What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Onkar Deokate’s case for adding AI to an existing CRM is straightforward: the bottleneck is often getting from a business question to an answer, not a lack of stored data. Rather than ask a company to replace its CRM, he is building a conversational layer intended to query existing department data and preserve the workflows people already use. This is his build rationale, not proof that every CRM has the same problem or that the proposed design has been validated across companies.
The problem he wants to solve is the path to an answer
Deokate describes a familiar kind of CRM task: answering, “Which leads from this university haven’t been followed up in 7 days?” In the workflow he recounts, a user sets filters, waits for a table, exports the results to CSV, and shares them. He estimates that this takes “three minutes minimum.” That is his estimate for this example, not an independently measured industry benchmark.
As an Amazon Associate I earn from qualifying purchases.
His alternative is to ask a question in ordinary language, with “show me this week’s drop-offs” as another example. As he puts it: “The problem isn’t the CRM. The problem is that there’s no fast path from a question to an answer.” The claim is about friction between a question and CRM data, rather than a general finding that CRMs lack useful information.
What the proposed AI layer is meant to do
Deokate describes Company/OS as a plug-and-play AI layer over existing CRMs, offering a natural-language interface to department data. The proposed approach is to add business context and query capability while leaving the underlying CRM in place. He names Salesforce, HubSpot, and custom-built internal tools as possible targets, but does not establish integrations, compatibility, or adoption for any of them.
The design he describes brings together four components: a business knowledge graph, five scopes of memory, department packs, and five reasoning modes. They are parts of his proposed system, not evidence of measured accuracy or proven performance across organizations.
A graph intended to connect business context
The system is described as ingesting code, product requirements, database schemas, Slack conversations, and Jira tickets, then connecting entities, relationships, and business rules in a property graph using Neo4j. Deokate’s architectural distinction is that vector search returns semantically similar text chunks, while a graph can be traversed to follow relationships between pieces of business context. That is his framing of the design trade-off; it does not establish that graphs universally reason better than vector stores.
Five scopes of memory
Deokate separates memory into five scopes so that business context can be represented at different levels:
Rank #2
- Global: the system-wide scope.
- Company: local terminology and business rules.
- Department: what Sales, Finance, or Audit can see and query.
- User: individual preferences.
- Conversation: ephemeral state for the current session.
These scopes are the author’s described design. The account does not establish how permissions are enforced or provide evidence about security properties.
Department Packs for different work
Department Packs are intended as bundles of workflows, query templates, and reasoning skills. The examples Deokate gives are:
- Sales: lead scoring, follow-up cadence, and conversion funnels.
- Finance: invoice reconciliation, commission calculations, and multi-currency settlements.
- Audit: compliance checks, document verification, and anomaly flags.
He says this structure drew on experience at one company and needs more research with users at other companies before it is expanded. These examples therefore describe a proposed model, not a demonstrated fit for every department or organization.
Five reasoning modes
The system assigns tasks to five named modes:
- Search: find a record.
- Advisor: suggest a next action.
- Executor: take an action.
- Analyst: aggregate insights.
- Developer: explain code or schema.
Deokate reports that Advisor handled 70% of queries in his account, but gives no sample size or independent validation. He also says he spent two weeks optimizing LangGraph routing before realizing he should first instrument which modes users actually needed. The lesson in his account is to measure demand before spending time refining routing.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Why add a layer instead of migrating to a new CRM?
The distribution argument is continuity: keep the CRM, workflows, and user habits, and add an interface over the data rather than undertaking a replacement. Deokate contrasts what he calls a one-day deployment for a layer with a 6–18 month enterprise sales cycle for a new CRM. Both are his claims, not universal timelines or a measured comparison of implementation effort.
A replacement may involve migration and retraining; a layer avoids making replacement the first step, but it does not eliminate technical work. The author says existing schemas can contain opaque columns, missing foreign-key constraints, or absent documentation, so ingestion still needs cleaning and annotation. His account does not quantify that effort or show that the layer works with every CRM.
For an organization assessing the idea, the meaningful questions are practical rather than categorical:
- How much data, process, and user behavior would a replacement require the company to move or change?
- Can the existing CRM’s schemas be understood, cleaned, and connected to the needed business context?
- How should access differ by department, and how will those boundaries be enforced?
- Have intended users been consulted about department-specific workflows and query patterns?
- Will the team instrument which reasoning modes people use before optimizing their routing?
Those considerations follow from the trade-offs Deokate identifies. His article does not provide a measured comparative evaluation of an AI layer versus CRM replacement, nor establish customer adoption or query accuracy.
What the build rationale establishes—and what it does not
Deokate’s argument is a case for changing the interface to CRM data before changing the CRM itself. The described architecture offers one way to organize context, permissions by department, workflows, and query tasks. It remains a proposed system in his account: the source does not establish cross-company validation, confirmed integrations, measured accuracy, or security performance. His own stated cautions—messy schemas, limited user research for packs, and the need to instrument routing—are important parts of the rationale, not minor implementation details.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
