Featured image: AI-generated conceptual illustration. Sources checked 26 October 2026.
One of the clearest ways an AI agent can contribute to revenue is by helping a business serve demand it already receives. The challenge is not always finding more prospects. Sometimes the company has interested buyers waiting for accurate answers and too little human capacity to respond well.
OpenAI described that situation in its September 29, 2025 account of an internal inbound sales assistant. The case is useful because it connects a business bottleneck, a feedback process, and a reported commercial outcome. It is also easy to overstate. Read it as a company-reported example, not an independently audited blueprint for guaranteed agent earnings.
What the source actually supports
In OpenAI's published account, the assistant used internal information to answer inbound questions and hand qualified prospects to sales representatives. The company reported that its first-email evaluation accuracy improved from about 60% to more than 98% through a feedback process involving representatives. It also reported that the initiative unlocked multimillions in annual recurring revenue.
Those statements have important boundaries. The accuracy figure concerns the company's described first-email evaluation, not universal factual accuracy or an overall conversion rate. Annual recurring revenue is not net profit, cash collected, or the provider's revenue from selling an agent product. The source does not disclose a controlled causal estimate or the full cost of implementation and operation.
That is enough to motivate a practical question: could better handling of existing inquiries improve a particular business's sales process? It is not enough to promise that a small agency can reproduce the same result by connecting a model to an inbox.
Diagnose the demand bottleneck
Before building anything, inspect a sample of recent inbound inquiries. How many received a timely, relevant answer? Which questions recur? Which require a specialist? How often do prospects disappear after a generic response? The point is to identify an observable service gap, not to assume every unanswered lead was ready to buy.
Segment the workload. A small-business buyer asking which plan includes a feature differs from an enterprise buyer asking about security, procurement, or integration. A useful assistant can answer supported questions and route the rest. It should not improvise legal or compliance assurances because a prospect wants a quick answer.
Check whether the bottleneck is genuinely response preparation. If the product is not suitable, pricing is unclear, or the company cannot fulfill demand, faster replies may have limited value. An agent cannot repair a weak offer by making confident statements about it.
Build an answer system before a sending system
Create an approved source library containing current product facts, plan definitions, permitted claims, escalation rules, and customer examples that may be shared. Assign an owner to every high-impact source. A stale pricing page or obsolete security statement can turn a fast response into a costly mistake.
Prepare drafts with source links for internal review. Separate the prospect's question from the assistant's interpretation, and mark anything unsupported. Keep draft creation, approval, sending, and CRM recording as distinct states. This makes it easier to measure quality and prevents a failed send from being confused with a failed answer.
Use explicit handoff criteria. A prospect requesting a negotiated contract, an exception to policy, or an answer beyond the approved material should reach a qualified person with context intact. The assistant's job is to reduce repeated explanation, not trap the buyer in an automated loop.
Design evaluations around commercial risk
Build a representative test set from authorized, appropriately handled historical inquiries. Include ordinary product questions, ambiguous requests, contradictory context, unavailable features, and questions that must be escalated. Ask reviewers to grade factual support, relevance, tone, and correct routing separately.
Anthropic's agent-evaluation guidance is a useful technical reference for testing outcomes. For inbound sales, the outcome may include a correct answer and the right CRM handoff, not simply a fluent email. Evaluate the complete workflow before exposing real prospects to it.
Keep a holdout set and repeat tests after changes. A prompt that improves personalization may accidentally increase unsupported promises. Track the types of corrections representatives make. Frequent corrections about one feature may indicate a source-library problem rather than a model problem.
A small-business pilot with honest arithmetic
Consider an illustrative software firm receiving 300 eligible inquiries monthly. Suppose its current process produces 30 qualified conversations. A pilot may aim to improve response coverage and quality while keeping representative review time manageable. Do not begin by promising a particular number of sales; establish the baseline and the measurement plan first.
If the pilot generates five additional customers at US$200 monthly recurring revenue each, that is US$1,000 of additional monthly recurring revenue, assuming those customers are genuinely incremental. It is not US$1,000 of profit. Delivery costs, support, discounts, churn, implementation, and sales effort still matter. Without a comparison group or another credible design, even the incremental attribution remains uncertain.
Track response time, answer correction rate, handoff acceptance, qualified-conversation rate, conversion, and customer complaints. Compare similar lead sources and time periods. A marketing campaign that changes lead quality can make the agent appear better or worse even when its behavior is unchanged.
What an agency can sell responsibly
A practical offer includes discovery, source-library preparation, a reviewed pilot, CRM integration, and ongoing evaluation. Charge for those deliverables and defined operating responsibilities. Avoid selling a vague autonomous salesperson with unlimited outreach or a guaranteed revenue number.
The best-fit customers have genuine inbound demand, reliable product information, a clear human escalation path, and enough volume to justify the integration. Poor-fit customers may need a better website, clearer pricing, or a disciplined manual response process first.
The lesson from the case is a business-design principle: use agents where better service can unlock existing opportunity, then measure what actually changes. Keep company-reported results attributed, financial terms precise, and promises proportionate to the evidence. That is more credible than turning one successful internal deployment into a universal income claim.
Sources and research note
Converting inbound leads into customers at OpenAI. Demystifying evals for AI agents. Research checked October 26, 2026. Product behavior and prices can change. Business examples and proposed prices in this article are illustrative unless explicitly attributed.
Leave a Reply