All articles
5 min read

OpenAI Jalapeño: Is the ROI of Automation Changing for Small and Mid-S

OpenAI and Broadcom have announced the Jalapeño chip for faster and cheaper AI model operations. Does this change the profitability of automation for small businesses? Check the ROI thresholds and a simple calculator for three processes.

Cover illustration for article: OpenAI Jalapeño: Is the ROI of Automation Changing for Small and Mid-S

Key takeaways

  • Focus on the cost per task (per ticket, call, document), not just the 'token price'.
  • A small decrease in LLM costs (-10% to -30%) can unlock ROI in text-based support and data entry.
  • Voicebots typically require a high 'deflection rate' (the percentage of issues resolved without human help) and low fixed costs, not just a cheaper LLM.
  • Calculate the LLM cost threshold: savings per task minus fixed costs per task.
  • Start with a narrow pilot and measure: volume, minutes saved, LLM cost, fixed costs.

OpenAI and Broadcom have announced the Jalapeño chip optimized for 'inference' (the process of generating responses) of AI models. What does this mean for the budget of a small business? Let’s explore simple ROI thresholds and 'what if' scenarios for three everyday processes.

What Was Announced and Its Impact on LLM Costs

OpenAI and Broadcom introduced Jalapeño—a chip designed to run AI models faster and cheaper. Inference is the moment when the model actually calculates the answer to your question. Lower costs here could mean potentially lower prices or higher limits from providers.

LLM (Large Language Model) is a program that 'reads' and 'writes' like a human. Importantly, you pay for usage. Therefore, in business, the cost per task matters: one ticket, one call, one document.

ROI (Return on Investment) can be simplified to: savings from human labor + additional revenue – LLM costs – fixed tool costs. The takeaway: you don’t need to know the chip specifications. Just calculate your cost per task and the profitability threshold.

Calculator: Customer Support (Chat/Email)

Assumption: AI suggests a draft response, and a human approves it. We calculate the minutes saved per ticket.

How to calculate: LLM cost threshold per ticket = (minutes saved × hourly rate ÷ 60) – (fixed monthly cost ÷ number of tickets). If the actual LLM cost is lower than this threshold, then there’s ROI.

  • Example: 2,000 tickets/month, rate of $10/hour, savings of 2 minutes/ticket, fixed cost $375/month.
  • Savings per ticket: 2/60×10 = $0.33. Fixed cost per ticket: 375/2,000 = $0.19. LLM cost threshold: 0.33 – 0.19 = $0.14.
  • 'What if' scenarios (current LLM cost = $0.15/ticket):
  • -10%: $0.14 (on the edge). -30%: $0.10 (profit ~ $0.04/ticket ≈ $80/month). -50%: $0.08 (profit ~ $0.25/ticket ≈ $500/month).
  • Conclusion: In text support, a -30% reduction can tip the scales. Start with the simplest categories of issues.

Calculator: Voicebot for Call Centers

A voicebot is an AI agent (agent = a program that independently carries out task steps) that operates via voice. It requires speech recognition (ASR—converting speech to text), LLM (reasoning), and text-to-speech synthesis (TTS—converting text to voice). Here, the cost per call can be higher than in text.

A key metric is the deflection rate—the percentage of calls resolved without human intervention. We calculate: savings per call = deflection × human labor cost per call. Subtract the voicebot cost (fixed cost per call + variable audio + LLM costs).

  • Example: 3,000 calls/month, 3 minutes/call, rate $12/hour (human cost ≈ $1.00/call), deflection 40%, fixed cost $625/month, variable bot cost $0.20/call.
  • Savings: 40% × $1.00 = $0.40/call. Fixed cost per call: 625/3,000 = $0.21. Total bot cost: 0.21 + 0.20 = $0.41. Balance: -$0.01 (in the red).
  • 'What if' scenarios: -30% variable cost → $0.14 (balance -$0.03). -50% → $0.10 (balance +$0.01). Still in the red.
  • When does it make sense? For example, with a 60% deflection, savings would be $0.60. With a -30% variable cost, total cost ≈ $0.61, which is a slight profit. Or if you reduce fixed costs to ~ $250/month (using a ready-to
  • bullets: []},{

Calculator: Data Entry (Invoices, CRM)

AI reads a document and pre-fills fields; a human just checks. This is usually a stable process with predictable costs per document.

LLM cost threshold per document = (minutes saved × hourly rate ÷ 60) – (fixed monthly cost ÷ volume).

  • Example: 5,000 documents/month, rate $9/hour, savings 2 minutes/doc, fixed cost $375/month, current LLM cost $0.20/doc.
  • Savings per doc: 2/60×9 = $0.30. Fixed cost per doc: 375/5,000 = $0.08. Balance: 0.30 – (0.20 + 0.08) = $0.02 (in the green).
  • 'What if' scenarios: -10% → $0.18 (profit ~ $0.12/doc ≈ $600/month). -30% → $0.14 (profit ~ $0.16/doc ≈ $800/month). -50% → $0.10 (profit ~ $0.20/doc ≈ $1,000/month).
  • Conclusion: Here, even a -10% reduction can switch the outcome to positive.

Market signals (Jalapeño) suggest decreasing costs for running models. The safest approach: calculate your thresholds today and prepare narrow pilots. Want to run the numbers on your data in 30 minutes? Reach out—we'll walk through the calculator and identify the first, cheapest step.

Frequently asked questions

Will OpenAI Jalapeño immediately lower my bills?

Not automatically. Jalapeño is designed to speed up and reduce the cost of running models, but it's the providers who set pricing and limits. So, it’s wise to calculate your thresholds now to be ready when prices actually drop.

How do I calculate the LLM cost per task?

Take the usage cost from the tool or API (a connector between programs) and divide it by the number of tasks. If you don’t have data, run a small test on 100 tasks and calculate the average cost.

What is an AI agent and where does it make sense?

An AI agent (agent) is a program that independently executes the steps of a task, such as determining the reason for contact, checking status in the system, and responding. It works well in repetitive processes with clear rules.

Should I wait for lower prices or start a pilot?

Start with a narrow pilot. Determine volume, minutes saved per task, LLM cost, and fixed costs. When prices drop, you’ll know immediately if scaling is feasible. Without this data, decisions will be based on guesswork rather than numbers.

How can I limit the cost risk of a voicebot pilot?

Start with 1-2 topics that have a high deflection rate, limit operating hours, and use a ready-made solution to reduce fixed costs. Measure the cost per call and the percentage of issues resolved without human help.

Let's talk
about your project

The consultation is free and no-strings-attached. We'll review your needs and I'll suggest concrete solutions.

Send a message

Briefly describe your problem — I'll get back to you with concrete suggestions.