On this page
Last week was acts, where a wrong write can be found and put back. Decides is where the choice reaches a live caller first, so the gate becomes a written list and a record someone reads.
Reads and drafts keep a person between the agent and the world. Acts takes that person out for one step, and Chuck's piece last week argued that the step is safe when the change can be seen and put back. Decides is the next rung. The agent chooses between options, the choice reaches a live person, and nobody approved it first.
Our phone agent is a good example. A caller says what they need. The agent decides what kind of call this is and what happens next, while the caller is still on the line. The caller experiences it as it happens.
There is no queue and no review before the outcome.
I've been promising this story since the opener, including how the first version of the agent became the second. That gets its own piece next week. This one covers what the agent is allowed to decide alone and why I think that list is the right size.
What it decides
The agent sorts each call into one of a short list of kinds. A new inquiry from someone who wants to work with us. A question about billing or an invoice. A request for support. Something urgent. A solicitation. Each kind has a defined next step, and the agent's decision is which kind it is hearing and which next step follows.
The solicitation is the one people ask about (we get a lot of them). The agent listens for a caller whose real purpose is to sell us something, or to keep a conversation going until a pitch can fit in, with nothing in it of value to us. It is also the hardest of the kinds to get right, because it is a judgment about what the caller wants, and a topic can't settle that.
It has unlimited patience and no weakness for a friendly opener, which turns out to be a good combination for the job.
I kept the list short on purpose. Every kind I add is another place the agent can be wrong in front of a caller, and every kind has to be described well enough that the agent and a person would sort the same call the same way.
Five kinds a person can tell apart in a sentence are easier to get right than twelve that overlap.
What the agent captures matters as much as what it decides. A summary of the call, and the details of anyone new, go to email, to the team chat and into the pipeline. The decision it made and the information behind it are in front of whoever picks the call up. The agent hands the call to a person and leaves the next move to them.
Reversibility stops being the gate
Chuck's rule for acts was reversibility: if a write is wrong, can you find it and put it back. That doesn't carry over to this rung.
A caller who was sent down the wrong path has already had that experience, and no edit to a record changes what they heard.
Two other parts of the promotion rule carry the weight here. The first is that the set of outcomes is bounded. The agent chooses among a handful of next steps that someone wrote down, and it can't invent a new one. A wrong choice is then one of a known number of wrong choices, and I can name each of them before the agent takes a single call.
The second is that the record is complete enough to catch a miss quickly. Every call is captured, so a call sorted into the wrong kind doesn't disappear into the day. Someone reads the summary, sees where it went, and sees that it doesn't fit. The miss I watch for most is the solicitation filter turning away a real inquiry, because that's the one call where being wrong costs us a prospect. The fix has two parts. The description of that kind gets sharper for the next caller, and the caller who went down the wrong path gets a follow-up from a person.
That changes what review means. At acts, review looks at a change before anyone depends on it. At decides, review looks at yesterday's decisions, because they're already out. The version of the gate that still works is a review soon after the fact, from a record that shows what the agent heard and what it chose.
The paragraph that goes first
The third part of the promotion rule is that someone has written down what the agent may do alone. We introduced it in August as the intent test: before anything goes live, one paragraph that a person can read in under a minute and know what they agreed to.
For the phone agent that paragraph has three parts. The first is what the agent is for. It's the front door, and its job is to work out who is calling and why, and to make sure the right person hears about it. The second is what it decides alone: which kind of call this is, and which details to capture. The third is what has to go to a person. I spend the most time on the third, because it's where a front door can start behaving like an employee. Anything that would commit the company to something, a price, a date or a promise, is the kind of line that paragraph has to draw, in words the agent can't read two ways.
Writing it takes longer than it sounds. The first draft always describes what you hope the agent does. The useful draft describes what you can live with when the agent is wrong.
If you can't say what a wrong decision would cost on a bad day, the agent shouldn't be making that decision yet.
Two weeks ago I asked whether a message sent in someone's name should say so. On the phone the same question is about the caller: whether the agent says what it is, and when. That belongs in the paragraph too, because it should be a decision someone made and not a default nobody looked at.
What I'd like to hear
Think of a system in your business that makes a choice while a customer is waiting. Is there a written list of what it may decide, and who reads the record of its choices afterward? Post your answer in the comments. The answers that name the list and the reader are the ones I learn from.
Next week I'll cover how the first version of our phone agent became the second, and what moved us from one to the other.