Agentic AI in support: what "agents that take actions" actually do
The pitch is an agent that looks things up, issues the refund, and reschedules the delivery on its own. The reality in most support operations is narrower, more supervised, and more honest. Here is the gap, with the numbers.
The pitch for agentic AI is a verb. Not a bot that answers questions, but an agent that acts: it looks up the order, checks the return window, issues the refund, books the replacement, and updates the shipping address, all in one conversation, with no handoff to a human. It is a genuinely good pitch. It is also, for most support operations in 2026, still mostly a pitch.
Start with the number every deck quotes.
Eighty percent of common issues, resolved autonomously, by 2029. Here is the number that rarely shares the slide.
Both come from Gartner, three months apart. Hold them together and you have the honest shape of the moment: a very large promise and a very large reckoning, running at the same time.
Adoption is real. Autonomy is not.
The confusion starts with the word "use."
Adoption moved fast and it is not fake. But "we use an AI agent" and "we let it take actions unsupervised" are different sentences, and vendors are paid to blur them. Gartner has a name for the blur: agent washing — chatbots, scripted flows, and RPA relabelled as agents.
Roughly 130 genuinely agentic vendors, out of thousands claiming the word. Most of what gets called "agentic" in a sales call is a retrieval bot with a new sticker.
What agents actually do today
Sort real deployments by the kind of action taken, and three tiers appear.
Retrieve and answer. The overwhelming majority. The agent reads your knowledge base and writes a good reply. This is what "AI resolved 70%" actually means: in most reporting, "resolved" means the customer got an answer and did not come back, not that the software did anything. It is autonomous resolution of the question, not the task.
Read actions. Look up an order, check a balance, surface subscription status. Low blast radius, increasingly common, genuinely useful. The agent touches your systems, but only to read from them.
Write actions. Issue the refund, cancel the plan, change the address, move money. This is where the pitch lives, and where careful teams get quiet. Write actions in production sit behind confidence thresholds, dollar caps, and a human in the loop who approves anything consequential. The agent proposes; a person, or a hard rule, disposes.
Write actions sit behind confidence thresholds, dollar caps, and a human who approves anything consequential. The agent proposes; a person, or a hard rule, disposes.
That is not the technology failing. It is the technology being deployed by adults. The gap between "can draft a refund" and "is allowed to issue one unattended" is a governance decision, and most organisations are making it conservatively, on purpose.
The trajectory, read plainly
Stripped of theatre, the near-term forecast is modest and believable.
From zero to fifteen percent of routine decisions in four years is a real shift. It is also a long way from "eighty percent of issues, no human," and the distance between those two numbers is where most of the disappointment in this field is going to come from.
We have already watched the disappointment arrive once, at scale.
Klarna was the poster child for full automation, until the quality and empathy gaps showed up and the company started rehiring humans. The lesson was not "AI doesn't work." It was that AI which answers is not AI that owns the outcome, and the difference is expensive to learn late.
What to do with this
Define agentic AI by what it is allowed to do, not by what it is called. In any evaluation, ask the vendor for the literal list of write actions running in production, the confidence threshold behind each, and the rollback for when the model is wrong. Anything vaguer than a list is a demo, not a deployment — a distinction the vendor scorecard is built to force and the maturity model is built to place. If the honest answer is "it drafts, a human sends," you are looking at assist, not autonomy, which is fine, and worth buying, as long as nobody prices it as the 2029 dream.
The pitch is a verb. Today's ledger is mostly a noun that answers well, plus a short, supervised list of verbs it is trusted to take. The teams that win are not the ones who believe the 80% number. They are the ones who wrote down which actions their agent takes today, measured what actually happened, and added the next verb only once the last one was safe.