Your AI Does Not Know How Your Business Works, with Saahil Dhaka, Clientell

What we talked about.

In this episode of the Sunny Ray Show, host Sunny Ray talks with Saahil Dhaka of Clientell, calling in from Dubai, about why business AI is smart but does not understand how a company actually works. Dhaka explains the gap between AI capability and the tacit knowledge that experienced employees build over time, such as taste, gut feel and unwritten discount habits. Clientell, started in 2021 as a reporting layer on the CRM, now builds a governed context graph that captures that know-how from systems like Salesforce, SAP and NetSuite. Dhaka describes plugging into Claude, OpenAI, Gemini and other MCP clients, and offering custom agents and a managed service, which he says has driven the most growth. He covers early compliance work, a plan-first approval workflow, the future of the Salesforce admin, and the value of model-agnostic infrastructure amid rising usage costs. He also shares a real example of an AI report that summed euros and dollars without converting them, and advises documenting a clear don't-do list before giving agents system access.

AI can be brilliant and still not know how your business works. Saahil Dhaka explains how to fix that with a governed context graph.

The questions, and the answers.

What is Clientell building, and what problem are you solving?

The core problem is that AI doesn't understand business. It lacks the tacit knowledge and taste that someone who has done a job for a while builds up. Coding has verifiable feedback, but marketing or design doesn't. So we're building a context graph, or induction database, that extracts the know-how people never write down and gives it to AI.

Why did reports from your early CRM reporting layer keep falling short?

People had hundreds of reports they never looked at. The real issue was underlying business context. With multiple products and geographies you get different currencies, and even state-of-the-art models can add them without converting. Back then the goal was simply standardizing data from sheets and the CRM so we could measure anything before improving it.

Why does the most powerful tool belong with the person closest to the work?

The people at the edge of the organization hold most of the context. For example, your best account executive gave a 20% discount last year, then left, and now a renewal is coming. That knowledge isn't in the workflow or CRM notes. It's probably in a chat or a Google doc, and unless you extract it, you never know.

Why a governed graph instead of just guessing or using a vector database?

Vector databases dumped everything in, and people found it didn't work because business work has a lot of implied meaning. A deterministic graph is like Google Maps with annotations plus a local guide's knowledge. Baking that into a structure lets us push the AI down a predictable path instead of letting it wander.

Which offering do customers want most: MCP, custom agents or managed service?

Managed service has driven the most growth for us. Many businesses realize AI-native development isn't their core business. If you're a manufacturer in Texas with 200 people, you don't want to build revenue, operations and IT functions and also manage your own CRM. You can offload all of that to us.

What happens to the Salesforce admin job if agents can administer Salesforce?

It's a wakeup call, similar to what's happening with developers. Admins will need to adapt, learn a bit of coding and tools like Claude Code, and manage agents such as analytics and support agents. If they do, they'll have almost zero manual work and can focus on strategy and building new agents for the business.

What should a company document before giving an agent access to its systems?

A structured don't-do list. AI can lack common sense. I haven't worked on our front end in three months, and if I ask Claude to build something it may use an outdated version and waste twelve hours of tokens. People focus on what to do, but I also focus on what not to do.

Where should a human always give final approval?

At the very beginning, not the end. If you discuss the approach up front and approve a concrete plan, AI rarely makes mistakes afterward. Many people wait until the work is finished and then dislike the output, having spent hours of tokens. Approving at the plan level avoids that.