Why the long game matters for AI in GTM
Most revenue teams are still running AI one rep at a time. Airspeed co-founder and CTO Devang Agrawal thinks that phase is ending. On Lodestar with host Cliff Simon, Devang argued that 2027 belongs to teams that deploy agents consistently across the whole org, and that getting there takes the kind of long-term work many leaders under quota pressure keep postponing.
Devang came to sales from AI research: Siri at Apple, then the Google DeepMind team scaling large language models before ChatGPT. The moment that made leaving feel obvious was a Christmas pub quiz at DeepMind. Someone entered a language model, and it did surprisingly well across every category. One model answering questions from every domain was the signal Devang and co-founder Adam Liska needed to start a company.
Below are the ideas from the conversation that matter most to revenue leaders.
Why build a context graph before agents?
Airspeed started as a conversational intelligence platform with a bet behind it: language models would get very good, and what they would need to do real work was a context graph. That means every call and email in your org, captured and labeled, so an agent knows what it is looking at.
The labeling is the point. Plug a general AI assistant straight into a call recorder and it cannot tell a discovery call from a renewal. With unlimited tokens it could read every conversation and work that out itself. Airspeed labels each call as it happens, so an agent goes straight to the right calls instead of spending 100 million tokens reinventing the data model.
Devang said the agents were not ready until late 2025 and early 2026. With the latest generation of models, Airspeed put an agent harness on top of the context graph, so agents can navigate it and use tools to do the work.
What does that look like on a real question?
Devang walked through an analysis from Airspeed’s own board prep: why is my conversion rate dropping?
- The agent scanned the CRM and flagged a process change. A new kind of deal had started appearing, which was skewing the numbers.
- It suggested removing those deals and rerunning the analysis.
- The clean data showed the biggest drop sat between stage 1 and stage 2, and that it had grown over the last couple of quarters.
- It normalized for reps who had joined recently and found the signal still held.
- It then spun up 100 subagents to analyze every stage 1 call.
That is the vague leadership question every RevOps analyst knows, worked through with the full context in place. The value is not only the answer. It is getting the answer while people are still paying attention.
Why should GTM leaders think in systems now?
Early 2026 was the experimentation phase. Teams deployed a couple of agents, bought a vendor with a few predefined ones, or gave everyone a Claude license and let each person work their own way. Devang called that the right move for its time.
The problem is that it lifts individuals, not teams. Your top performers build agents for themselves and pull ahead, while the team benchmark barely moves. Systems thinking asks a different question: what can we automate across the whole org, so every rep’s numbers move?
The practical advice: learn what an agent harness is. Instead of building agents one at a time, you build a framework that holds the context, tools, and memory, then build as many agents as you need on top of it.
Cliff named the hard part: governance, enablement, and enough buy-in across the org to keep leveling it up. Devang agreed, and added that buying licenses and calling it done is not a strategy. By 2027, in Devang’s view, it will not be optional.
Why do short-term fixes hurt long-term capability?
Cliff described a pattern that keeps showing up. Companies miss their growth number, EBIT falls, and leaders go after short-term gains to keep the ship from sinking. Each quick fix makes the eventual turnaround harder.
Devang pointed to private equity. The old model was to hold a company for three years and flip it on short-term gains. That no longer works. You have to do the long-term work up front so the company you sell has better metrics, not a better quarter. Devang sees more leaders starting to think this way, and expects it to become essential.
Why does a CTO spend so much time with sales leaders?
Airspeed is a London company building a team in New York. Devang’s view: you cannot dip a toe into the US market. Airspeed went all in after its Series A, built a cluster in New York, and is seeing much more success since.
The travel is about being in the room with buyers. A product feedback call tells you what people say. A conference full of CROs and sales leaders shows you what they care about and how they talk to each other. “If I was building developer tools, I could have just sat at home,” Devang said, because an engineer already knows that persona. Building for sales and RevOps leaders means embedding with them.
The model Devang points to is the AppDynamics and Harness ecosystem: technical founders who grew fast by learning why buyers cared, not by leaning on the technology alone. Engineers sometimes see a small change as trivial, then watch it transform the customer’s experience.
What keeps a founder grounded?
The episode ends on the personal side. Devang has two boys, three and one. The goal is to be home for bedtime, then go back to work once the kids are asleep. Mornings, in Devang’s words, are “survival”: negotiating with two hyperactive kids to get them ready for school. A nanny and a supportive spouse make the travel possible.
Devang also reflected on a mix of cultures: the warmth of India, the calmer pace of the UK, the intensity of the US, and a soft spot for Dutch directness. As Cliff put it, saying what you mean is often the more caring choice.
The takeaway
Agents are no longer the hard part. The hard part is the long game: a context graph your agents can trust, a harness that turns one rep’s experiment into the whole team’s workflow, and leaders willing to invest before the quarter forces their hand.
Watch the full Lodestar episode, or read Devang’s take on why agent harnesses built for individuals break at enterprise scale. If you want to see the context graph and agent harness on your own pipeline, book a demo.
Frequently asked questions
What is a context graph in sales?
A labeled model of everything your revenue team says and does: calls, emails, deals, and the people on them. Airspeed tags every call as it happens, for example as a discovery call, so an agent can find the right conversations later without reading every transcript from scratch.
What is an agent harness?
A shared framework that holds the context, tools, and memory agents need, so you can build as many agents as you want on one foundation instead of wiring each one up by hand. Devang Agrawal's advice on Lodestar: stop thinking about individual agents and start thinking about the harness.
Why isn't a Claude license for every rep enough?
Because it lifts your best users, not the team. Top performers build their own agents and pull ahead while everyone else stays put. Systems thinking means deciding what to automate across the whole org, setting governance, and enabling every rep, so the team benchmark moves instead of one rep's.
Why can't an AI tool connected to a call recorder answer pipeline questions?
It can, slowly and expensively. Without labels it does not know which calls were discovery calls, so it has to read every conversation and work that out first. A context graph built ahead of time lets the agent go straight to the calls that matter and answer while the question still has someone's attention.
