Agents That Do the Work, Every Day, in Production

The main difference is how they fail. An RPA script follows a recorded path and breaks when the path changes: a moved button, a renamed field, a new dialog. An agent reads the screen or the data, decides what to do, and can handle the case nobody recorded. A practical test is how much of your automation team’s time goes on maintaining scripts instead of building new ones. Where RPA is working, leave it. We’re usually replacing the small number of scripts that take a disproportionate share of that time, and interface change is only one reason a script ends up in that group, alongside data-format changes, credential rotation and environment failures.

It depends on the workflow, and each route has a point where it stops. Your own engineers on a framework like LangGraph work well for one workflow in one system, but get harder at the second workflow, the security review and the day a provider retires a model. The RPA platform you already own suits high-volume, deterministic work on interfaces that don’t change, and it’s cheaper per run and easier to audit than an agent. Your software vendors’ own agents are the right choice for work that lives entirely inside one of their products. An agent engineering partner makes sense for workflows that cross several systems, take actions that are hard to unpick and must be explainable afterwards. BigStep’s one-week Agent Opportunity Blueprint checks which route fits your workflow, and it will tell you when the answer is one of the first three.

It depends on what the agent has to do. The main cost drivers are how many systems it connects to, how many actions it can take on its own, how strict the approval and audit requirements are, and whether it has to run for hours across hundreds of steps. BigStep publishes its prices by stage: an Agent Opportunity Blueprint to pick the workflow (1 week, $8K to $15K), an Agent Prototype Sprint (2 to 3 weeks, $25K to $45K), and an Agent Scale Launch for a production system (6 to 12 weeks, $80K to $200K). Keeping it running afterwards through AgentOps managed services starts from $5,000 a month.

Every action an agent can take is tiered by whether it can be reversed. The ones that can’t, like a payment, a regulatory filing, a clinical note or a message to your customer, wait for a named human approver. Confidence thresholds can escalate more cases on top of that, but the approval gate is the control, and it’s enforced in the tool layer instead of only being asked for in the prompt, so it holds even when the model behaves unexpectedly. BigStep sorts every action by whether it can be undone and whether it leaves a record someone will read later, and you decide where those lines sit.

Prompt injection is when text an agent reads, such as a document, an email or a supplier’s portal, contains instructions written to steer it. It’s the main security problem with agents, and it’s different from the one your application security team already handles: an agent with tool access is an authenticated actor that text nobody at your company wrote can steer. The defences are structural. Retrieved content is kept separate from instructions, the agent’s actions are limited to what the workflow needs, credentials are scoped so an injected instruction can’t reach data the agent has no business touching, outbound traffic is limited to an allow-list, and anything that triggers an action is validated, not trusted. BigStep tests this before go-live and again whenever the model changes.

Yes, and systems with no API are the normal case in enterprise operations. A computer-use agent is an AI agent that operates a browser or a desktop application the way a person does, reading the interface and reasoning about it instead of following recorded coordinates. That reaches payer portals, carrier platforms, government filing sites and thick-client legacy applications that integrations can’t touch. It runs in a sandboxed environment, checks visually that the action it took is the action that happened, and records every run so it can be replayed step by step when someone needs to check it. Where a system is already automated reliably, such as a mature RPA estate against an interface that doesn’t change, leave it alone. This is for the systems nobody has managed to automate, or where the automation keeps breaking, and it’s what BigStep’s computer-use agents are built for.

More often than most people expect. If the work is high volume and genuinely deterministic, a script is cheaper, faster and easier to audit. If it happens four times a month, the return won’t cover the engineering. If nobody can describe how the work is done today, the problem is the process and not the automation, and an agent will encode the confusion. The Agent Opportunity Blueprint exists to find these out in a week for $8K to $15K, and telling you not to build something is a valid outcome of it.

An agentic workflow is a business process in which an AI agent, a system that decides what to do next instead of following a fixed script, carries out several steps on its own. It uses tools to read and write in real systems, and it stops for a human where the rules say it must. AgentOps is what keeps one working after launch: tracing every run, clustering failures instead of counting them, attributing cost per workflow and re-testing against a recorded baseline whenever a provider changes a model. We build the first and run the second, and the second is the part most agent projects find they needed about six months too late.