the benchmark: five production analyses, comprising twenty sections, across three reference portfolios in skincare and pet food. Analyses were regenerated at six reading depths. Two sections were also compared with a high-end reasoning configuration using a larger context: 122 generations in total, October 2026. Selection reach was measured for four questions through 24 read-only reconstructions, separate from those generations. Main findings were assessed by comparative reading. The benchmark evaluates the current engine in several configurations; it is not a controlled comparison with the previous generation. LLM usage costs are calculated using the provider’s published prices and do not represent the full cost of an enterprise programme. Full measurements and results are available in the benchmark report.