Dovetail Software Just Named the AI Skill That Separates the Winners From the Noise

“The technical term is context engineering. Essentially just trying to figure out how to apply the general-purpose model in a specific way.”

That line, from Benjamin Humphrey, co-founder and CEO of Dovetail Software, cuts straight to a problem most teams are quietly struggling with.

They have deployed an AI agent. They have connected a data source, set up a workflow, and waited for the insights to roll in.

What arrived instead was noise. Generic summaries. Themes that don’t map to reality. Outputs no one trusts enough to act on. The agent technically worked. The intelligence did not.

The companies getting real value from agentic AI, Humphrey argues, solved the data problem before they deployed anything.

The Prompt Was Never the Problem

A general-purpose model knows a great deal about the world and almost nothing about your business.

It cannot tell you why support tickets spiked last Tuesday. It has no memory of the product decision your team reversed three quarters ago. Feed it a raw export of sales calls and it will return something that sounds plausible. Whether it is accurate is a different matter. This is the ceiling every team hits when they treat AI as a shortcut rather than a system.

Context engineering is the practice of closing that gap, structuring data, defining scope, and designing the conditions under which a model can reason reliably about your specific situation. It is less about what you ask the model and more about what you have taught it to understand before the ask arrives.

Fifty support tickets can be skimmed manually. Fifty thousand cannot. “You can’t really pass all of that into a language model,” Humphrey explains, “because it’s greater than the context window that the model can handle.”

So That’s What the Clustering Is Actually For

The process Dovetail Software uses to turn raw, unstructured feedback into boardroom-ready intelligence involves several discrete steps, and skipping any one of them is where the output falls apart.

It starts with sampling. Incoming data, whether pulled from Salesforce, Gong, Figma integrations, or direct uploads, is broken into its component signals. Each support ticket, call transcript, or survey response is reduced to its key points rather than processed wholesale.

Those key points then go through clustering. The model identifies the broad themes that exist across the full dataset, without anyone having to define those themes in advance. Once the clusters exist, each data point gets assigned to the theme it belongs to.

“We do what’s called clustering to determine the actual broad themes across that large data set,” Humphrey says, “and then we work through it and bucket things into those themes.”

What comes out is a structured, time-stamped body of intelligence that can be tracked week over week, queried in natural language, and pushed automatically to wherever the relevant team works.

The quantitative shape of qualitative data, volume of a theme rising or falling, sits alongside the raw feedback that explains why.

“The cool thing about it is it’s kind of like the quantitative meets the qualitative,” Humphrey says.

What an Agent Actually Does on a Monday Morning

Getting the architecture right is what makes agentic AI useful rather than decorative.

Dovetail Software’s agents do not wait to be asked. A CEO’s inbox receives a weekly voice of customer report, compiled from everything that entered the platform that week, with week-over-week memory intact. A head of sales gets a top-deals briefing in Slack, pulling from call transcripts to surface the objections appearing most frequently and the competitors coming up in live conversations. A product lead gets flagged before a pattern becomes a crisis.

“You might want to get a weekly voice of customer report in your email inbox,” Humphrey explains. “What the agent would have done is basically look at all the data within the product, synthesise it into some themes and then send that to you, and it has memory as well, so you can track things week over week.”

This is what separates a genuinely agentic system from a scheduled report. The agent is reasoning over a live, structured dataset that has been deliberately engineered to make that reasoning reliable.

Teams not doing that groundwork will keep running into the same problem. The agent produces an answer. The answer cannot be traced to a source. No one knows whether to trust it. The insight dies before it reaches a decision.

Why Good Data Has to Come Before Good Agents

None of this happens automatically, and Humphrey is direct about that.

The companies winning with agentic AI are the ones that treated data consolidation as infrastructure before they deployed a single agent. That means pulling sales calls, support tickets, and survey responses out of their separate silos and putting them somewhere that a model can reason across them coherently.

“Executives need to break down the silos between their functional teams,” Humphrey says. “What happens is you get these departmental managers that gatekeep access to their data. The sales team doesn’t really want to share their calls. The customer success team doesn’t want to share their calls.”

Image from Dovetail Software