By James W. Moore
Salesforce Just Bet Its AI Business on Outcomes, Not Activity.
On June 25, Salesforce announced Agentforce Help Agent, a pre-packaged customer service agent that can be deployed across chat, voice, web, and portal channels. The bigger news wasn’t the product. It was the pricing. Starting in July, customers using Help Agent pay $2 per resolution, and only when the agent actually closes an issue without a human escalation or an unhappy customer walking away. No resolution, no charge.
This is the third pricing model Salesforce has tried for Agentforce in under two years. It launched with a flat “$2 per conversation” fee that the market found confusing and easy to game, since a conversation wasn’t the same thing as a result. That gave way to consumption-based Flex Credits last year. Now Salesforce is tying cost directly to a defined outcome: a resolved issue, full stop. Kishnan Chetan, the Salesforce executive who leads Agentforce Service, said the company is running the same model on its own help portal, where it has handled 4.3 million inquiries and closed 70% of them without human involvement.
The shift is not without risk on Salesforce’s side. Every interaction still costs Salesforce tokens and compute, resolved or not, and a customer motivated to game the system by claiming dissatisfaction could drive up Salesforce’s costs without ever paying for a resolution. Salesforce is effectively betting its own margin on the reliability of its model.
Why This Matters for Insurance:
Outcome-based pricing has been discussed in insurance AI circles for a while, mostly in the context of claims automation and underwriting copilots. Salesforce putting real dollar figures and a public commitment behind it changes the conversation from theoretical to standard-setting. Every carrier, MGA, and agency management system vendor now selling AI tools into this industry is going to face the same question from buyers: why am I still paying a subscription or seat fee for a tool that’s supposed to replace work, rather than paying for the work it actually completes? Agencies evaluating AI vendors this year have new leverage to ask for pricing tied to resolved claims, completed quotes, or successfully bound policies, not licenses per user.
Sources:
- Huge Agentforce Pricing Shift: Salesforce Introduces Pay-Per-Resolution — Salesforce Ben
- Salesforce Unveils AI Help Agent With Pay-Per-Resolution Pricing — CIO
AWS Is Betting $1 Billion That AI Deployment Should Take Days, Not Months.
On June 30, AWS announced a new Forward Deployed Engineering organization backed by a $1 billion investment. The idea: embed AWS engineers, working alongside purpose-built AI agents, directly inside customer teams to build and ship production AI systems in days rather than the months a traditional consulting engagement takes. Francessca Vasquez, the AWS vice president who announced the effort, described it as agentic-first, timeline-compressing, and designed so customers are self-sufficient once the engagement ends. Named early customers include the NFL, the NBA, Southwest Airlines, and Cox Automotive.
The mechanism is worth understanding. AWS teams deploy a semantic layer into the customer’s own AWS account that connects to internal data sources and builds a governed knowledge graph, so AI agents reason over the customer’s own institutional knowledge rather than a generic model. Customer engineers move through the engagement from observers to co-builders to independent operators, and AWS says every project leaves behind documentation, runbooks, and trained internal staff.
Why This Matters for Insurance:
AWS explicitly named regulated industries, financial services, and government as its target for this model, citing the governance and speed-to-production demands those industries carry. That’s a direct pitch at carriers and larger MGAs still stuck between pilot programs and production AI. But the self-sufficiency framing deserves a second look. It answers, at least on paper, the dependency risk Microsoft’s Satya Nadella raised last week when he warned that AI concentration in a handful of vendors is politically and economically unsustainable. A carrier’s decision about whether to buy that self-sufficiency promise, or whether it becomes another form of vendor lock-in with extra steps, is worth scrutinizing before signing anything.
Sources:
Agentic Commerce Is Coming for Every Industry With a Middleman.
A McKinsey report published June 10 on the state of European e-commerce lays out a shift that reaches well past retail. The core claim: AI agents are starting to shop on behalf of consumers, researching, comparing, and in some cases completing purchases without a human ever browsing a page. McKinsey estimates that by 2030, $3 trillion to $5 trillion in global B2C retail revenue could flow through this kind of agentic commerce, and cites its own research showing 38% of European consumers already use generative AI tools to help decide what to buy.
The structural consequence McKinsey draws out is the one worth sitting with. Retailers are no longer competing primarily for a customer’s attention. They’re competing to be selected by an algorithm acting on that customer’s behalf. Product data, pricing logic, and fulfillment reliability become inputs to a machine decision rather than marketing to a person. Boris Ewenstein, CEO of the German retailer Otto, compared the shift to earlier moves from catalogs to online to mobile, calling AI the next paradigm change in how customers shop.
Why This Matters for Independent Agents:
Swap “retailer” for “agent” and the argument reads almost unchanged. If AI assistants increasingly research coverage options, compare quotes, and shepherd a consumer or small-business owner through a purchase decision, the agents who get selected will be the ones whose offerings, appetite, and pricing are structured and legible enough for a machine to evaluate confidently. The agents who aren’t legible to that system risk becoming invisible to it, regardless of how good their actual advice is. This isn’t a five-year-out problem. It’s the same mechanism already reshaping retail, arriving in a market where trust and relationship have always been the agent’s strongest asset. The open question is whether that asset survives translation into a form an algorithm can recognize.
Sources:
From the AI World: How Do You Trust an Answer From an AI You Didn’t Build?
Two stories this week, from very different corners of the AI industry, are circling the same problem: how do you know an AI’s answer is actually trustworthy, and what happens when someone finds a way to make it lie convincingly?
Anthropic published a detailed account on June 30 of the process behind restoring access to its Fable 5 and Mythos 5 models, which the U.S. government had ordered pulled on June 12 over export control concerns tied to a reported cybersecurity workaround. Beyond the timeline, the more interesting part is what Anthropic revealed about how it defends against misuse. The company uses layered safety classifiers with a deliberate “safety margin,” meaning some clearly benign requests get blocked anyway to reduce the odds that a genuinely harmful one slips through. Anthropic is now working with Amazon, Microsoft, Google, and other partners on a shared industry framework for scoring how severe a given AI “jailbreak” actually is, based on how much capability it unlocks, how broadly it applies, how easy it is to pull off, and how widely known the technique becomes.
Meanwhile, insurance technology vendor Vertafore held a webinar introducing ReferenceConnect AI, an enhancement to its longstanding insurance knowledge platform. The pitch is narrower but points at the same problem: rather than pulling from the open internet, the tool grounds its answers in vetted insurance publications and carrier-approved content, and shows users the source material behind each response. Vertafore’s early testing found users located key information up to 80% faster than with traditional keyword search, while still being able to verify the answer against a citation.
Why This Matters for Insurance:
Every carrier and agency now evaluating an AI tool, whether it’s a general-purpose model or a niche platform built for underwriting or claims, is going to run into a version of this same question: what happens when the tool is wrong, and how would anyone know? Anthropic’s approach is a severity framework for classifying how bad a failure could be. Vertafore’s is source citation baked into every answer. Different scales, same underlying instinct: an AI tool without a visible way to check its work is a liability in an industry built on defensible decisions. When evaluating vendors, ask what “grounded” actually means in their system and how they know when it’s wrong before you do.
Sources:
- Redeploying Fable 5 — Anthropic
- How AI Is Helping Carriers and MGAs Make Faster, More Defensible Decisions — Carrier Management
If You Found This Useful
This newsletter exists because insurance executives and independent agents need sourced, direct analysis on AI, not vendor marketing dressed up as news. If a colleague would benefit from receiving it, forward this issue or share the link.
And if you have not yet downloaded The Independent Insurance Agent’s AI Playbook, it is available free at insuranceindustry.ai/ai-playbook. More than 33,000 words of practical guidance, no cost, no catch.
Sources
- Huge Agentforce Pricing Shift: Salesforce Introduces Pay-Per-Resolution — Salesforce Ben
- Salesforce Unveils AI Help Agent With Pay-Per-Resolution Pricing — CIO
- AWS Invests $1 Billion to Embed AI Forward Deployed Engineers With Customers — About Amazon
- Europe’s New E-Commerce Agenda: How AI Is Resetting Growth and Competition — McKinsey
- Redeploying Fable 5 — Anthropic
- How AI Is Helping Carriers and MGAs Make Faster, More Defensible Decisions — Carrier Management
AI Disclaimer: This content was created with assistance from artificial intelligence technology. While content is based on factual information from the source material, readers should verify all details directly with the respective sources before making business decisions.
AI Disclaimer: This content was created with assistance from artificial intelligence technology. While content is based on factual information from the source material, readers should verify all details directly with the respective sources before making business decisions.
