Key takeaways
- Start from a costly decision or task, not from what the model can do.
- Design for the unhappy path: narrow scope, clean handoffs, auditable actions.
- Grounding the agent in your real data and systems is where the ROI lives.
- Measure resolution rate, time saved and cost per task, not how human it sounds.
- The model is the easy part; workflow, data and integrations create the value.
The demo trap
Every week we see another AI agent that wows a room and then quietly disappears. It summarises, it chats, it looks like the future. Then someone asks the uncomfortable question: what did it actually change in the business? Usually the answer is nothing. The demo was the product, and the product was theatre.
We've learned, building agents for e-commerce, services and operations teams, that shipping value is a completely different discipline from impressing a stakeholder for ninety seconds. A demo optimises for the best case. Production is defined by the worst case, repeated thousands of times, by people who didn't attend the kickoff.
Start from a decision, not a capability
Most AI projects begin with "what can the model do?" We start with "what decision or task is costing you money, time, or trust?" An agent earns its place only when it removes real friction from a specific workflow.
Concretely, before writing a line of orchestration we insist on three answers:
- The job to be done. One sentence, in the client's own words, describing the outcome, not the technology.
- The baseline. How is this handled today, how long does it take, and what does an error cost?
- The definition of good. A measurable bar the agent must clear to be worth deploying.
If we can't fill in these three, the project isn't ready, and no amount of clever prompting will save it.
Design for the unhappy path
The difference between a toy and a tool is how it behaves when things go wrong. Real users send ambiguous requests, incomplete data, and edge cases nobody imagined. A production agent needs guardrails that assume failure is normal.
In practice, this means:
- Bounded scope. A narrow agent that does one thing reliably beats a general one that does ten things unpredictably.
- Graceful handoff. The agent should know what it doesn't know and escalate to a human cleanly, rather than hallucinating confidently.
- Verifiable actions. When an agent writes to a database, sends an email, or issues a refund, every action is logged, reversible, and auditable.
We treat the language model as one component, not the whole system. Around it sit validation, retrieval from trusted sources, business rules, and the integrations that let the agent actually do things instead of merely talking about them.
Ground it in the client's reality
An agent that doesn't know your products, your policies, or your tone is a stranger answering your phone. Value comes from grounding: connecting the model to your real catalogue, your CRM, your knowledge base, your constraints. That plumbing is unglamorous and it is exactly where the ROI lives.
This is also where trust is won or lost. We'd rather an agent say "I'll check and come back to you" than invent a delivery date. Over time, a slightly more cautious agent that is right builds far more adoption than a confident one that is occasionally wrong.
Measure, then let it earn more responsibility
We ship agents on a leash and lengthen it as they prove themselves. Early on, a human reviews outputs. As accuracy holds against the "definition of good", we expand autonomy where the data supports it. This staged trust model protects the business and gives teams confidence instead of fear.
The metrics that matter are rarely the ones demos celebrate. Not "how human does it sound" but resolution rate, time saved, error rate, and cost per task. When those numbers move, the agent has earned its keep. When they don't, we kill it fast and redirect the budget. Honesty here is cheaper than hope.
Why most of the work isn't the model
The uncomfortable truth is that the model is now the easy part. The hard, valuable work is understanding the workflow, cleaning the data, defining the rules, building the integrations, and designing for the humans who will live with the result. That's where we spend our time, because that's where business value is manufactured. A great model with poor context and no integration is a very expensive chatbot.
AI agents are genuinely transformative, but only when they're engineered like infrastructure rather than showcased like magic tricks. The companies winning with AI right now aren't the ones with the flashiest demos. They're the ones who quietly automated something that used to hurt.
If you have a workflow that's costing you more than it should, we'd be glad to look at it with you and tell you honestly whether an agent is the right answer. Let's talk about your project.
