The center of gravity in enterprise AI has shifted from chat to agents: systems that take multi-step actions, call tools, and connect to the software teams already use. A chat window answers a question and stops there, waiting for a person to act on it. An agent can look something up, decide what to do next, and carry it out: filing a ticket, updating a record, drafting and routing a follow-up. That is a meaningfully bigger promise, and a meaningfully bigger risk: a chat answer that is wrong sits on a screen where someone can catch it before it matters. An agent that is wrong can act several steps into a workflow before anyone notices.
What an agent really is
Strip away the marketing and an agent is AI that does more than answer: it retrieves information, calls tools and APIs, and completes multi-step tasks toward a goal, ideally with checkpoints where a human stays in control. That is a different engineering problem than a chatbot. A chatbot’s worst failure is a bad sentence. An agent’s worst failure is a bad action, taken on your behalf, inside a system of record. The agents that hold up in production are narrow and reliable, not general and flashy. They are built to do one job well within clearly defined limits, rather than to improvise across an open-ended range of tasks.
Where it pays
- Workflow automation. Connecting AI into existing systems (CRM, ticketing, and ERP), so work flows through end to end, instead of stopping at a chat window where a person has to manually copy the answer somewhere else.
- Multi-step research and drafting. Tasks that span several sources and steps, from pulling data from a few systems to reconciling it and producing a draft, can collapse from an afternoon into minutes.
- Triage and routing. Classifying, summarizing, and routing incoming work (tickets, applications, and requests) to the right queue or person, consistently and immediately.
The architecture that makes agents reliable
Every durable agent deployment we have seen shares the same underlying design choices, regardless of the vendor or model behind it. Access is scoped tightly. An agent that can read a customer record should not automatically be able to modify billing. Actions above a defined risk threshold route to a human checkpoint before they execute, rather than after. The system is evaluated continuously against real examples, not demoed once and left alone, because agent behavior drifts as the underlying model, prompts, or connected systems change. And every action is logged in a form someone can audit later: which tool was called, with what input, and what the outcome was.
The failure mode is autonomy without guardrails. Reliable agents are scoped tightly, evaluated continuously, and keep a human in the loop where the stakes are high. It is the same discipline that separates production software from a demo, applied to a system that can now take action on its own.
Human checkpoints that actually work
A human-in-the-loop design is useless if the review UI is a wall of JSON nobody reads under deadline pressure. Design checkpoints around decisions people already make: approve a draft email, confirm a CRM field update, accept a routed ticket category. Show the evidence the agent used. Make rejection cheap and informative so the evaluation set improves. If reviewers rubber-stamp everything, you do not have a control — you have theater.
Separate read tools from write tools in permissioning. Many valuable agents can do most of their work with retrieval and drafting, then ask a person to commit the write. That pattern captures speed without giving the model unsupervised authority over systems of record.
Why most agent pilots stall
Agents raise the bar on evaluation and governance because they act, not just answer, and that is exactly where most pilots run aground. MIT NANDA’s 2025 State of AI in Business research found that while a large share of organizations evaluate enterprise AI tools, only a small minority reach production with measurable P&L impact. Agent projects are more exposed to this than simple chat tools, because the cost of an ungoverned failure is higher and more visible. Unclear ownership, missing success metrics, and governance bolted on after the demo are still the usual stall points.
BCG’s 10-20-70 pattern is especially relevant here: tool calling and orchestration are not most of the work. Redesigning the workflow, training the people who supervise the agent, and defining who is accountable when an action is wrong. That is where programs succeed or quietly die.
Adopt deliberately
Start narrow, prove reliability on one workflow, then expand, and choose a platform whose agent capabilities actually fit your roadmap (see choosing the right vendor). Adoption still decides the return more than the technology does; see pilot to production. The organizations getting real value from agents today are not the ones that moved fastest; they are the ones that scoped tightly, measured honestly, and expanded only once a workflow had proven itself.
Start with an AI Decision Sprint.
Production readiness checklist for agents
Before an agent touches a live system of record, confirm each of these in writing:
- Named owner. A business owner who can kill or pause the agent, not just a technical contact.
- Scoped tools. The minimum APIs and permissions required, with read vs write called out separately.
- Risk thresholds. Which actions auto-execute, which require approval, and who approves.
- Evaluation set. Real historical cases the agent is measured against before launch and after every change.
- Failure mode. What the agent does when a tool call fails, a lookup is empty, or confidence is low: halt, escalate, or ask; never invent.
- Audit log. Every tool call, input summary, and decision recorded so a bad run can be reconstructed.
- Rollback. A way to disable the agent in minutes without taking down the surrounding systems.
McKinsey’s 2025 State of AI survey found that AI high performers are nearly three times as likely as others to say they have fundamentally redesigned individual workflows, and more likely to have defined processes for human validation of model outputs. Agents amplify that finding: the ones that create value are the ones embedded in a redesigned process with checkpoints, not the ones bolted onto an unchanged queue.
MIT’s 2025 NANDA research put measurable P&L impact from integrated GenAI initiatives at around 5%. Agents that skip evaluation and permissions tend to join the other 95%, not because the model is weak, but because the operating model never got designed.
Sources
- MIT Project NANDA, “The GenAI Divide: State of AI in Business 2025”
- McKinsey, “The State of AI: Global Survey 2025”
- BCG, “The Leader’s Guide to Transforming with AI”
- NIST, “AI Risk Management Framework: Generative AI Profile”
Accessed July 2026. Vendor terms and benchmark methodologies change; verify current primary documentation before making a decision.
Turn the pattern into a production system
Build covers architecture, integration, evaluation, delivery, and handoff against written acceptance criteria.
Explore Build →