Executive summary: On November 11, 2026, Databricks CEO Ali Ghodsi showed a reporter how he uses the company's updated AI agent, Genie, to get the things he would otherwise never hear. Genie analyzes the parts of an employee-satisfaction survey where the company scored low, summarizes a sales pitch that went poorly (including the exact features the prospect needed before buying), and even preps him for interviews. The capability is not "another dashboard." It is an always-on data agent that runs multistep jobs unsupervised against the information a company already has, and returns the unvarnished truth about what is not going well. The number behind it is Dunbar's number - roughly 150 people - the maximum you can keep an honest relationship with. Below it, a founder hears the bad news directly. Above it, nobody tells you the bad stuff anymore. For a small business, the practical opportunity is a truth agent you can point at the data you already have, with a cost guardrail you control. The playbook below is four steps: name the number you are most afraid to look at, point an agent at your existing data, set the spend guardrail, and schedule the one-page truth weekly. EP54 of The AI Guy Podcast walks the same four moves in audio.
Every agent so far has been built to make you faster. The truth agent is built to make you honest, and that is a different and rarer capability. The obstacle is not a people problem - it is a physics problem. Dunbar's number, named after the British anthropologist Robin Dunbar, is about 150 people: the maximum number of stable relationships a human can maintain. In a startup of forty, the salesperson who bombed the big pitch tells the founder straight up, because the relationship is direct. At 12,000 employees, that same salesperson will not confess a disastrous call to the CEO. The HR manager will not flag that their idea for raising satisfaction will fail. The dashboard shows the number you want to see, not the number that needs attention. You find out about the real problem two weeks after it happens, from the customer who is leaving. Ghodsi put it plainly: "Anyone who has more than 150 employees, which is a lot of companies, a lot of people can get value out of this." The implication for a small business is direct. You are past the point where you can know everyone, long before you are past the point where it matters. The truth agent is the thing that closes the gap while you are still small enough to act on it.
Tools & pricing: Databricks Genie (enterprise data platform; pricing is platform-based and usage-scaled - verify current pricing and availability for your data residency). The capability - an agent that answers the uncomfortable question in plain English against your existing data - is the part you should replicate at SMB scale, not necessarily the specific product.
This is the whole move in one sentence. Not the number that looks good on the slide. The number that tells the truth. The churn rate on your three biggest accounts. The gross margin on the product you thought was your best. The median response time on your support queue, not the average (the average hides the tail where the angry customers are). The percentage of new customers who came back for a second order. Write it down. One number. The reason this is step one and not step four is that a truth agent without a named target is just a chatbot. Give it the question it should keep asking, and it has something to do. The number you are most afraid to look at is, by definition, the one most likely to be quietly wrong.
Tools & pricing: A notebook, a whiteboard, or your existing BI tool. Free. The cost is the five minutes it takes to name the number honestly.
You do not need to buy a data warehouse to run a truth agent. You already have the data. The invoices, the support tickets, the survey results, the CRM, the churn log. What you need is an agent that can read what you have and answer the uncomfortable question in plain English. The mechanism Databricks describes is an ontology - a layer that gives the agent the context it needs from the information customers already store on the platform. For a small business, the practical version is simpler: connect the agent to the one or two systems of record behind the number you named, and ask it to explain what moved and why. The constraint to respect: do not wire up twelve tools on day one. One system of record, one number, one question. If the agent can tell you, in plain English, why the churn rate ticked up last month, you have more truth than you have had in years.
Tools & pricing: Your existing BI or the agent your team already uses (a general assistant with tool access, pointed at your database or CSV exports). No new system required to start.
Here is the part that decides whether this pays or quietly becomes an expense line nobody watches. Databricks handles it with its Unity Gateway product, which lets a company set usage limits and swap in less expensive open-weight models. The principle transfers directly to a small business. You control spend two ways. First, cap usage - a monthly token or query budget, with a hard stop. Second, route by difficulty: the cheap, high-volume work (pulling last week's support tickets, counting returns) goes to the cheap open-weight model; the hard work (diagnosing why a big account is churning) goes to the strong model. Set the rule before you scale it. The guardrail is not a limit on capability. It is the difference between a tool that saves you money and a bill that surprises you. Baseline it: how much does the agent cost per month at the cadence you plan to run, and does the one decision it surfaces pay for that ten times over?
Tools & pricing: Usage caps + model routing (a cheap open-weight model for volume, a strong model for diagnosis). The numbers are yours to set; the rule is what matters.
An agent you have to remember to ask is a dashboard in disguise. The move is to make the truth automatic. Weekly. One page. The number you named, what moved, what the agent thinks it means, and one thing to do. Five minutes, every Friday, before you open the rest of your inbox. The cadence is the point. On-demand truth is a report; scheduled truth is a habit, and a habit is what changes the decision you make. The realistic ceiling is not dramatic - it is the difference between finding out about the churned account at the exit interview and finding out about it the week it started drifting. That is a two-week lead time on the single most expensive thing in your business. And because the agent runs in the background, the marginal cost of running it weekly is close to zero once the guardrail is set.
Tools & pricing: A calendar invite + your agent. No new software. The cost is five minutes a week.
Two things worth stating plainly. First, the Genie capability is an enterprise feature on Databricks' platform, aimed at organizations with data already on that platform; it is not a drop-in for a small business on a SaaS stack. The transferable insight is the pattern - an always-on agent that answers the uncomfortable question against data you already have - not the specific product. Second, an agent is only as honest as the data behind it. If your numbers are wrong, the truth agent returns a confident wrong answer. So step two is really "make the number you named trustworthy, then point the agent at it." The agent does not remove the need for judgment. It removes the need to be the last to know.
The direction is clear. The next wave of agents is not about doing more of what you already do. It is about telling you the things you would not hear otherwise. That is the capability that compounds, because the companies that survive are not the fastest - they are the ones that find out about the problem while it is still small. And the number behind it is the one that matters. 150. Below that, you are fine. Above that, you are already guessing, and the truth agent is the thing that stops the guessing. Baseline it this week: write down the one number you are most afraid to look at, point the agent you already have at the data behind it, set the spend guardrail, and schedule the one-page truth every Friday. The fastest company wins the quarter. The honest company wins the decade.
Your next move is one number. Write down the metric you are most afraid to look at this month. Point your existing agent at the data behind it, set the usage cap, and schedule the one-page truth for next Friday. To get the full SMB agentic-AI strategy, including how the truth agent stacks with the in-suite agents and where the decision intelligence actually pays for itself, start here:
Get the SMB AI adoption strategy