Leave the IT Tasks to ServiceHub: Governed, Agentic Automation

The promise is simple: take the repetitive back-office work that eats your team's hours and let agents do it, so people spend their time on the business problem instead. What makes the promise safe to keep is governance — agents do the work, but inside guardrails you configure, with a full audit trail behind every decision.

The question is never "can an agent do this?" It is "what happens when it gets something wrong?" Everything below is an answer to the second question.

Three ways to author a workflow

Start wherever suits the person doing the work. None of these is the lesser path:

• Draw it. The Visual Workflow Builder lays out steps, approvals and branches on a canvas. A readiness check verifies the design is genuinely runnable before you can submit it, so an incomplete workflow is caught at authoring time rather than halfway through its first run.
• Describe it. State the outcome in plain language and the platform drafts the steps, forms and tool selections — grounded in your own workflows and tools, not a generic template.
• Start from a template. Begin from a workflow that already works and adapt it. Templates carry their intake form with them, so the questions come too.

Two independent dials

People conflate these, and conflating them is why "agentic" makes risk teams nervous. They are separate.

How a step runs — deterministic or agentic. A deterministic step does the same thing every time. An agentic step reasons about the input it was given.

How closely a human watches it — three levels:

• Autonomous — the agent acts and completes the step on its own.
• Confidence-based — the agent proceeds only when its confidence clears the threshold you set. Below it, the step goes to a person. A technical failure always goes to a person, whatever the confidence.
• Always-review — a human reviews before the action is taken, every time.

These combine freely. A step can be agentic and still require approval; deterministic and fully autonomous. Both dials are configuration, not code — including the thresholds and the retry limits.

On top of the mode, a workflow can declare approval checkpoints: deterministic human gates that pause for an explicit approve or reject regardless of what the model decided. The agent cannot talk its way past them, because it is not consulted.

Three things that are true of every run

• Tool-truth. A step records the tools it actually invoked and what they returned. If it cannot prove the required tool ran, it fails to human review. It never reports a silent false "done". This exists because the most dangerous failure in agentic software is not an error — it is a confident summary of work that never happened.
• Preview before send. Anything with a side effect outside the platform composes a draft and holds. Nothing leaves until someone approves it.
• A full audit trail. Reasoning, tool calls, approvals and the deliverables produced are recorded against the task, so the evidence sits with the approval instead of in a log nobody opens.

When the agent gets it wrong, you correct it in words

A reviewer who disagrees with an agent's output does not have to accept it or start over. They comment on the task in plain language — "you have mapped the tax column to the wrong header" — and re-run the step. The agent re-runs with that correction as context, and the loop repeats until it is right or the reviewer approves it directly. Every iteration stays visible on the task.

That turns review from a gate into a conversation, and it is available to every agentic workflow rather than being special-cased into one.

On demand, or on a schedule

A workflow runs when you submit it, or on a schedule you set, and one workflow can carry more than one schedule. Scheduled runs are ordinary runs: same governance, same audit trail, same pauses. A scheduled step that needs a human still waits for one — the schedule does not buy the agent extra authority.

Reading what your systems already produce

Agents draw on 47 built-in tools and read the documents you already have — Excel, CSV, HTML, PDF and scanned images, including photographed documents, with no manual re-key.

Where a document format is genuinely new, ServiceHub can be taught it through a guided request rather than a development project. You supply a sample, describe it in your own words, and give a known-correct example — a closing balance, an entry count — that the learning has to reproduce before it is allowed to go live. If it cannot reproduce your example, it does not go live. That gate is the whole point: without it, "the system learned your format" is a claim nobody can check.

Metered, so it stays economical

Agentic work consumes real resources, so it is metered — and the balance is checked before a run starts, not after. A runaway process is prevented rather than invoiced. How that works is its own post.

What teams get back

Leaving the IT tasks to ServiceHub is not about removing people from the loop. It is about putting them in charge of the loop, and letting agents do the repetitive work inside it.