PRWire: Executives rarely lack information. They lack time to find the right information, reconcile competing accounts and determine what deserves attention. Financial results sit in one system, sales forecasts in another, and the explanations behind both may be scattered across presentations, meeting minutes and email threads.
AI-powered knowledge retrieval and summarization tools offer a way to bring that material together. Used well, they can help management understand what has changed, examine the evidence and prepare for decisions with less manual research. Their usefulness depends on the quality of the information they retrieve, the controls governing access and their ability to distinguish facts from interpretations.
Documented deployments at Cemex, McKinsey, Morgan Stanley and Smartsheet show how these capabilities are being applied. They also reveal an important boundary: evidence of faster information access and strong adoption is more common than evidence of improved strategic outcomes or independently verified financial returns.
Where executive value begins
A conventional dashboard shows performance against selected metrics. A knowledge assistant can add context by retrieving the reports, decisions and operational explanations behind those metrics. An executive asking why a regional business is behind plan may need financial data, pipeline changes, customer developments and management commentary in the same answer.
The most valuable use cases therefore address recurring management questions. Executive briefings can combine financial performance, strategic initiatives, risks and previous commitments into a concise account of material developments. Performance reviews can bring together budget variances and operational commentary. Meeting preparation can retrieve earlier decisions, stakeholder context and unresolved actions before a discussion begins.
Strategic initiative monitoring is another strong application. A project report may describe an initiative as on track while its milestone log shows delays and its steering committee minutes record unresolved dependencies. A system that surfaces that inconsistency provides more value than one that simply shortens the project report. It gives management a specific issue to investigate without prematurely concluding why the discrepancy exists.
Risk reviews, investment assessments and acquisition due diligence can benefit from the same approach. AI can organize evidence across large document collections, compare assumptions and identify unanswered questions. The resulting analysis remains dependent on the completeness of the underlying material and the validity of the methods used to assess it.
Cemex: financial information at executives’ fingertips
Cemex provides a particularly relevant example because its LUCA Bot was developed directly for senior management. A Microsoft account published in February 2026 describes approximately 100 senior leaders with access to natural-language financial and operational analysis. The system covers more than 120 key performance indicators across regions, countries and plants, drawing on a decade of data.[1]
The published technology stack includes Microsoft Foundry, Azure OpenAI, Azure AI Search, Azure Cosmos DB, Azure App Service, Microsoft Teams and Azure Storage. Access is restricted according to each executive’s authorized region and business line. Although responses are delivered quickly, the underlying KPI information is updated monthly—a distinction that matters when interpreting claims about immediate access to information.
Previously, executives could depend on phone calls, email chains and lengthy report searches to obtain answers. LUCA Bot provides self-service access, including while executives are travelling or colleagues are offline. The account reports 400–500 queries per month and internal accuracy measurements of 92% for data and 82% for analysis, assessed using a recurring benchmark.
Those figures demonstrate both usefulness and remaining limitations. They should not be interpreted as guarantees for every question or as proof of financial ROI. They also show why retrieving a figure and producing an analytical explanation need separate evaluation. A system may locate a number correctly while drawing an incomplete conclusion from it.
McKinsey: making organizational knowledge easier to use
McKinsey’s Lilli addresses a different problem: finding and synthesizing accumulated expertise. At launch, the firm described more than 40 curated knowledge sources encompassing research, case studies, interview transcripts and other material, alongside access to its expert network.[2]
For management users, the relevance extends beyond document search. Institutional knowledge can support strategic research, meeting preparation and the identification of comparable business situations. A leader considering a new market or operating model can investigate what the organization already knows before commissioning another research exercise.
A subsequent McKinsey publication reported that 74% of staff used Lilli regularly and that users saved more than 30% of their time on information gathering and synthesis.[3] These are company-reported productivity results across staff, rather than executive-specific measures of decision quality. The cited launch account also does not disclose the complete production architecture.
McKinsey’s adoption approach is instructive. Leadership participation, onboarding, training and the continual introduction of practical features helped make the platform part of everyday work. Access to a capable tool alone does not establish a useful working habit.
Morgan Stanley: building trust through evaluation
Morgan Stanley’s Assistant and Debrief illustrate how retrieval and summarization can support demanding professional workflows. The Assistant provides access to an internal corpus of approximately 100,000 documents. Debrief uses Whisper and GPT-4 to transform consented meeting recordings into notes, action items and follow-up drafts, with CRM integration and advisor review.[4]
Reported adoption of the Assistant exceeded 98% of financial advisor teams. These tools primarily support advisors rather than executive committees, but the implementation lessons apply to management systems: answers must be dependable, meeting outputs need review and quality must be tested continuously.
Morgan Stanley describes expert-graded evaluations, improvements to retrieval methods and daily regression testing. For Debrief, evaluations examine whether the system captures important action items without introducing errors. This is a more meaningful assurance process than checking whether a generated summary reads fluently.
The public evidence supports strong adoption and workflow usefulness. It does not establish a quantified improvement in executive strategic decisions, and the complete retrieval infrastructure is not specified in the cited account.
Smartsheet: delivering knowledge within an existing workflow
Smartsheet embedded its Amazon Q Business assistant, @AskMe, in Slack. The implementation connected public help documents, training materials and hundreds of employee help channels, supporting a workforce of more than 3,300 people.[5]
The announcement describes deployment within weeks without custom code. Executive leaders reported better visibility into projects and less need to request information directly from employees. Quantified time savings and financial ROI were not disclosed.
The example highlights a practical design choice: place the service where users already work. For executives, a familiar collaboration channel or management portal can make retrieval more convenient than a separate application. Convenience must still be paired with appropriate permissions and reliable source coverage.
Convene: supporting decisions at board level
Convene brings AI-powered retrieval and summarization into the boardroom through its AI Companion and Automated Minutes capabilities. According to its published product information, directors can obtain summaries of meeting materials, retrieve context across minutes and document libraries, and access related references before making a decision. These capabilities can help board members prepare for discussions, examine proposals in the context of earlier deliberations and ask more informed questions. Convene also supports drafting minutes from meeting artifacts and creating actions with assigned owners and deadlines, helping connect board decisions with accountable follow-through. Built on Amazon Bedrock, the AI capabilities operate within a board management platform rather than a standalone research tool. Their potential value lies in reducing preparation and information-retrieval effort while preserving continuity between meetings; directors retain responsibility for assessing the evidence and exercising judgment.[8]
How a multi-source solution should work
An executive knowledge system needs to handle documents and structured business records differently. Board papers, policies and meeting minutes contain narrative evidence. Finance systems and CRM platforms contain records that require governed definitions, filters and calculations.
For documents, a common architecture is retrieval-augmented generation, or RAG. Relevant passages are retrieved and supplied to a language model as evidence for an answer. Hybrid retrieval combines keyword search, which is useful for exact names and terminology, with semantic search, which finds related concepts. Ranking then prioritizes the most relevant material. Microsoft’s Azure AI Search documentation describes these patterns alongside incremental indexing and document-level access controls.[6]
Content preparation is a substantial part of the work. Connectors must preserve source identity, document versions, timestamps and permissions. Scanned files may require optical character recognition. Tables need their headings, units and reporting periods retained. Documents split into searchable sections must keep enough surrounding context to preserve meaning.
Structured data requires a separate analytical path. A question about quarterly revenue or forecast changes should use approved queries, APIs or a business intelligence semantic layer. Calculations should be performed by the data or analytics system. The language model can explain the results using relevant commentary, but should not reconstruct material financial totals from scattered paragraphs.
An orchestration layer brings these paths together. It identifies the question’s scope, selects authorized sources, retrieves evidence and invokes approved analytical tools where needed. Complex questions can be broken into smaller searches, but each additional step needs limits, monitoring and evaluation.
A hybrid integration approach is often appropriate: index relatively stable documents, query frequently changing transactional systems through controlled interfaces and maintain consistent business identifiers across both. This reduces the chance of presenting an old indexed snapshot as a current business position.
Managed products can reduce the infrastructure burden. Amazon Q Business, for example, documents enterprise connectors, permission-aware responses, citations and integrations with collaboration tools.[7] A custom application offers more control over business logic and presentation. The choice should follow the sources, security requirements and management workflows involved.
What executive users need from the interface
Executive users need a short answer with a clear route to the evidence. A useful briefing starts with material developments, explains their potential implications and identifies decisions or questions requiring attention. Supporting detail should be available without overwhelming the initial view.
Citations are central to that experience. A material claim should lead to the relevant document passage, record or calculation. A citation is valuable only if the source actually supports the statement. Reporting periods, refresh dates and source coverage should also be visible so users understand the answer’s limits.
Comparison features are especially useful: what changed since the last review, which forecast assumptions moved, and whether agreed actions were completed. Where sources disagree, the system should show the discrepancy rather than silently select the most convenient account.
Decision-support outputs should distinguish observed facts, interpretations, scenarios and recommendations. A budget overrun is an observation. An explanation of its cause requires further evidence. A proposal to reduce scope depends on assumptions about future costs and benefits. Presenting these separately helps executives assess how much weight each deserves.
Meeting continuity also matters. Summaries should preserve decisions, accountable owners, deadlines and unresolved issues. They must retain material qualifications and dissenting views; removing those to produce a cleaner narrative can distort the discussion.
The difficult work: integration, security and accuracy
Data integration problems often precede model problems. Multiple versions of a report, inconsistent customer names and competing definitions of revenue can produce contradictory answers even when retrieval works correctly. Organizations need accountable source owners, shared identifiers and explicit rules for authoritative information.
Information security must operate before retrieved content reaches the model. A user should receive only information they are authorized to access, with permission changes and deletions propagated through the system. Protection must also extend to generated summaries, caches, conversation histories and exports, which may contain the same sensitive information as the original sources.
Retrieved documents must be treated as evidence rather than instructions. Malicious or misleading content embedded in a document should not be allowed to change the assistant’s behavior or authorize tool actions. Restricted tool permissions and targeted testing are necessary parts of the design.
Accuracy testing should examine retrieval and generation separately. Did the system find the right evidence? Did it preserve dates, numbers, caveats and reporting periods? Did it miss a material exception? Does each citation support its associated claim? Can it acknowledge insufficient evidence instead of producing a plausible answer?
The evaluations described by Morgan Stanley and Cemex show why this is an ongoing operating responsibility.[1][4] Changes to models, connectors, documents or permissions can affect behavior after launch. A production service needs regression testing, feedback handling and clear ownership of failures.
Adoption and value must be measured in real work
Executives may experiment with a new assistant and still return to established support teams when decisions matter. Adoption requires a dependable service that addresses a recurring need. Involving executive assistants, chiefs of staff, finance teams and strategy staff can help identify those needs and define acceptable outputs.
The strongest starting point is a specific management process, such as a monthly performance review or a strategic initiative committee. Limit the initial source set, establish a preparation-time baseline and evaluate whether the system produces a verified briefing with less effort.
Success measures should include material errors and omissions, citation quality, preparation time, repeat use during actual reviews and the amount of correction required before an output is accepted. Reduced information requests to support teams and faster assignment of accountable actions can provide additional evidence of usefulness. Operating cost should be assessed per accepted briefing or completed analytical task.
Time saved is a credible early benefit. Better decisions, reduced risk and increased revenue require stronger evidence and careful attribution. High usage alone does not demonstrate those outcomes.
AI knowledge retrieval can make management preparation faster and organizational evidence more accessible. Its lasting value depends on whether executives can understand where an answer came from, assess what remains uncertain and use it within a disciplined decision process.
Sources
- Microsoft, “Faster decisions: How an AI agent is redefining executive workflows at one of the world’s largest building materials companies,” February 12, 2026. https://news.microsoft.com/source/latam/features/ai/cemex-luca-bot-ai-agent/?lang=en
- McKinsey, “Meet Lilli, our generative AI tool that’s a researcher, a time saver, and an inspiration.” https://www.mckinsey.com/about-us/new-at-mckinsey-blog/meet-lilli-our-generative-ai-tool
- McKinsey, “Reconfiguring work: Change management in the age of gen AI.” https://www.mckinsey.de/capabilities/quantumblack/our-insights/reconfiguring-work-change-management-in-the-age-of-gen-ai
- OpenAI, “Morgan Stanley uses AI evals to shape the future of financial services.” https://openai.com/index/morgan-stanley/
- AWS, “Smartsheet Integrates Amazon Q on AWS to Implement New Generative AI-Powered Assistant for its Employees,” July 9, 2024. https://press.aboutamazon.com/aws/2024/7/smartsheet-integrates-amazon-q-on-aws-to-implement-new-generative-ai-powered-assistant-for-its-employees
- Microsoft Learn, “Retrieval-augmented generation (RAG) in Azure AI Search.” https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview
- AWS documentation, “What is Amazon Q Business?” https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/what-is.html
- Convene, “Convene 10.0: Latest Board Portal Release.” https://www.azeusconvene.com/new-releases
