Building AI Agents that Actually Work: A Deliberate Framework for Success

Something unexpected is happening in enterprise AI right now: many organizations are pulling back. After months, or sometimes years, of investment in AI, companies are rolling back subscriptions and pulling the plug on projects that never delivered what was promised.

We're seeing this across industries, and the reasons are consistent. Adoption lags behind expectations, token consumption blows past projections, success hinges on subjective opinions, and essential iteration is forfeited in favor of frictionless demos and expedited implementation. 

The instinct is to blame the technology. But in most cases, the technology isn't the problem. It’s the sequencing: moving through the wrong stages in the wrong order, or skipping important steps entirely in the rush to show results. The truth is, the path from pilot to production often is not a straight line.

Skipping ahead may come back to bite you

There's a gap between what AI could do at full capacity and what it's actually ready to do right out of the gate after an initial investment, and that gap is where most implementations break down.

When organizations jump straight to autonomous agents that act, respond, and make decisions without human review, they’re skipping the foundational work that makes those actions trustworthy. Prompts aren't engineered for the specific workflow. Edge cases haven't been accounted for. Users haven't had a chance to build confidence in what the AI is doing. And when something goes wrong (and something typically does, early on), there's no safety net.

At Thunder, we recently saw this play out with a company managing thousands of support cases per day. Prior to engaging with us, they recognized that a significant portion of those cases could be resolved by pointing customers to the right knowledge article, so they deployed a chat agent with an aggressive mandate: before a customer could open a case, they had to go through the agent first. The agent was developed and implemented quickly, and because of that, the prompts lacked proper engineering and the agent did not behave as expected.

The backlash was immediate. Customers who wanted human help with a problem were instead met with an AI bot trying to troubleshoot them out of creating a case. Customer satisfaction dropped significantly, costs climbed, and internal users disengaged. Not a great situation, but luckily, they reached out to us, and this is a challenge Thunder knows how to correct.

‍A deliberate path forward

AI is all about acceleration, but as the saying goes: sometimes you need to slow down to go fast. The key to making rapid, measurable progress with AI is to take a more deliberate approach.

When Thunder’s AI team works with our customers on Salesforce Agentforce projects, we follow a maturity model: a structured progression that builds AI capability in stages, with each step earning the right to move to the next.

Our model has four stages :

• Diagnose (Crawl): In this stage, AI works entirely in the background. It might tag inputs, categorize cases, or draft potential responses, but it does so without any customer-facing exposure. Leadership and operators observe how it performs and build a fact-based picture of where it's effective before anything goes live.

• Suggest (Walk): AI surfaces recommendations to service representatives, who review and accept or reject them. Users see what the AI is doing, trust gets established, and ROI becomes measurable. This stage is where you prove the model before you extend it.

• Act (Run): With confidence established, AI takes action autonomously, sending responses, routing cases, and following up on the workflows where it has already earned that trust. Not everywhere at once; just where the groundwork is solid.

• Orchestrate (Fly): AI agents across different parts of the business begin working together, calling on each other to complete cross-functional tasks. This is AI at scale, and it’s only sustainable because each agent has proven its effectiveness as it matured over the previous stages.

The timeline looks different for every organization, but a 12–18 month roadmap is typical from initial discovery through full orchestration.

The framework in action

Going back to our customer example above: our client was pushing an undesirable customer experience and we knew they needed to take a step back and reassess. When Thunder came in, the first fix was experiential. Restoring the ability to manually create a case directly through a submission form functioned as a pressure release valve. Once we gave customers the choice, the hostility toward the AI agent dropped significantly.

From there, we rebuilt the agent stage by stage. We started by identifying the top five reasons customers were opening cases and engineering a structured troubleshooting framework for each one. We addressed one case driver at a time, moving deliberately up the maturity model, until the agent was genuinely equipped for the work it was being asked to do.

Over time, as the agent performed better, customers learned to trust it with the tasks and questions we focused on, and the agent was able to deflect more than 50% of the cases that previously would have been opened via the traditional live chat. 

Thunder's Amplify practice: a roadmap, not a shortcut

Deliberate doesn't mean slow. Organizations that move through the maturity model thoughtfully often scale faster than those that rush, because they’re not stopping to untangle problems that shouldn't have happened in the first place.

Doing this well requires a specific mix of expertise:

• Strategic thought leadership to identify and prioritize the right use cases
• Prompt engineering to build agents that actually perform
• Implementation depth to make it all work inside your existing systems

These aren't skills you need for a single project; they're capabilities your organization needs on an ongoing basis as your AI program evolves. For many companies, assembling that team in-house isn't practical or cost-effective.

Thunder's Amplify practice is built to fill that gap. Every engagement starts with a discovery workshop where we map your current workflows, identify where AI can genuinely reduce friction, and prioritize use cases based on ROI potential. From there, we build a 12–18-month roadmap and stay with you as you move through each stage, not just to implement, but to refine, optimize, and scale.