What is agentic automation? It is automation in which an AI model works out how to reach an outcome, instead of running a fixed list of steps, inside limits the business sets. Salesforce’s architecture guidance describes it as automation that takes user intent or unstructured data as input and applies reasoning to reach a desired outcome. What decides whether it survives production is how much autonomy the system has earned.
Key takeaways
- How agentic automation differs from rule-based automation and from plain workflows.
- Why many agentic projects stall, and how permissions are widened in stages.
- Who builds and hands over these systems, and how that work is measured.
What agentic automation means in practice
The practical line sits between a path written in advance and a path the model chooses. According to Anthropic, workflows offer predictability and consistency for well-defined tasks, while agents are the better option when flexibility and model-driven decision-making are needed at scale.
Anthropic also offers a caution worth repeating. For many applications, optimising a single model call with retrieval and in-context examples is usually enough. Agentic automation is therefore not a label for every AI feature. A classifier that routes a refund request into the right queue is a workflow. A system that reads the customer’s history, chooses which tool to call and issues the refund is closer to an agent. Anthropic’s own customer support example has tools pulling customer data and order history, with actions such as refunds handled programmatically.
Most real systems mix both. Salesforce’s guidance says architects must decide when to use probabilistic reasoning on unstructured data and when to enforce guided, deterministic workflows to limit unpredictable outcomes. That decision, made case by case, is the core design skill.
Where agentic automation came from
The lineage runs from rules to workflows to agents. According to Automation Anywhere, a vendor of robotic process automation, traditional tools follow fixed rules and linear paths. That works well for stable, structured tasks and badly for anything with variation.
The next step put language models inside code-defined paths. Anthropic’s catalogue of workflow patterns includes routing, which sends different query types to different downstream processes, and parallelisation, where several model calls run at once and are aggregated programmatically. The code still owns the control flow.
Agents hand that control to the model. Anthropic notes that at each step the agent needs ground truth from the environment, such as tool call results or code execution, to assess its progress. Automation Anywhere says this suits work that is dynamic, long-running or judgment-heavy, the kinds of process that have historically been hard to automate. The shift is real, but it moves risk from the script to the system around the model.
Why so many agentic automation projects stall
The evidence for stalling is consistent. According to IHL Group, Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. The same article reports that only 17% of organisations have deployed AI agents so far, while more than 60% expect to within two years.
Part of the confusion is vocabulary. MiHCM reports that Gartner also flagged agent washing, the rebranding of existing assistants, RPA and chatbots as agents, and estimates only about 130 of the thousands of agentic AI vendors are real.
The failures tend to be unglamorous. IHL’s retail example is a good one. An agentic replenishment system running on inventory data that is only 60 to 70 per cent accurate will place autonomous orders at the wrong quantities, times and locations. The model is not at fault there. The agent was given write access to a process whose inputs nobody had verified, and the cost of correcting machine-speed errors arrived before the savings did.
How autonomy is earned in production
Treat autonomy like access rights: granted in stages, on evidence. Gartner’s advice, as reported by YourStory, points the same way. It advises a phased approach that includes clearly defined use cases, realistic goals and strong monitoring. One working rule for the phases is graduated autonomy.
A production agent earns write permissions after read-only trust is established.
In practice the first phase is read-only. The agent proposes actions, people compare them with what they would have done, and the gaps become the test set. Only then does the agent get a narrow write permission, with approval, then without it for the low-risk cases. Anthropic’s guidance supports the mechanics: agents can pause for human feedback at checkpoints or blockers, and stopping conditions such as a maximum number of iterations maintain control. Anthropic also recommends transparency, meaning the agent’s planning steps are shown explicitly.
Permissions answer what an agent may do. Accountability answers who stands behind it. The Signature Test puts that question to any deployment:
Who signs the risk statement before go-live?
If nobody will sign, the system has not earned write access, however well it demos. That is a useful finding, and it is far cheaper to reach before go-live than after.
Who builds agentic automation and hands it over
These systems are rarely built from a distance. They need someone inside the client’s environment who understands the data, the permissions and the process. A press release from UNKAPPED, published on WebDisclosure, describes forward-deployed engineering as engineers embedded in client teams who connect data, fix access and permissions, wire agents into existing systems and take them to production. That is the Forward Deployed Engineer role, and AIU describes the Certified Forward Deployed Engineer (CFDE)™ as the first certification standard for it.
Agentic work follows a lifecycle, and the stages are not equally glamorous.
AIU teaches the deployment lifecycle in seven stages: post-sale handoff, discovery, scoping, build, deploy, iterate and hand-back.
Hand-back is the stage most often skipped. An agent the client team cannot inspect, pause or adjust has not been handed over. It has been left behind. For agentic systems that means the client’s own people can read the logs, understand why a refusal happened and change a permission without calling the original team.
Judging that work needs more than a demo, so the measures matter.
AIU teaches that Forward Deployed Engineers are measured on three things: deployment velocity, customer success and product feedback quality.
Speed alone rewards shortcuts, which is why it sits beside customer success and the quality of feedback sent back to the product. Together they favour engineers who ship, stay accountable and report honestly what the system could not do.
If this is the work you want to do, the next step is to see how CFDE is structured.
The CFDE credential and who issues it →
Frequently asked questions
What is agentic automation?+
Agentic automation is automation in which an AI model decides how to complete the steps of a business process, calling tools and adapting to what it finds, instead of following a fixed script. The business still sets the limits: what the agent may read, what it may change, when a person must approve, and who is accountable for the result.
How is agentic automation different from a workflow?+
In a workflow, code defines the path and the model fills in individual steps. In agentic automation, the model chooses the next step and the next tool within boundaries. Workflows are more predictable and easier to audit, so many production systems combine the two, using an agent only where the path cannot be written in advance.
How much autonomy should an AI agent have at first?+
Start with read-only access. Let the agent propose actions while people compare them with their own decisions, then grant narrow write permissions with human approval, and widen them only as evidence builds. Autonomy should be earned in stages, and a named person should be accountable for the risk before the agent goes live.
