Sell Outcomes, Not Hype: Agentic AI for Leaders With Warren Schilpzand

Warren Schilpzand is the Regional Director for SambaNova Systems. It was co-founded in 2017 in Palo Alto, Calif., by three industry luminaries in artificial intelligence: Kunle Olukotun, Rodrigo Liang and Christopher Ré. They develop a full-stack AI platform, including custom hardware and software, to run enterprise AI applications efficiently.

Agentic AI For Sales Leaders:

The first time I heard Warren explain inference, I realised I’d been lumping all “AI stuff” into one mental bucket. Training, usage, models, magic — all stirred into a single soup. Then he said, “Inference is the execution of the model,” and my sales brain finally clicked. Training is building the engine. Inference is turning the key, driving the car, and winning the race.

This post is your simple, practical guide to agentic AI — the next phase that turns LLMs from neat demos into revenue machines. You’ll get the difference between training and inference, the agentic architecture that makes systems feel intuitive, real use cases (mortgages, expense policies, retail, travel, government), and a concrete plan to get your team pitch-ready. My aim is to make you dangerous in the room — the kind of leader who can sell, shape, and steer outcomes when AI shows up on the agenda (which is now… everywhere).

1) Stop Blurring “Training” And “Inference” — They’re Different Sports

Training: the lab.

Warren drew a hard line. Training is the behind-the-scenes marathon: massive datasets, long runs, huge energy and water consumption. The audience at this stage is data scientists. The output is a model — a capable brain that’s learned how to perform a task.
Inference: the arena.

Once trained, the world shifts to inference — using the model inside applications. This is where customers (and impatience) live. Inference must be fast, cheap, repeatable, and available to millions. If the system hesitates, users bounce. If it’s clunky, they never come back. Inference is what we sell.
Why sales leaders should care.

Budgets will increasingly be split by phase: training platforms vs. inference infrastructure. Different KPIs, different buyers, different value stories. If you sell inference right, you’ll attach to customer experience, conversion, and cost-to-serve — not “AI research.” That’s a bigger, closer-to-revenue conversation.

Field anecdote: the latency line

Warren’s voice example is brutal and brilliant: if your AI is in a live conversation, you’ve got ~200ms to respond or you break the illusion. Past that, humans feel the lag, the spell snaps, and adoption craters. Your deal will ride on whether your stack can hit that number reliably.

Field anecdote: the abandoned journey

We’ve all rage-quit a bot that took too long. That’s not “user impatience,” that’s an inference throughput problem. If you can quantify tokens-per-second, concurrency, and cost-per-1k tokens at peak, you can translate engineering into pipeline impact: fewer drop-offs, more completed journeys, higher NPS.

Field takeaway: sell outcomes that infer controls

Speed-to-answer (SLA) on AI flows
Concurrent sessions handled at peak
Cost-per-inference vs. CX metrics (AOV, conversion, CSAT)

If you can draw a line from inference metrics to revenue or returns reduction, you’ll beat any “AI strategy” deck.

2) Agentic AI: When The System Decides The Steps (And Feels Human)

Agents choose the path.

Agentic AI is an architecture where agents decide how to get the job done. You give them the tools, rules of engagement, and goals. They sequence actions, iterate as needed, and hand back results. One agent can work solo, but the real magic is orchestration — dozens or hundreds collaborating inside an experience.
Context is the fuel.

The secret isn’t re-training — it’s context engineering. You feed agents the right history (customer data, policy docs, prior decisions, best-practice steps). The system uses that context to make decisions. Change the context, and you change the behaviour — without another expensive training cycle.
Why does this matter operationally?

This lets business owners change outcomes by updating the “source of truth” (a policy doc, a playbook, a product catalogue), not raising tickets for engineering sprints. Faster iteration, safer governance, fewer brittle rules engines.

Story: the finance policy that became the system

Warren’s bank example is a gem. Two agents: one that “reads” the finance policy doc and interprets it; another that coordinates expense reviews. No rules coded. No fine-tuning. If Finance updates the policy doc, the system adapts. That’s a governance dream — and an IT backlog diet.

Story: the sales dinner that was compliant

The agent didn’t just see “$1,000 crab.” It read the justification, inferred a client dinner with seven people, matched that to per-head allowances, and green-lit it. That’s nuance you don’t get from a rigid rules engine — and it prevents the painful “manual exception” merry-go-round.

Field takeaway: sell “business-owned logic.”

“Update the document, not the code.”
“Adapt in hours, not sprints.”
“Governance in the hands of the policy owner.”

3) Use Cases Your Buyers Already Feel (And Will Fund)

Mortgage & applications: get it right in one sitting.

Forms are fine; back-and-forth is not. As users enter details, agents ask for missing pieces in real time, clarifying subjective answers (“tell us more about this expense”). Result: a complete submission first time, faster decisions, and a measurable lift in “first responder wins.”
Retail & digital sales: bring the shop assistant online.

Websites make you know what you want. Great sales assistants don’t — they ask, guide, and recommend. Agentic AI turns your PDPs and search bars into a helpful human-like conversation that reduces cart abandonment and nudges buyers to the right product the first time.
Travel: sell the trip, not the ticket.

Warren’s story of booking LA flights then getting irrelevant hotel suggestions is classic. An agent could infer Disneyland or Universal from “two adults + two kids to LAX,” ask one smart question, and keep the entire wallet in one flow. That’s not “AI magic,” that’s better inference + context.

Government & services: the endless loops end here

Think Service NSW on good days — then everywhere else. Matching citizens to services is a context problem with subjective inputs. Agentic flows that prompt, clarify, and assemble a complete case upfront reduce case time and human ping-pong. Same headcount, faster outcomes, happier citizens.

Bottom-line levers buyers recognise

Warren’s caution is fair: “productivity” without headcount change is hard to book on a P&L but returns reduction is concrete. Help buyers pick right, return less — that’s a line item you can defend. Ditto call deflection with high CSAT. Sell the levers that finance can see.

Field takeaway: lead with top-line, close with bottom-line

Top-line: higher conversion, higher AOV, faster decisions
Bottom-line: fewer returns, fewer manual exceptions, lower cost-to-serve

4) The Inference Stack: Latency, Throughput, And Cost-Per-Outcome

Latency: protect the illusion.

Voice and live-chat require sub-200ms round-trips to feel conversational. That’s end-to-end: model + network + orchestration. Bake SLAs into proposals and reference real-world concurrency.
Throughput: millions of tokens, many sessions.

Your customer won’t test it with ten users; they’ll hit peak before lunch on a promo day. Size for the surge. The killer slide: “At 5× peak traffic we maintain <200ms and £X per 1,000 tokens.”
Cost-per-outcome: the metric the CFO understands.

Move from “pennies per 1k tokens” to “cost-per-conversion, cost-per-completed-application, cost-per-deflected-contact.” Tie the inference economics to business goals and you’ll outmanoeuvre point-solution competitors.

Mini-playbook: discovery questions I now ask

Which journeys must feel conversational (<200ms)?
What’s peak concurrency, and what breaks today?
Which P&L line item do we move first (conversion, returns, deflection)?
Who owns the context (policies, catalogues, playbooks) and how often does it change?

Mini-playbook: proof you can demo

Expense policy agent reading the actual Finance doc
Retail assistant who narrows the choice in three turns
Mortgage pre-check that completes the case in one session

Field takeaway: pitch the SLA, not the sizzle

If you anchor the conversation on SLAs and outcomes, you’ll feel like an operator, not a futurist. Buyers relax when they see numbers.

5) Context Engineering Beats Perma-Fine-Tuning (Most Days)

Fine-tuning is expensive — and slow to iterate.

Every policy change shouldn’t trigger a training job. Keep training for when you truly need new capabilities. For behaviour, prefer context.
Agentic context gives you levers everyone understands.

“Want a different result? Update the policy/playbook/product data.” That’s change management your stakeholders already know how to run.
Governance and audit live where they should.

If Finance edits the finance doc, Legal reviews it, and the agent follows it — you’ve aligned AI with existing control frameworks, not invented a new one.

Story: the two-agent expense checker

One agent “knows” the policy because it reads it. The other orchestrates cases. Together, they interpret nuance (the seven-person client dinner) and still enforce guardrails. Fewer escalations, faster resolution, clearer audit.

Story: forms that talk back

Whether it’s mortgages or benefits, the agent prompts for clarity the moment ambiguity appears. That’s fewer incomplete submissions and faster cycle time — the two KPIs operations directors crave.

Field takeaway: teach your team the phrase “agentic context.”

It signals you understand how behaviour changes without retraining. It’s the secret handshake for the serious crowd.

6) How To Get Your Team Ready To Sell Agentic AI

Get passionate or get passed over.

Warren’s first principle is blunt: if you don’t love what you represent, your customers won’t either. Your energy is the transmission medium.
Embodied discovery.

If you’re pitching an airline, talk like an airline exec who cares about load factors, ancillaries, and on-time performance — then map agentic AI to those. You’re selling outcomes in their language, not acronyms in yours.
Learn, teach, move on.

The space shifts weekly. Build a learning cadence: master a use case, teach your team, then free yourself to chase the next frontier. If you hoard, you stagnate.

Team drills (I run these now)

90-minute lab: Turn a policy PDF into an agent behaviour change.
Role-play: 5-question discovery that lands on a measurable P&L lever.
Deal review: Replace “productivity” fluff with one quantified line item.

Assets to create this quarter

One-page talk track: Training vs. Inference (with SLAs)
Demo script: Expense agent reading a live policy doc
ROI sheet: Retail returns reduction model (before/after)

Field takeaway: certify for outcomes, not buttons

Shift your enablement from “how to click” to “how to land value.” Buyers remember the latter.

7) Your First 30/60/90 Days: A Sales Leader’s Agentic OS

Days 0–30: Anchor the fundamentals.

Define your top three journeys where latency matters. Map owners of context. Quantify peak loads. Write the one-page “inference SLA” you’ll take into every meeting.
Days 31–60: Build credible demos + numbers.

Create two agentic demos that change behaviour by editing context. Pair them with a simple calculator that models conversion lift or returns reduction.
Days 61–90: Land lighthouse wins.

Target one account per vertical where the P&L lever is obvious. Propose a 6–8 week pilot with hard SLAs and a crisp exit to production.

Governance & risk talking points

We keep logic in the documents owners already control.
We log decisions and surface rationales for audit.
We size for peak, test for 200ms, and price by outcome.

Field takeaway: sell small, scale fast

Start with a narrow, valuable journey. Win the week. Then win the year.

Summary

Agentic AI is what happens when LLMs stop being a lab project and start acting like your best colleague: fast, contextual, and capable of choosing the next best step. Training builds the model; inference wins the customer. Context engineering lets business owners steer behaviour without retraining. The use cases your buyers will fund first are the ones they already feel — completing applications in one sitting, buying with confidence, getting answers without the phone queue, and stopping the endless “exception” emails. Lead with top-line, close with bottom-line, and anchor your pitch on SLAs and measurable outcomes. Then coach your team to love the work, speak the customer’s language, and keep moving. That’s how you’ll sell agentic AI in 2026 and beyond.

Sell Outcomes, Not Hype: Agentic AI for Leaders With Warren Schilpzand