There is a chart you can draw on a whiteboard that explains almost every strategic decision DRS Automation Group has made in the last twelve months.
The horizontal axis is time. The vertical axis is agent capability — how well the system handles the actual conversations a business has. You plot two lines.
The first line is a horizontal AI agent — a general-purpose assistant configured for a particular business. It starts low. It climbs steeply for the first two weeks. It reaches roughly the eightieth percentile of the conversations the business actually has — and then it stays there. Forever. Every gain after that is marginal, expensive, and political.
The second line is a vertical agent — an agent built from the ground up for a single industry. It starts lower than the first line. It takes longer to climb. It reaches the eightieth percentile around month three. Then, while the first line stays flat, this line keeps going. Slower than at the start. But it doesn't plateau. At month twelve it is at ninety-three. At month twenty-four it is at ninety-seven. It compounds.
The horizontal agent is a good employee. The vertical agent is a discipline.
This note is about why those two lines diverge. The mechanism is not mysterious, but it is rarely articulated, and almost every decision an AI consultancy makes — what to build, what to refuse, how to price — flows from understanding it.
01 / Why the 80% Is Easy
Every industry has a long tail of conversations and a short head. In pest control, the short head is: what services do you offer, when can you come out, where are you located, how much does it cost, do you treat for mosquitoes. Six or seven questions, asked a thousand different ways, accounting for the majority of inbound traffic. In auto dealerships, it's: is this vehicle still available, what's the out-the-door price, can I trade in my current car, what financing do you offer, do you take Apple Pay.
Modern general-purpose language models handle the short head almost out of the box. Give a frontier model a competent prompt, a basic knowledge base, and a couple of tools, and it will be conversational on the first ninety percent of caller intent within a week. This is what gets a demo built. This is what gets the first contract signed.
It is also where the trouble starts — because the operator now believes the system understands pest control. It does not. It understands the questions pest control customers ask the most often. Those are not the same thing.
02 / What Lives in the 20%
The remaining twenty percent of conversations is where every dollar gets won or lost. It is also where vertical knowledge actually lives. To make this concrete, consider three real moments from production:
A pest-control caller says: “I think we have termites. There's like, sawdust on the windowsill.”
A general-purpose agent hears this and offers to schedule an inspection. A vertical agent hears this and knows two things the operator knows: first, the “sawdust” is almost certainly frass, and frass on a windowsill is diagnostic of a specific kind of termite activity. Second, this caller needs to be routed to a specific technician with a specific license type — not any tech in the area. The general agent books an appointment. The vertical agent books the right appointment, with the right person, in the right priority slot. Same conversation, different outcome, hundreds of dollars of margin difference per call.
An auto-dealership caller says: “What's my payment if I put five thousand down?”
The general agent says: “That depends on financing terms.” The vertical agent knows the dealership runs three lenders for sub-prime, two for prime, and a captive finance arm for the manufacturer in question. It also knows that asking “what's the payment” before asking “what credit tier do you fit” is the second-worst move on the floor. The vertical agent guides the conversation to the question that comes before the payment question — which is the question every veteran salesperson would ask. Same conversation, different conversion rate.
A field-services caller says: “Same as last time, Tuesday is fine.”
The general agent stalls — it does not know what “last time” refers to. The vertical agent stalls too, on day one. But on day forty it has been trained on the operator's CRM and recognizes that “same as last time” from a known phone number means a specific service code at a specific recurring frequency, which it can confirm and book without escalation. This is not a model upgrade. It is a vertical-knowledge upgrade. The horizontal agent will still be stalling on this question in year two.
03 / The Learning Surface
The reason the vertical line keeps climbing is that every edge case the vertical firm encounters becomes part of the system. The general-purpose agent can't do this. It serves a thousand verticals. An edge case from pest control is noise to it. The vertical agent serves one. Every edge case is signal.
Three things compound for a vertical:
- Domain ontology. The list of things that exist in the industry — frass, termite swarmers, IPM protocols, GLB designations, ECC codes, captive finance, recourse vs. non-recourse paper. The general agent has surface knowledge. The vertical agent has a working ontology. Every new edge case adds a node.
- Conversation patterns. The specific ways customers in that vertical phrase confusion, hesitation, objection. “Is this the actual price” means something very specific at a car dealership. The vertical agent has heard it ten thousand times. The general agent treats it as one phrasing among many.
- Integration depth. The CRM. The scheduling tool. The lender APIs. The route optimizer. The accounting system. The technician's phone app. The general agent connects to none of these natively. The vertical agent is built with them. Every additional integration deepens the moat.
Each of these compounds independently. Each new client in the vertical adds material to all three. The vertical agent at month thirty-six is qualitatively better than the same agent at month six in a way the horizontal agent simply cannot match.
04 / Why The Model Doesn't Save You
The objection is: but the models keep getting better. Won't a future general-purpose model just handle the twenty percent natively?
The answer is: the models are getting better at language, not at your industry. A frontier model in 2027 will still not know whether your particular pest-control franchise charges by the linear foot or the perimeter, whether your dealership's used-car policy allows the appraisal number over the phone, or that your field-services dispatcher considers Friday afternoons a soft-block. These are not language facts. They are operational facts. No amount of pre-training reaches them.
The model is a substrate. The vertical is an application. Better substrates make better applications cheaper to build. They do not eliminate the application.
If anything, the cheaper substrates make the vertical layer more valuable, not less. When everyone has the same model, the only thing that matters is what you've built on top of it. That is — almost by definition — vertical.
05 / What This Means If You're Buying
If you're an operator evaluating an AI agent, the question to ask is not which model are you using. The question is: how does this system get better at my industry specifically, over time, after I sign?
The bad answers are: we'll upgrade you to the next model when it comes out, we have a knowledge base you can edit, our agent is trained on industry data. These describe horizontal systems.
The good answers describe a learning loop. Every customer conversation the agent gets wrong becomes a training case. Every operator correction becomes a rule the agent applies forever. The agent's domain ontology grows month over month. The agent's integrations deepen. If you can't see the compounding mechanism, there isn't one.
06 / What This Means If You're Building
If you are building an AI consultancy, this note is the strategy memo. Pick a vertical. Dig.
Don't pick three verticals. The compounding only works if you stay long enough to compound. Pick the one vertical where the operator language, the integrations, the regulatory environment, and the customer-conversation patterns are deeply legible to you, and where your reputation as a domain operator will compound alongside your technical reputation.
Then build the same agent for every client in that vertical. Not a bespoke build per client — the same agent, configured per client. Every client's edge case becomes the next client's default behavior. Every client makes the agent smarter for the rest. The first client subsidizes the second's onboarding. The hundredth client is paying for a system that already understands their industry at a depth no horizontal vendor can match.
This is the holding-company logic that explains why DRS Automation Group separates its work into vertical products — PestPilot for pest control, LotPilot for independent auto dealerships — rather than offering a single “DRS AI Agent” that adapts to whatever vertical the next prospect happens to be in. The vertical brand is a forcing function. It commits us to staying long enough in one industry that the compounding actually shows up.
Horizontal AI looks faster at the start. It costs less to ship the first version. It can quote a lower price. It can demo a wider set of use cases. In the first three months of any engagement, the horizontal vendor will win the comparison.
The vertical bet is a bet on month thirty. By month thirty the horizontal vendor is selling the same eighty-percent agent it sold on day one, and the vertical operator is sitting on three years of compounding domain knowledge no competitor can replicate without paying the same tuition.
Most AI vendors will not be around in month thirty. The ones that are will be vertical.