Until now, almost every AI tool was sold as a subscription: you pay monthly whether it helps you or not. Salesforce, the world's largest vendor of sales and customer relationship software, announced at the end of August that it will let customers pay for AI based on results, meaning the revenue it brings in or the costs it cuts (CX Today). You do not need to be a Salesforce customer for this to matter: the principle behind the change is exactly what you can ask of any vendor or agency pitching AI to you.
What changed: you pay when the AI solves, not when it talks
For its support agent, Salesforce now bills only resolutions completed from start to finish without a human stepping in. If the customer asks for a person or leaves unhappy, nothing is charged (Salesforce Ben). Cost is tied to outcomes, not activity.
Their own numbers also show how realistic that outcome is. On the company's support site, the agent handled more than 5 million conversations and resolved 64% of them on its own (CX Today). The remaining 36% still ended up with a human (Outsource Accelerator). Keep that ratio in mind: even at a company with enormous resources, one customer in three needs a person. Nobody can honestly sell you an AI that "solves everything".
Why this matters for a small business
You will probably never buy Salesforce. But when a major vendor agrees to be paid by results, it changes what you can ask of smaller vendors too. When someone proposes a chatbot for your website, an assistant that answers your e-mails, or an automation for quotes, you now have a new question ready: "What happens to the price if it does not work?"
There is also a reason vendors can afford to say yes: the underlying costs are falling. On September 1, Anthropic released the Claude Fable 5.1 model and cut the price by four times for information the model re-reads repeatedly (for example your company instructions or product catalogue), from 1 dollar to 0.25 dollars per million text fragments processed (TechSpot). The company estimates that for assistants working through many steps, the bill can drop by up to 45% (TechSpot). When a vendor tells you AI is expensive, you can ask whether that discount has reached you.
The trap: who decides what a "result" is
This is where you need to pay attention. When you pay by outcome, the definition of the outcome becomes the entire contract (CX Today). A "resolved ticket" might mean the customer got a correct answer, or just that they closed the chat window out of boredom. A "sale brought by AI" might be a sale that would have happened anyway.
Even large companies take the problem seriously. Stripe, the payment processor, warned that a sales lift can come from product tweaks, marketing campaigns or seasonality, not necessarily from AI (AI to ROI). If a vendor asks for a percentage of "AI-generated revenue" without saying how it is measured, you have a contract where you may end up paying for your own sales team's work.
How to apply the principle in your company, in four steps
Here is what it would look like for a plumbing and heating company with three people in the office, receiving 40 quote requests a week by e-mail and phone.
1. Measure the starting point. For two weeks, count: how many requests come in, how fast you reply, how many turn into orders. Without this baseline nobody can prove anything, neither the vendor nor you.
2. Write down what "resolved" means. For example: "the request gets a reply with an estimated price in under 2 hours, and the customer does not write back a second time asking for clarification". One clear sentence anyone can check.
3. Ask for a 30-day pilot with a human in the loop. The AI drafts the reply, a colleague approves it with one click. You count how many drafts were sent without changes. If fewer than 60% pass, it is too early to take the person out of the loop.
4. Price the measured result, with an exit. A small subscription plus a fee per request resolved according to the definition from step 2, with the right to cancel at 30 days. A vendor who refuses any link to results is telling you something about how confident they are in their product.
Data caution: what you hand to a vendor working for results
To deliver results, an AI vendor will ask for access to your e-mail history, tickets or quotes. Before you hand it over, a few simple rules:
- Do not send customers' personal data (names, phone numbers, addresses, ID numbers), signed contracts, passwords or confidentially negotiated prices. For testing, an anonymised export where names become "Client 1", "Client 2" is enough for the AI to learn your tone and the types of requests.
- Sign a data processing agreement (DPA) with the vendor. If your company is in Romania or anywhere in the EU, you are the data controller under GDPR and you are responsible for what your vendors do with customer data.
- Ask where the data is stored, how long it is kept and whether it is used to train models. Get the answer in writing.
- When the AI replies to customers on behalf of your company, tell them. It is both fair and required by European transparency rules.
A well-run pilot does not need all of your company's data. It needs a clean sample and a clear definition of success.
Conclusion: one concrete step for next week
You do not need to buy anything now. Pick a single repetitive process (quote requests, frequent questions, appointments) and measure it for two weeks: volume, response time, conversion rate. With those three numbers in hand, every conversation with an AI vendor changes: you bring the baseline, they have to prove the improvement. That is exactly what large customers have started demanding from Salesforce. You can do it too.
Sources
- CX Today, "Salesforce's Outcome-Based Pricing: Strategic Masterstroke or Defensive Scramble", August 27, 2026: https://www.cxtoday.com/ai-automation-in-cx/salesforces-outcome-based-pricing-strategic-masterstroke-or-defensive-scramble-against-ai-seat-compression/
- Salesforce Ben, "Huge Agentforce Pricing Shift: Salesforce Introduces Pay-Per-Resolution": https://www.salesforceben.com/huge-agentforce-pricing-shift-salesforce-introduces-pay-per-resolution/
- Outsource Accelerator, "Salesforce's outcome-based pricing tests CX budgets", September 2, 2026: https://news.outsourceaccelerator.com/salesforce-outcome-pricing/
- TechSpot, "Anthropic releases Claude Fable 5.1, cuts cached token pricing by 75%": https://www.techspot.com/news/113701-anthropic-releases-claude-fable-51-stronger-coding-science.html
- AI to ROI, "News & Analysis: September 4, 2026": https://ai2roi.substack.com/p/ai-to-roi-news-and-analysis-september