What Is Deep Analysis?
Deep Analysis ≠ Deep Search.
You’ll have come throughout “Deep Search” options in instruments like ChatGPT or Claude — designed to reinforce retrieval and concise solutions. Whereas Deep Search focuses on retrieval efficiency and short-form solutions, Deep Analysis is about understanding, reasoning, and synthesis — combining adaptive planning, retrieval, evaluation, and context engineering to supply long-form, well-cited analysis outputs.
Consider it as shifting from “discover me one thing” → “clarify and cause by this subject for me.”
In contrast to easy search duties, Deep Analysis requires persistence, iteration, and strategic depth — just like how a human analyst or marketing consultant would work.
Conventional AI search methods are constructed to reply fast factual questions:
“What’s Salesforce’s income in 2024?”
However Deep Analysis asks:
“How is Salesforce’s income progress correlated with generative AI adoption within the enterprise sector, and what can we be taught from rivals’ go-to-market shifts?”
This shift adjustments all the pieces — the time expectation, width of exploration, and depth of reasoning are far higher.
Deep Analysis blends planning, reasoning, and writing — not simply retrieval.
Key constructing blocks embrace:
- Adaptive Planning: dynamically decomposing advanced analysis objectives
- Retrieval: gathering from various, multimodal sources
- Evaluation & Reasoning: connecting dots throughout proof
- Context Engineering: curating context for LLMs to remain constant
(see this nice piece) - Lengthy-Kind Synthesis: writing coherent, well-cited stories
What Is Enterprise Deep Analysis — and Why Does It Matter?
In an enterprise setting, analysis doesn’t dwell in isolation. Data is scattered throughout:
- Inner methods comparable to Salesforce, Slack, Google Docs, Calendars, inside information bases, and so on.
- Exterior sources comparable to LinkedIn, GitHub, public information, stories, internet information, and so on.
Enterprise Deep Analysis bridges each worlds — combining inside information and exterior insights to serve strategic enterprise objectives.
Instance enterprise functions embrace:
- Gross sales: Account analysis and aggressive evaluation
- Service: Rising difficulty triage throughout assist information and boards
- Advertising and marketing: Market development synthesis from CRM and exterior media
- Management: Strategic determination briefs and forecasting
- Engineering: Benchmarking rivals’ tech stacks and repos
The end result isn’t simply solutions — it’s insights that drive motion.
Enterprise Deep Analysis stories are highly effective instruments that serve a number of functions. They act like clever consultants—guiding each worker from particular person contributors to senior executives in making higher selections.
These stories distill advanced information into accessible insights that may gasoline real-time providers, speed up reply discovery, and uncover hidden patterns or root causes behind enterprise challenges.
Distinctive Challenges in Enterprise Deep Analysis
- Planner Intelligence
- Does the system know the place to seek for what?
- Can it stability inside vs. exterior information sources?
- How does it handle time-sensitive, contradictory, or incomplete data?
- How does it coordinate throughout instruments like Salesforce, Slack, Google Workspace, and LinkedIn?
- Device and Knowledge Entry
- Are the best APIs and connectors in place?
- Can the system parse and retrieve information precisely from structured and unstructured sources?
- Privateness and Entry Management
- Inner information isn’t open to everybody. Who’s allowed to see what?
- How can the system respect permission hierarchies and information residency guidelines?
- Quotation and Analysis
- How will we guarantee each perception is traceable to its supply?
- How will we consider the high quality of analysis when human experience is uneven or fragmented?
- How will we detect duplicated or conflicting data throughout methods?
An Instance of Enterprise Deep Analysis
In one among our inside use instances for gross sales, we designed a modular, multi-graph structure that mirrors how human researchers function, dividing and conquering by specialised sub-systems that collaborate intelligently.
1. Planner Sub-Graph
The Planner is the mind of the system — decomposing high-level analysis objectives into actionable subtasks.
- Activity Enter: Accepts pure language analysis requests (optionally paired with a predefined or LLM-generated template).
- Background Investigation:
The Background Investigator Agent scans a number of information layers:- Public internet (search, crawlers)
- Computational instruments (code execution, evaluation modules)
- Inner methods (Salesforce MCPs, CRM information connectors)
- Activity Decomposition:
Utilizing findings from the background stage, the planner breaks down the issue into well-defined subtasks mapped to acceptable instruments. - Subtask Execution:
Every subtask is dispatched to the Orchestrator Sub-Graph, both in parallel (for unbiased duties) or sequentially.
2. Orchestrator Sub-Graph
The Orchestrator is the challenge supervisor of Deep Analysis — overseeing every subtask and synthesizing partial findings right into a unified report.
- Step 1: Creates a tough define of the ultimate report and assigns a set of N analysis steps to specialised Activity Researcher/Executor Sub-graphs (e.g., Public Net Researcher, Inner Salesforce Researcher, Coder, and so on.).
- Step 2+: Iteratively refines the define by analyzing the outputs from executors, figuring out lacking items, and planning the following analysis wave.
- Reporting: As soon as protection is ample, the Reporter Agent consolidates findings, generates a long-form report, and attaches citations.
- Human-in-the-loop: Optionally available human suggestions might be built-in at any stage to refine route or validate conclusions.
3. Activity Researcher / Executor Sub-Graph
Every executor agent focuses on a single activity — whether or not that’s querying Salesforce information, summarizing a GitHub repo, or operating a code experiment.
They act because the arms of the system, executing with precision, feeding outcomes again to the orchestrator for synthesis.
4. Instruments
Our framework employs specialised instruments designed to navigate distinct information landscapes. These instruments act because the arms of the system, executing focused queries to construct a complete view. The first instruments embrace:
- Net Search: This device queries the general public web to collect exterior data, comparable to public information, market information, and competitor stories. This enables the system to include a broad, exterior perspective into its evaluation.
- Enterprise Information Search: To faucet into proprietary firm intelligence, this device searches inside information bases like Trailhead, Highspot, and Confluence. It retrieves essential data from onboarding paperwork, product particulars, and aggressive battle playing cards, immediately addressing the necessity to leverage inside information.
- CRM Search: This device connects on to inside methods like Salesforce to acquire account particulars and different buyer relationship information comparable to alternatives and conversational information.
- Dialog Search: Inner Slack conversations from completely different channels and messages.
The Planner Sub-Graph intelligently coordinates these instruments, figuring out whether or not to deploy them in parallel or sequentially primarily based on the analysis activity.
Enterprise Deep Analysis Analysis
Having established the first capabilities of those methods, it turns into essential to assess Deep Analysis fashions utilizing analysis strategies distinct from these utilized to look or Q&A fashions.
In Deep Analysis, accuracy is important, however not ample. A system might be 100% right on particular person information factors but fail to supply a helpful strategic evaluation. Subsequently, our analysis should shift.
The last word purpose is to measure how successfully an agent can perceive a posh purpose, cause throughout disparate sources, and synthesize data right into a coherent and actionable report.
Why Does Benchmarking Matter?
With out structured analysis, Deep Analysis stays anecdotal to customers, the place “it feels smarter” or “this report appears good.”
This isn’t sufficient for the enterprise. Enterprise enterprise contexts demand repeatable, explainable, and quantitative benchmarks. This rigor is particularly important when a number of brokers (OpenAI, Gemini, Slackbot, and so on) are producing business-critical analyses.
Benchmarking Deep Analysis methods is not only about leaderboard scores. It’s about constructing belief. For any enterprise consumer to behave on an AI-generated perception, they want clear solutions to a few basic questions:
- Traceability: The place an perception got here from
- Transparency: The way it was derived
- Consistency: Whether or not it aligns with company fact
Benchmark Practices in Pipeline
Our current analysis: SFR-DeepResearch [1], DeepTrace [2], HERB [3], and LiveResearchBench [4], present complementary views on methods to measure these capabilities:
- SFR-DeepResearch [1] (https://arxiv.org/abs/2509.06283) focuses on coaching autonomous single-agent researchers with an RL recipe that improves planning and power use; it demonstrates functionality features on exterior reasoning benchmarks fairly than proposing a brand new analysis pipeline.
- DeepTRACE [2] (https://arxiv.org/abs/2509.04499) contributes an audit framework that measures traceability and factual assist on the assertion–quotation stage, turning identified failure modes into eight measurable dimensions and revealing giant fractions of unsupported claims throughout methods.
- HERB [3] (https://arxiv.org/pdf/2506.23139) benchmarks Deep Search over heterogeneous enterprise information (Slack, GitHub, conferences, docs), quantifying retrieval problem at scale (39k+ artifacts; answerable & unanswerable queries) and exhibiting retrieval as a main bottleneck for downstream reasoning.
- LiveResearchBench [4] (https://arxiv.org/pdf/2510.14240) is a benchmark of 100 expert-curated duties spanning each day life, enterprise, and academia, every requiring intensive, dynamic, real-time internet search and synthesis. In addition to an analysis suite masking each content- and report-level high quality, together with protection, presentation, quotation accuracy and affiliation, consistency and depth of study.
Collectively, they encourage an enterprise analysis that spans protection/recall, quotation accuracy & auditability, reasoning coherence, and readability, whereas holding a transparent boundary between Deep Search (discovering the best proof) and Deep Analysis (planning, reasoning, and long-form synthesis).
In our inside benchmarks, impressed by these frameworks, we suggest on 5 core dimensions for enterprise analysis:
DimensionWhat It MeasuresWhy It IssuesProtectionHow broadly and deeply the agent explores related data throughout sourcesEnsures completeness; vital for strategic and aggressive analysesQuotation Accuracy & ThoroughnessWhether or not insights are verifiably grounded in credible inside or exterior evidenceBuilds belief and accountability in enterprise selectionsReasoning CoherenceLogical consistency and readability of multi-step reasoningReflects analytical depth; reveals the mannequin “thinks” like an analystReadability & ConstructionReadability, group, and fluency of the ultimate reportMakes outputs usable by enterprise stakeholders and managementInner Data RichnessHow successfully the agent leverages proprietary inside enterprise dataMeasures the system’s potential to mix inside fact with exterior perception
Implementation and Ends in Enterprise
To validate our analysis framework, we benchmarked a number of methods on a various set of Enterprise Deep Analysis stories, that are long-form stories that mix Salesforce’s inside information with exterior information.
Primarily based on our enterprise use case of gross sales report era and motivated by the above enterprise benchmark dimension, we choose for dimensions that want analysis for our pipeline: Report Readability & Construction, Inner Richness / Alignment, Quotation Accuracy, and Protection.
Our outcomes beneath present that whereas readability is basically a solved problem for many fashionable LLM brokers, enterprise grounding and traceability clearly distinguish our Deep Analysis system. It not solely generates fluent, well-structured stories but additionally anchors each perception to verifiable sources and Salesforce’s inside information graph: a basis for decision-grade belief in enterprise AI.
Analysis
DimensionGeminiOpenAISlackBot (Salesforce)Salesforce AIRQuotation Accuracy45.8percent39.5%*79.2%Protection
3.28 / 52.98 / 53.28 / 53.02 / 5Inner Data
Richness42.3 %30.2percent61.4%73.7%Readability & Construction3.6 / 53.6/53.6/53.6/5
*inaccessible org citations
* On this desk, the numbers from Salesforce AIR are utilizing OpenAI GPT-4.1-mini
With belief on the core of Salesforce’s values, we took an additional step to make sure our benchmark outcomes weren’t solely quantitative but additionally human-verified. To validate consistency, we carried out a parallel analysis with skilled human annotators and achieved a Fleiss’ κ (kappa) rating of 0.6, indicating sturdy alignment between mannequin judgments and human evaluations.
Instances like these spotlight the following frontier of Enterprise Deep Analysis, the place safe organizational information might be accessed, reasoned over, and synthesized to energy various enterprise use instances.
In addition they showcase the necessity for enterprise-oriented benchmarks and frameworks that measure not simply accuracy or fluency, however the real-world impression and reliability of AI-driven analysis.
Conclusion
Enterprise Deep Analysis represents a big leap past conventional search, reworking how companies harness data. By integrating adaptive planning, various retrieval, subtle evaluation, and long-form synthesis, it strikes from merely answering “what” to deeply explaining “why” and “how.”
The distinctive challenges of enterprise settings—spanning inside and exterior information, entry management, and strong quotation—necessitate a specialised method. Our multi-graph structure, designed to imitate human analysis, addresses these complexities by orchestrating specialised brokers and instruments. Finally, the analysis of such methods should transcend mere accuracy to concentrate on belief, traceability, transparency, and consistency.
Our inside benchmarks, validated by human consultants, exhibit that our Deep Analysis system excels in offering correct, well-cited, and contextually wealthy insights from each proprietary and public information, setting a brand new customary for decision-grade AI within the enterprise.
Quotation
Please cite this work as:
“`
Chien-Sheng Wu, Prafulla Kumar Choubey, Kung-Hsiang Huang, Jiaxin Zhang, Pranav Narayanan Venkit, “In the direction of Reliable Enterprise Deep Analysis”, Salesforce AI Analysis, Oct 2025.
“`
Or use the BibTeX quotation:
“
@article{wu2025sfrdeepresearch,
writer = {Chien-Sheng Wu, Prafulla Kumar Choubey, Kung-Hsiang Huang, Jiaxin Zhang, Pranav Narayanan Venkit},
title = {In the direction of Reliable Enterprise Deep Analysis},
journal = {Salesforce AI Analysis: Weblog},
yr = {2025},
}
“`
