Webinar Catch-Up: Five Lessons on Agentic Revenue Operations

By Anna Zavalishina, Business Development Lead, Europe, ideaHelix
Anna leads business development for ideaHelix in Europe, working with Salesforce account teams and customers across EMEA. ideaHelix is a Salesforce Summit Consulting Partner specializing in Agentforce Revenue Management and Salesforce Industries, with automation tools available on AppExchange.

Introduction

At Salesforce Tower London, I hosted Michael Kiruba-Raja, Senior Director, Agentforce Revenue Management at Salesforce, and our own Sasha Morozov, Principal Architect at ideaHelix, for an honest conversation about what it really takes to make AI work across revenue operations.

Key Takeaways

1. Sometimes you shouldn’t use AI

One of the most useful lessons was also the least fashionable.  Plenty of revenue problems are better solved with tools Salesforce customers already have: a flow, an integration, a report or a list view.  As Sasha put it, when you have a shiny new tool, you start trying to solve everything with it.  Don’t.

Traditional automation is often faster, cheaper and more predictable than an agent.  And sometimes AI simply isn’t an option. Security, data residency, GDPR and cost can all be hard constraints.  Where AI gets interesting is when you need to reason across data that has never lived in one place: ERP, product catalog, contracts, quotes and billing.

Bring that together and an agent can answer questions no report ever could.  That’s the use case worth funding.

2. Keep the initial scope narrow

“We need to bring AI into the business” is not a project.  Start with one use case you can explain in two sentences and measure with a number.  Then build the evaluations around it.  That evaluation layer is often skipped. We make it part of the implementation methodology for a reason.

3. Build for the day it fails

AI will fail.  Models go down, get things wrong and change over time.  So build the fallback from the start.  For us, that means people.  Every agent in a revenue process needs a clear route back to someone who understands the job.  That doesn’t weaken the automation. It makes it trustworthy enough to use.  And it doesn’t stop at go-live.  You launch, measure, evaluate and keep improving.

4. Out-of-the-box gets you most of the way

Michael described the split between Salesforce and partners as an 80/20 rule. It’s a useful way to think about it.  Salesforce provides the platform and increasingly capable out-of-the-box agents.

A catalog agent can generate and localise product information. A customer-facing agent can answer a question about an invoice, trace it back through the order and quote, and bring in a person when needed.

Then comes the last 20%.   That’s where industry knowledge matters.  Telecom product modelling doesn’t look like insurance. Insurance doesn’t look like manufacturing.  So we package that experience into accelerators: industry-specific product catalog agents, deal review agents and tools built around the real bottlenecks in quote-to-cash.

Because configuring the quote often isn’t the problem.  Getting it approved and contracted is.  And because Agentforce Revenue Management (ARM) exposes so much through APIs, those accelerators can plug into the platform rather than work around it.  Claudeforce takes this further, bringing quote-to-cash workflows into where people already work.

5. AI-native delivery is more than code generation

Everybody talks about AI-native delivery.  Usually they mean generating code faster.  We looked at the whole implementation lifecycle instead: where AI helps, where it doesn’t, and where we don’t want it involved.  On top of Salesforce’s ARM capabilities, we’ve added skills around solution design, industry-specific catalog work and CI/CD.

The results vary, but we’ve seen improvements of up to 70%.  In some cases, work that might previously have needed 15 people can be handled by three.

But there’s a catch.  Code generation gets much faster. Human review doesn’t disappear.  People are still in the loop, and if the underlying process is unclear, AI won’t fix it.

The human part is the point

None of our customers are really talking about AI as a way to reduce headcount.  They’re talking about making the teams they already have more effective.  And customers still prefer dealing with people when it matters.  Agents build trust when they solve the problem and hand over cleanly when they can’t.  They destroy it when they’re designed to frustrate customers until they give up.  That’s a design choice, not a technology limitation.

Thanks to Michael and Sasha, and to everyone who joined us live.