The decisions behind the product
Practical perspectives on Claude, MCP, enterprise data, model operations, and the decisions behind trustworthy analytics.
63 articles · newest first
Connector workflows
Why ServiceNow SLA breach counts disagree
Separate incidents from SLA instances, choose a clock and reconcile assignment groups before interpreting a breach rate.
· 3 min read
Connector workflows
Find SAP fulfillment gaps without double counting
Analyze SAP order lines, deliveries and billing at the right grain, using a worked partial-delivery example.
· 3 min read
Connector workflows
Oracle Fusion aging starts with installments
Use installment-level due dates and remaining balances to make an Oracle Fusion receivables question reproducible.
· 3 min read
Connector workflows
Find the delay in a Snowflake analytics answer
Separate application work, warehouse queueing and query execution before changing models or warehouse size.
· 3 min read
Connector workflows
Give BigQuery questions a scan budget
Turn an open-ended business question into a bounded BigQuery query without hiding incomplete results.
· 3 min read
Connector workflows
Test the identity behind a Databricks answer
Check Unity Catalog policy behavior, application credentials and cached answers with a two-persona evaluation.
· 3 min read
Cross-system workflows
Build a Salesforce to NetSuite revenue bridge
Explain CRM-to-finance differences through identifiers, event dates and allocation rules rather than forcing totals to match.
· 3 min read
Cross-system workflows
Connect Salesforce demand to SAP fulfillment
Compare commercial demand with order-line fulfillment using explicit mappings, dates and quantity conventions.
· 3 min read
Cross-system workflows
Review service incidents beside renewal exposure
Combine ServiceNow service issues and Salesforce renewal context without inventing a causal churn score.
· 3 min read
Cross-system workflows
Reconcile Oracle Fusion with a Databricks model
Separate refresh lag, transformations and permission scope when a lakehouse finance measure differs from its operational source.
· 3 min read
Buying guide
A practical scorecard for an enterprise analytics pilot
An enterprise analytics pilot needs a scorecard that reflects the intended business workflow.
· 4 min read
Buying guide
Bring three business questions to an analytics demonstration
A tailored analytics demonstration is more useful when the buyer supplies three questions from an existing workflow.
· 4 min read
Solution walkthroughs
Solution walkthrough: a regional business review in Claude
This walkthrough describes an illustrative business meeting, not a named customer deployment or measured result.
· 4 min read
Solution walkthroughs
Solution walkthrough: from sales pipeline to cash questions
This solution walkthrough uses a hypothetical business review to show how pipeline and cash questions can be organized.
· 3 min read
Innovation Notes
Design human review around the decision at risk
Human review works best when it focuses on the decision at risk, not every low-impact step. Design it around what a person must understand and authorize.
· 4 min read
Innovation Notes
Evidence should travel with an AI answer
Business answers move from an application into meetings, documents, messages, and decisions.
· 4 min read
Innovation Notes
A useful refusal is a product capability
A refusal is useful when it explains the relevant boundary and helps an authorized user take an appropriate next step.
· 4 min read
Innovation Notes
Innovation begins with stable business meaning
Enterprise AI becomes more useful when the meaning of a business question remains stable across interfaces, teams, and model changes.
· 4 min read
Evaluation
Follow-up questions should keep their scope visible
Follow-up questions often omit details because the user expects the previous scope to remain in effect.
· 4 min read
Evaluation
An ambiguous question should produce a useful next step
People ask questions such as "How are we doing?" because shared context makes them useful in human conversation.
· 4 min read
Evaluation
Measure the time from question to decision
The time a model takes to generate text is only one part of an analytical workflow.
· 4 min read
Evaluation
Compare models without changing the question set
A model comparison should hold the business task and acceptance criteria steady while recording the configurations being tested.
· 4 min read
Evaluation
Build an evaluation set from decisions people make
An evaluation set should represent the business decisions the application is intended to support.
· 4 min read
Governance
Keep enough evidence to reproduce a decision
An evidence record should make a material analytical decision reviewable without creating an uncontrolled copy of the underlying data.
· 4 min read
Governance
Permissions are part of answer correctness
An analytical answer is not correct for a user if it includes records that person is not authorized to access.
· 4 min read
Governance
Review refusals alongside successful answers
Review analytical refusals alongside successful answers. A refusal may correctly protect a permission boundary or acknowledge missing evidence.
· 4 min read
Integrations
Prioritize connector requests with evidence
A connector roadmap should distinguish an available integration, an evaluated pilot, and a candidate request.
· 4 min read
MCP engineering
Keep identity boundaries explicit in a custom MCP server
A custom MCP server sits between an assistant and one or more protected systems.
· 4 min read
MCP engineering
A useful MCP tool contract is smaller than an API catalog
A useful MCP tool contract names a business or analytical operation that can be described precisely.
· 4 min read
MCP engineering
Design an MCP adapter for an app without one
An application without an MCP server may still expose an API that can support a carefully designed adapter.
· 4 min read
Models & MLOps
Plan a model rollback before changing production
A model rollback needs more than an old model name. Preserve the prior configuration, prompt, tool contract and dependencies so you can restore a tested state.
· 4 min read
Models & MLOps
Make evaluation part of a model release
Changing a model can alter factual accuracy, tool choices, response format, latency, and refusal behavior even when the application code stays the same.
· 4 min read
Models & MLOps
An OpenAI-compatible endpoint still needs compatibility testing
An OpenAI-compatible endpoint can reduce integration work, but the label does not guarantee identical behavior for every model or API feature.
· 4 min read
Models & MLOps
Evaluate NVIDIA NIM as an inference option
Evaluate a specific NVIDIA NIM model and hardware profile against your application's contract, workload, recovery needs, and operating responsibilities.
· 4 min read
Models & MLOps
Size an inference GPU around the workload
An inference GPU should be selected against the model, context length, concurrency, latency target, and serving software.
· 4 min read
Perspective
Why we do not publish an accuracy number
Every vendor in this category leads with a percentage. Here is the arithmetic that convinced us ours would be misleading, and what we publish instead.
· 6 min read
Perspective
What a refusal should sound like
Most systems fail by answering anyway. The wording of a decline is a design decision, and the difference between useful and useless is specificity.
· 5 min read
Perspective
Connect first, model later
The conventional order is model everything, then let people ask. That sequencing is why so many analytics programs stall before anyone gets an answer.
· 5 min read
Models & MLOps
What local inference does and does not keep local
Running a model locally changes where inference executes. It does not automatically keep every part of an AI application local.
· 4 min read
Models & MLOps
Choose an open model by its operating requirements
Choosing an open model requires more than comparing a leaderboard score.
· 3 min read
Models & MLOps
Evaluate Anthropic and OpenAI on the same business questions
Evaluate Anthropic and OpenAI using the same business questions, authorized data scope, reference answers, and acceptance criteria.
· 4 min read
NetSuite
A finance-first checklist for validating AI answers
Finance validation should compare an AI answer with an approved report under the same definition, period, entity scope, currency, and access context.
· 4 min read
NetSuite
Why credit memos belong in the question
Credit memos and other adjustments can materially change a revenue, receivables, or customer-balance answer.
· 4 min read
NetSuite
Make an accounts-receivable aging question reproducible
An accounts-receivable aging answer depends on a cutoff date, an age basis, bucket boundaries, and the treatment of credits and payments.
· 4 min read
NetSuite
Subsidiaries and currencies need explicit boundaries
A financial answer across subsidiaries needs an explicit entity scope and currency basis.
· 4 min read
NetSuite
Ask finance questions against the right close calendar
Finance questions often depend on posting periods, fiscal calendars, and close status rather than simple calendar dates.
· 4 min read
NetSuite
Revenue, invoices, and cash are different questions
Revenue, invoicing, and cash collection describe different financial events.
· 4 min read
Salesforce
Make record visibility part of your Salesforce evaluation
A Salesforce total is a statement about the records visible under a particular access context.
· 4 min read
Salesforce
Reconcile a conversational forecast with Salesforce
A conversational forecast should be reconciled with the exact Salesforce view the business uses.
· 4 min read
Salesforce
A useful question about stale opportunities
Define which Salesforce signals warrant a review, validate false positives with managers, and keep record hygiene separate from sales judgment.
· 3 min read
Salesforce
Account hierarchies can change a revenue answer
A Salesforce account can represent a legal entity, a branch, a buying unit, or a parent group.
· 4 min read
Salesforce
Agree on the denominator before asking for win rate
Choose the opportunity population, denominator, period, and count or amount basis before comparing Salesforce win rates across teams.
· 4 min read
Salesforce
Why pipeline needs an as-of date
A pipeline answer is incomplete without an as-of date. Opportunity stages, expected close dates, amounts, and ownership can change after a report is produced.
· 4 min read
Claude & MCP
What should a Claude and MCP pilot prove?
A Claude and MCP pilot needs an explicit decision at its end: what evidence justifies wider adoption, a narrower scope, or stopping the rollout.
· 4 min read
Claude & MCP
Build a repeatable Claude analytics demonstration
A repeatable analytics demonstration uses a defined question set, known reference results, approved identities, and a recorded configuration.
· 4 min read
Claude & MCP
Separate analytical questions from business actions
Asking why a number changed is different from authorizing a change in a business system.
· 4 min read
Claude & MCP
Test Claude with two different business roles
Permission testing belongs in the business evaluation of an AI integration.
· 4 min read
Claude & MCP
Review the evidence behind a Claude answer
An AI-generated explanation can be fluent even when its underlying scope differs from the reader's intent.
· 4 min read
Claude & MCP
When Claude connects but the data question fails
Separate Claude sign-in, tool availability, source access, and question-definition failures before changing permissions or reconnecting accounts.
· 4 min read
Claude & MCP
Choosing remote or local MCP for enterprise analytics
Remote and local MCP connections solve different operating problems. A remote connector reaches a hosted service over the network.
· 4 min read
Claude & MCP
A practical approval checklist for Claude connectors
A connector approval should identify the business use, the people allowed to use it, and the data they need.
· 4 min read
Claude & MCP
MCP connects tools. Who defines revenue?
MCP lets an assistant discover and call tools, but it does not decide what revenue or churn means. Pair the connection with an explicit metric definition.
· 4 min read
Claude & MCP
Claude and MCP: start with one business question
A useful first question for Claude and Nexatron is one whose answer already has an owner.
· 4 min read