What makes a system agentic rather than generative?
A system is agentic when it plans, takes actions, observes the results, and adjusts — a loop rather than a single response. A generative system produces an output when prompted. An agentic system decides which steps to take toward a goal, uses tools to take them, and reacts to what happens.
For a business the distinction is operational rather than academic. A generative system needs a person to decide what to ask for, judge the output, and carry it to the next step. An agentic system carries the work between steps itself, which is what makes function ownership possible and what makes guardrails necessary.
Which business functions can agentic AI own today?
Agentic AI can own functions that are continuous, evidence-driven, and anchored in a system of record. Those three properties matter because the loop needs somewhere to look for truth and something to be measured against.
Search visibility fits: the evidence is search data and published pages, the system of record is the site and the search console. Media buying fits: the evidence is spend and conversions inside ad platforms. Reconciliation fits: the ledger is the system of record. Support triage fits: the help desk is. Outbound fits: the CRM is. What these have in common is not that they are easy, but that they are checkable.
- Search visibility and content production, measured in rankings and citations
- Paid media buying, measured in spend against a target return inside hard caps
- Reconciliation and exception handling, measured against the ledger
- Support triage and knowledge maintenance, measured in the help desk
- Outbound sequencing and CRM hygiene, measured in the CRM
Why are approval gates the enabling control, not the brake?
Approval gates are what let a business give an agentic system real scope, which makes them an enabler rather than a constraint. Without a gate, the only safe scope is a trivial one, because any mistake reaches a customer, a ledger, or a budget directly.
With a gate, the system can be given a genuinely broad remit, because the blast radius of a wrong proposal is a rejected proposal. On ftes.ai this is the default state: a Digital FTE starts at the level where it proposes everything, and a manager promotes it for specific classes of action once the pattern has proved itself. The autonomy is earned against evidence rather than assumed at purchase.
What does agentic AI still fail at?
Agentic AI fails where the remit is ambiguous, where the evidence is missing, and where the decision belongs to a person. An undefined process does not become defined because software is pointed at it; the system inherits the ambiguity and produces confident work in the wrong direction.
It also fails quietly when nobody checks. A loop that runs without review will optimise for whatever proxy it was given, which is why a measured target and a visible audit trail matter more than model choice. The most common cause of a disappointing deployment is not the model; it is a vague goal, no baseline, and no one reviewing the output.
How should a business start with agentic AI?
Start with one function, one property, and one target that can be measured against a baseline you record before you begin. A property here means a specific website, store, ad account, or ledger rather than the whole estate, so that results are attributable and mistakes are contained.
Set the guardrails before the first run: what it may never do, what must escalate, and what the spend ceiling is. Then review the output on a fixed rhythm rather than watching it continuously. The rhythm is the point — an agentic system you check on twice a day is a direct report; one you watch all day is a tool you are operating.
How do you know it is working?
You know an agentic deployment is working when the manager's time shifts from producing to deciding, and when the measured target moves against the baseline you recorded. Those two signals together separate real progress from activity.
Watch the approval queue as a leading indicator. If proposals are increasingly approved unchanged, the system has learned the standard and is ready for more autonomy in that class of action. If they are repeatedly sent back for the same reason, the remit or the guardrails need work — which is a management problem with a management fix, not a model problem.
What does a first agentic deployment cost a business in time and money?
The first agentic deployment costs more human time than money. On ftes.ai the money is a plan block of metered tokens shared by every Digital FTE, with a hard stop when it is exhausted, so the financial exposure is bounded by a purchase the organization chooses. The time is the manager's: an hour or two to review the Brand Brain and write the remit in the first-run wizard, minutes per proposal in the first weeks while every action is gated, and a weekly read of the send-back reasons to tune the remit.
That time falls quickly as the approval record accumulates and routine classes of action earn autonomy, and it is the same time a manager would spend onboarding a careful new hire. The honest way to plan for it is to record a baseline of what the function costs today before the pilot, run two full cycles, and compare outcomes in Targets, tokens per role, and human hours against that baseline. Agentic AI pays back when the repetitive work it absorbs was real; the record is what shows whether it was.