A.I. Agents & Skills

What are A.I. Agents and Skills?

An A.I. Agent is an intelligent teammate that works alongside your Staff. It has its own name, its own character, its own way of writing, and its own set of Skills. A Skill is one specific job you hand to A.I.: choosing which Auto-Template runs, or writing a secure message to the person an automation is about.

The split matters. The Agent is who is doing the work and who is accountable for it. The Skill is what gets done. You build Skills once, then assign them to whichever Agents should run them.

IMPORTANT: A Skill can be set up at any time, but nothing actually runs until your account has an active A.I. Service subscription or a verified connection of your own under A.I. Connections. Until then the setup is saved and simply waits.

How do I get started?

There are four one-time steps.

Turn on the A.I. In A.I. Connections, activate the A.I. Service subscription or connect your own provider. Nothing runs until one of those is in place.
Build a Skill. Describe the situations you already recognise, point each one at the Auto-Template it should run, choose a fallback, and publish. Only published Skills can be assigned.
Add an Agent. From your Staff list, use the create button and choose Add Agent rather than Add Staff Member. Give it a name, a Manager, and a Voice, then assign it the Skill.
Put it to work. Add an A.I. Auto-Decision step to an Automation, point a Kickoff form at the Agent, or hand a single record over with Run Agent.

TIP: Set the certainty floor high to begin with and turn Wingman approval on, so every decision reaches you before anything runs. Lower it once the Run History has earned your confidence.

What is an Agent?

You add an Agent from your Staff list: open the create button and choose Add Agent, which sits alongside Add Staff Member. An Agent is a worker rather than a person, and it cannot sign in.

An Agent runs only the Skills you assign to it. You give it:

A name and character. The name appears on everything the Agent does, so pick something a client would not find strange, for example Onboarding Assistant.
A Manager. Every Agent answers to one staff member, who is accountable for the decisions it makes. That person cannot be offboarded while they are still responsible for an Agent, so accountability never quietly disappears.
A Voice. Start from a Tone and then edit it freely. The voice shapes how the Agent writes. It does not change decisions or facts.
An A.I. Execution setting. Which model the Agent thinks with each time it runs a Skill. Leave it to follow your Default Provider, or pin it to a specific provider and model so it stays put when the default changes.
Skills. Only published Decision and Compose Skills can be assigned.

Deactivating an Agent stops it running every Skill assigned to it, everywhere it is used. Its configuration is kept and it resumes exactly where it left off when you switch it back on. Removing an Agent is different: that deletes its identity and its Skill assignments, and cannot be undone.

What is a Skill?

A Skill gives A.I. one job. There are three kinds, and two of them are the ones you assign to Agents.

Decision Skill: picks which Auto-Template runs.
Compose Skill: writes a secure message to the person the automation is about.
Image Skill: not assigned to Agents. It is a route that lets a text-only provider still generate images. See the A.I. Connections article.

Skills we design and ship carry a System badge, and keep it after you configure them. Configuring the prompt or the conditions does not make it a different Skill; it makes it yours to steer.

Decision Skill: pick the template

A Decision Skill chooses between Auto-Templates you already built, so the outcome is always something you wrote and approved. You describe each situation in plain business language and say which Auto-Template it should run.

Applies To. The kind of record the Skill decides about. Each kind has its own Auto-Templates, and this is fixed once the Skill is published, because the conditions and templates were written for it. To decide about a different kind of record, create a second Skill.
Conditions. Written the way you would explain them to a colleague, for example the account is newly signed and onboarding has not begun. The A.I. reads intent rather than matching exact values.
The fallback. What runs when no condition matches. It is also used when the A.I. cannot return a usable answer. Leave it empty and nothing runs in those cases.
Data Sharing. The fields you switch on are the only ones sent with the request, and the only ones a condition can refer to.

Every condition is graded and the best match wins. Order only breaks ties between conditions that scored equally, so dragging a condition up the list does not give it priority over a better match.

In practice this is what replaces a person stopping to decide. A Kickoff form can route each submission to the right onboarding path overnight. An invoice follow-up can match its tone to the account, gentle for a good payer and firmer for a persistent one. A finished project can choose between an upsell sequence, a testimonial request, or a quiet close-out, based on how the work actually went.

TIP: Fewer, clearer conditions decide better than many overlapping ones, and sharing only the fields your conditions actually mention keeps the decision reliable. If two conditions overlap, combine them.

Before you publish, use Test this Skill. Pick a real record as a sample and the Skill runs for real against the model, but delivers nothing. You see which condition it settled on, how certain it was, and the model’s own answer. It also tells you when a condition was never chosen across your samples, which usually means it is too specific or an earlier condition is catching those records first.

Compose Skill: write the message

A Compose Skill writes one thing: a secure message to the person the automation is about. It does not choose between templates and it does not offer a menu of outcomes. The prompt is the adaptation.

The prompt. Describe how the note should read and which situations it covers. There is no separate list of scenarios: the situations you describe here are exactly what each run is matched against.
Data Sharing. Optional. The fields you switch on are the only ones the message may mention.
Message formatting. Allow paragraphs, emphasis and lists where the writing calls for it, or send as plain text.
Message Is From. Choose a specific staff member, or let it follow the coordinator or salesperson recorded on the person when the step runs. If they have nobody in that role, the message comes from the Agent’s Manager.

Wingman: the approval and certainty layer

Wingman is what keeps A.I. from acting on a guess, and it is where you decide how much autonomy to give a Skill.

Every run is graded for certainty. You set a certainty floor, which is the weakest match the Skill will act on. On a Decision Skill, anything below the floor is too weak to act on and the Skill takes its fallback instead. On a Compose Skill, anything below the floor is not recognised well enough to write from: nobody is messaged, the step stays on, and you are told.

You can also turn Wingman on as an approval step. With it on, a Decision Skill emails its decision to the staff you nominate and nothing runs until someone approves. A Compose Skill holds the finished message and sends it only when a person approves it.

Approve and the message is sent exactly as written, with your name on the record.
Decline and nothing is sent, nobody is contacted, and your name is on the record for that too.
Nobody answers in time and the message is dropped. Nothing is sent and there is no fallback message.

Opening an approval link only shows you the message. Nothing goes out until you approve it.

The grades themselves have meanings you can edit, so very confident, fairly confident and not confident mean what they mean in your business rather than in the abstract. The A.I. grades each run against those descriptions.

Where a Skill can run

Each Skill declares where it is allowed to run, and that is set by its type rather than chosen by you. Each place only offers the Skills that can actually run there, so you are never shown a writing Skill on a step that only picks templates.

A.I. Auto-Decision automation step: Decision Skills.
A.I. Compose automation step: Compose Skills, as their own step.
Kickoff Intelligent Automation: one slot that can offer both kinds, where they are allowed there.

On a Kickoff form, each submission is handed to the Agent and the Skills it holds decide what happens next, instead of every submission taking the same path.

You can also hand a single record over by hand. Choose Run Agent on the record, pick the Agent, and it lists the Skills it will run. Nothing is triggered until a Skill decides, and a Skill below its certainty floor pauses for approval instead of acting.

How do I see what the A.I. actually did?

Every Skill has a Run History, and every decision an Agent makes is recorded in the Run History of the Skill that made it. It shows how many runs there were, how many fell back, how many did not complete, how many are waiting for review, and whether the work arrived from Automations, from forms, or was handed over by hand.

For a message, the run opens to show who it was written for, the situation it recognised, how certain it was, and the message itself. Where a run stopped early, it says why, and where no message was written it says that plainly rather than showing an empty result.

IMPORTANT: Nothing is ever sent silently on a guess. A run below its certainty floor either takes the fallback or stops and tells you, and a Skill with Wingman approval on will not act until a person says so.