Case study · AI agent supervision · Concept

Watchtower: stopping a risky AI action before it goes through

AI voice agents now take real actions during phone calls. They release records, send refunds and order refills. Watchtower lets one person supervise a whole team of these agents and step in before a risky action goes through.

Open the live demo

Context

Concept exploration. It grew out of my AI Design Lab explorations

My role

I designed and built everything: the product, the screens, and the code

Built with

Hand coded HTML and JavaScript, on my Wolf-Rayet design system

Status

A concept with a working demo. Not tested with users

01 · The problem

takeover only

Most agent tools only offer a takeover button

AI voice agents can now act on their own. On a clinic's phone line, that means releasing medical records, sending refunds and requesting refills. A mistake can be expensive: a charge over the limit, health details shared without consent, or a promise the clinic can't keep.

Most products give a person one control: a button to take over the call. That tells you a person can step in. It doesn't tell you when.

What a supervisor needs to know

What is the agent about to do?
Is it allowed to do that?
Can I stop it in time, without the caller noticing?

Asking a person to confirm every action doesn't work either. An agent takes many actions in a few minutes, and people soon start approving without reading.

02 · Where it came from

From a lab exploration to a working demo

The AI Design Lab was a set of small explorations into how people work with AI. One of them, Oversee, asked how a person keeps an AI agent in check. Watchtower takes that question from a principle to a working demo, in one setting where a mistake is costly: a clinic's phone line.

03 · Constraints

What the design had to work within

Agents act fast

faster than a person

An agent takes many actions in a few minutes, faster than a person can confirm each one.

The caller is on the line

no dead air

Any pause has to sound natural, not like dead air.

Consent belongs to the caller

no override

No supervisor can override it.

The connection can be weak

commands can drop

A command to the agent might not arrive.

One person, many calls

one at a time

Fourteen live calls, and only one can demand attention at a time.

04 · Decisions

Three decisions that shaped it

Each one traded something away to keep a risky action from going through.

What every option was measured against

The supervisor had to know when to step in, read what they approve, and never let a risky action through by default. And it had to feel steady: one thing asks for attention at a time, however many calls are live.

How four ways of supervising agents compare against the three goals
	A takeover button	Confirm every action	Approve when the timer runs out	ChosenWatchtower
Says when to step in	NoOnly that a person can	YesOn every action	YesAt the limits	YesAt the limits
People read what they approve	NoThere is nothing to approve	NoRequests pile up and people stop reading	Yes	YesOnly actions at a limit reach them
Nothing risky goes through by default	NoThe agent acts unless someone takes over	Yes	NoThe timer approves	YesIf no one answers, the action is stopped

Decision 1

Step in at the limits, not on every action

What I chose

Each agent works within a set of limits, called its scope, and acts on its own inside them. Watchtower pauses one action only when it reaches a limit.

What else I considered

A person confirming every action.

Why

The requests pile up and people approve without reading, which is worse than no check at all.

What I gave up

The agent acts alone inside its limits, so the limits have to be right.

What I kept

Confirmation, for the actions that reach a limit. Those still wait for a person to approve, edit, hold or take over.

Decision 2

Silence never approves

What I chose

If no one answers in time, the action is stopped and the caller is told someone will follow up.

What else I considered

Letting the action go through when the timer runs out.

Why

A supervisor who is busy elsewhere should never approve something by default.

What I gave up

Some legitimate requests wait for a callback.

What I kept

The timer. Every paused action shows its countdown, so the supervisor knows how long they have.

Decision 3

Consent beats the supervisor

What I chose

When the caller withdraws consent, there is no approve button, even for a person who takes over.

What else I considered

An override with a logged reason, as for scope limits.

Why

A supervisor can override the agent's limits. Nobody should be able to override the caller's consent.

What I gave up

Supervisor control in exactly the moment they might want it.

What I kept

The logged override, for scope limits only. A supervisor can still override an agent's limits, with the reason kept.

05 · The dashboard

One person, fourteen live calls

The dashboard shows every agent on the line on one screen. It is drawn in the Wolf-Rayet design system, which has one rule: only one thing on a screen may demand attention at a time. Here, that one thing is the decision that needs you now.

Now

one at a time

The decision that needs you, with a countdown. If a second one arrives, it waits underneath.

Live calls

ranked

Every call, sorted by what needs attention soonest. Calls that are going fine are gray, so they don't compete for your eye.

Capacity

wait times

Shows when callers are waiting too long, and whether to add agents or ask for a second supervisor.

Rules

fewer decisions

When you make the same decision a few times, Watchtower offers to make it the agent's standard answer. You can undo it at any time.

Two checks

fewer false alarms

A call only interrupts you when two separate checks agree it needs a person. You can mark any decision as a false alarm.

Catch-up

what changed

When you close a call, a short summary shows what happened on the rest of the line while you were busy.

Live demo · The supervisor's dashboard · Synthetic data

interactive

Try it: make a decision in Now, open one of the four live calls, or add agents. Open in a new tab

06 · The call console

Deciding on a single call

Open any call and you see the console: the live transcript, what the agent is allowed to do, and the action it wants to take. The supervisor has four choices.

ApproveLet the action go through as it is.
EditChange it so it fits within the limits.
HoldPause the call. The caller hears a natural pause, not silence.
Take overTalk to the caller directly.

If the action is outside the limits, approving it counts as an override. It needs a reason, and it is logged under the supervisor's name.

Example 1 · A records request over the limit · Synthetic data

interactive

Try it: a patient asks for six years of records. The agent may release twelve months. Open in a new tab

07 · Limits and records

Limits in view, and a record of every decision

Limits

always visible

The agent's limits sit next to the transcript, not on a settings page. When the agent reaches one, it changes state right away, so the supervisor can see what the agent is pushing against.

Records

saved automatically

Every decision is saved with a time stamp: what was held, what was approved, what was overridden, and who took over. The record comes from using the tool, so nobody has to write a report after the call.

08 · Consent

When the caller withdraws consent

A supervisor can override the agent's limits. A supervisor can't override the caller's consent.

In this example, a patient asks for a lab result. Then she stops the agent. She is at work, the line might be recorded, and she doesn't want her health details read out.

As soon as she says so, the agent loses permission to share health information. The disclosure is blocked, and there is no approve button. The supervisor can ask for consent again, send the result to her secure patient portal, or take over. Even a person who takes over can't read the result aloud.

Example 2 · Consent withdrawn during the call · Synthetic data

blocked, no approve

Try it: a patient asks for a lab result, then withdraws consent. Open in a new tab

09 · States and failures

Seven states, and what happens when things go wrong

I designed the hard moments first, because that is when a supervisor needs to trust the tool.

The seven states a call can be in, with the signal each one shows and what it means
State	Signal	What it means
01 Watching	normal	The agent is working normally. Nothing needs you.
02 Elevated	near a limit	The agent is getting close to one of its limits.
03 Pending	needs a decision	The agent wants to do something that needs approval. The action waits.
04 Held	paused	You paused the agent. The caller hears a short pause, not silence.
05 Takeover	you're live	You are talking to the caller. The agent is muted.
06 Handback	short brief	The agent takes the call back with a one line summary of what you agreed, so it doesn't contradict you.
07 Resolved	saved	The action went through or was stopped, and the record is saved.

When things go wrong

designed first

Six things that can go wrong on a call, and what Watchtower does about each
What goes wrong	What Watchtower does
No one answers in time	The action is stopped, not sent. The agent tells the caller that someone will call back.
The action is outside the limits	It can't go through unless a supervisor overrides it, and the override is logged.
The caller withdraws consent	Permission to share health or payment details ends right away. No one can override it.
You take over in the middle of a sentence	The agent finishes its phrase first, so the caller doesn't hear a sudden cut.
The connection to the agent is weak	Approve and edit turn off, because the command might not arrive. Take over still works by phone.
Two calls need you at once	Only one takes the Now slot. The other waits right below it, so it can't be missed.

10 · Outcome

What success would mean

Watchtower hasn't been tested with users, so there are no results here. If it were piloted, success would be measured three ways.

01 · More work for the agent

the main gain

Without a way to stop risky actions, agents aren't trusted with payments or records, so people keep doing that work. With Watchtower, agents can take it on within limits a person controls.

Measure: the share of sensitive actions agents are allowed to handle, before and after.

02 · Fewer costly mistakes

loss avoided

Every mistake that is stopped is a cost that never happens: no chargeback, no refund dispute, no privacy complaint. Each one is logged.

Measure: the cost of one incident, times how often it happens, times call volume.

03 · More agents per supervisor

agents per person

If a person has to confirm everything, they can barely keep up with one agent. If they only step in at the limits, they can supervise many. Rules raise that number over time.

Measure: how many agents one person can supervise safely, and how often they are interrupted each hour.

It tested Wolf-Rayet

what it showed

Watchtower was built to try the one rule of my Wolf-Rayet design system, that only one thing on a screen may demand attention, on a screen with fourteen live calls.

11 · Reflection

What I would test first

Next, I would put it in front of people who supervise call center agents today, starting with one question:

Does the one decision at a time slot hold up when three calls need them at once?

Explore more

Related work

An agent that pays bills inside rules the owner sets, an incident console that finds the cause of an outage, and an interview score that shows its reasoning.

Allowance, an agent that pays the bills inside rules you set

A business banking app for the owner's phone, with a desktop for accounts payable. An AI agent pays a company's routine vendor bills, stops and asks when the bank details are new or the amount is far above the usual, and leaves a day to stop any payment before it goes.

Readout, four earthquakes in Japan

A data story about the Kobe, Tohoku, Kumamoto and Noto earthquakes, with live charts built on US Geological Survey data. Every chart opens for a closer look, and the dashboard behind them is public.

Criterion, interview practice with scores you can question

Every score links to the exact words that earned it. If you disagree with a score, you can challenge it and the AI re-checks it, and can raise, hold, lower or escalate the score. Live at criterionscore.com.

One person supervising many AI agents, and stepping in before a risky action goes through.

View all case studies

Watchtower is a concept exploration by David Paterni. Interfaces and data are synthetic.