How to redesign workflows to reduce AI costs effectively

TL;DR

Redesigning workflows is essential before buying more AI agents, because better interfaces and integration can make adoption practical without inflating costs. The key is to map value delivery, replace costly bottlenecks, and ensure agents support core business outcomes rather than merely automate steps. A thoughtful harness design matters for sustainable cost control.

Transcript

agents keep getting better and better and better, but they can still make your bill go up. At Dreamforce, I got to ask Mark Beni off a question about how we get more people to adopt AI. And later that day, I was on a panel talking about the cost that we're all facing when we roll out agents of scale. By the end, you're going to know how to look at ... Read More

Key Insights

• AI costs rise when more capable agents perform more per request, so the real test is whether the work being done by agents creates new value that could not be achieved otherwise.
• Adoption hinges on putting agents into places where people already work, such as Slack, to lower the barrier to try AI and gradually scale usage across teams.
• A thick harness combined with a thin harness helps balance capability and cost by controlling how much data and tooling an agent uses for each task.
• Evaluations and classifiers determine when a cheaper model is worth using, preventing overreliance on expensive fronts while preserving value.
• Designing workflows from a clean sheet and focusing on value streams keeps costs in check by avoiding unnecessary handoffs and redundant steps.
• Payments for AI should be tied to a measurable business outcome, not the number of model calls, to ensure sustainable ROI.
• Agents can read, reformat, and hand off requests automatically, but the envelope of the process still exists; the goal is to streamline it without changing the ultimate outcome.
• A broad adoption path requires visible reliability and trust in agents, so businesses monitor performance and reduce friction to scale across the organization.

Questions & Answers

Q: How to start reducing AI costs without sacrificing outcomes

To start reducing AI costs without sacrificing outcomes, identify the few value-driving paths in your business where AI can deliver measurable improvement. Begin by mapping the complete workflow that leads to a customer outcome, then determine which steps must exist and which ones can be streamlined or removed. Focus on eliminating redundant handoffs and reducing manual reentry. Implement a design where agents handle only high-value decisions and routine tasks are automated in a controlled, auditable way. Continuously monitor performance, costs, and the business impact to ensure that each AI action contributes to the desired result. This approach prevents overspending on token-wasteful paths and keeps adoption aligned with real value.

Q: What is meant by a thick harness and a thin harness in AI adoption

A thick harness refers to a robust AI system that has access to broader data, tools, and decision-making capabilities, enabling complex tasks but potentially increasing cost and risk. A thin harness, by contrast, uses limited data and fewer tools, focusing on simple, well-defined tasks to minimize cost and risk. The key is to balance these harnesses: start with a thin harness for core functions to validate value and safety, then add a thick harness selectively for higher-impact workflows where the business case justifies the investment. This staged approach helps control costs while expanding capability.

Q: How do interfaces influence AI adoption in organizations

Interfaces influence AI adoption by meeting people where they work, reducing friction to try AI, and increasing perceived usefulness. If agents are embedded in familiar tools like Slack, users can experiment with AI in real time without a steep learning curve. This proximity to daily tasks builds trust and demonstrates incremental value, making it more likely that teams will expand usage over time. The result is broader adoption across departments without a sudden, expensive rollout.

Q: Why is it important to redesign a workflow before buying an agent

Redesigning a workflow before buying an agent helps you target the exact steps where AI can add value and avoid automating nonessential tasks. This upfront analysis clarifies the business goals, identifies bottlenecks, and determines where an agent can reduce cycle time or improve accuracy. It also prevents paying for capabilities you do not need and helps craft an evaluation plan to measure impact after deployment, ensuring the investment aligns with strategic outcomes.

Q: How can cost be justified when AI agents are more capable

Cost can be justified when AI agents enable work that could not be done before or significantly improve speed and accuracy for high-value tasks. The justification rests on a clear link between the agent’s work and meaningful business outcomes, such as faster customer response, better decision quality, or increased throughput. It requires a principled approach to harness design, selecting tools and data carefully, and monitoring usage to avoid waste. When these conditions are met, higher costs can be defensible as ROI improves.

Q: What role do classifiers and evaluations play in model selection

classifiers and evaluations play a key role by determining when a cheaper model is sufficient and when a more capable model is necessary. They provide a governance layer to steer cost and performance by assessing task complexity, risk, and required accuracy. This helps prevent overspending on expensive models for simple tasks while ensuring that critical decisions receive appropriate support. Regularly updating criteria as workflows evolve keeps the system aligned with business needs.

Q: How do you measure whether AI adoption is moving the business value

Measuring whether AI adoption moves business value involves tracking the end-to-end impact on the value streams that matter most, such as customer acquisition, fulfillment, retention, and revenue. Establish clear KPIs for each workflow, monitor throughput, error rates, and time to value, and compare outcomes before and after AI deployment. Use these measurements to refine the harness, adjust the task mix, and decide where to scale. A data-driven approach ensures the AI program stays focused on value rather than activity.

Q: What is the risk of paying six agents to summarize each other

The risk of paying six agents to summarize each other is inefficiency and token waste that does not add new value. When multiple agents perform repetitive, redundant tasks, costs can surge without corresponding benefits. To mitigate this, design workflows so that agents have clear responsibilities, implement handoffs that minimize duplication, and use evaluation gates to stop unnecessary cascades of summarization. Regular audits help ensure the system remains lean and productive.

Summary & Key Takeaways

• A business should start with the end value and map the steps required to reach it, then decide where to deploy agents to maximize impact without unnecessary expense.

• Interfaces that fit into existing work environments, like Slack, raise adoption and reduce resistance, enabling broader use without huge upfront costs.

• The cost discussion shifts from choosing cheaper models to redesigning workflows and harnesses so that agents perform only valuable, high-impact tasks.