When you create an AI Detection marker, or add an AI step to a multi-condition marker, you tell Odie what to look for in a few sentences. Odie then reads each account's conversations, alongside what it already knows about the account, and decides whether that account shows the signal. A short instruction like "detect churn" feels clear to you, but it leaves Odie guessing about what churn sounds like in a real conversation, and the marker ends up on the wrong accounts. This guide gives you a simple way to write instructions that Odie applies the same way your team would.
How Odie reads your instructions
- Odie reads the account's most recent calls, emails, tickets, and notes from the sources you chose, one at a time, and decides whether each one shows the signal.
- It then looks at those findings together and decides whether the account has the signal right now.
- When the evidence is unclear, Odie doesn't apply the marker.
- A name or keyword on its own isn't enough. The conversation around it has to match your instructions.
- One clear statement can be enough. Several vague hints usually aren't.
- Odie also knows the account itself, such as its renewal date, plan, and product usage, so your instructions can combine what customers say with where the account stands.
Answer four questions
Good instructions answer four questions. Each answer becomes part of a single sentence, so you can see exactly how the pieces fit together.
| Question | What to write | Example answer |
|---|---|---|
| What's happening? | The situation, described as what the customer says or does | Evaluating, trialing, or getting pricing from another vendor |
| Who has to say it? | Whose words count | Someone at the customer |
| How clear does it need to be? | How directly they have to say it | They say so directly, or ask us to match another vendor's price |
| Where's the line? | What separates this signal from situations that sound similar | The other vendor would replace us |
Put the answers together, and you have your Detection instructions:
Apply when someone at the customer says directly that they are evaluating, trialing, or getting pricing from another vendor that would replace us, or asks us to match another vendor's price.
If one of your answers is hard to write, the signal is usually too broad. Narrow it until each answer is easy, or split it into separate markers.
Write the line as what counts
Describe the line as part of what counts, rather than as a list of things to avoid. AI follows a clear description of what to look for more reliably than a list of exceptions.
| Instead of | Write |
|---|---|
| Don't apply when they mention a tool they used before switching to us. | Apply only when the other vendor would replace us. |
| Ignore pricing questions during onboarding. | Apply when pricing questions are tied to replacing us. |
If a similar situation still gets the marker, explain it in Additional Context as background rather than a command. For example: "Customers often mention the tool they used before us during onboarding. That's history, not an evaluation."
Describe what the customer says, not a conclusion
The first question trips people up most. "Churn," "risk," and "unhappy" are conclusions your team draws, not something a customer says on a call. Describe the words or behavior that lead you to that conclusion.
| Instead of | Write |
|---|---|
| Detect churn | Apply when a customer says they plan to cancel, won't renew, or will cut seats or licenses at renewal. |
| Detect competitor mentions | Apply when a customer says they are evaluating or getting pricing from another vendor to replace us. |
| Detect unhappy customers | Apply when a customer raises the same unresolved problem more than once, or asks to escalate it. |
| Detect expansion | Apply when a customer asks about adding users, teams, or products, or asks for pricing on a higher plan. |
| Detect champions leaving | Apply when our main contact says they're leaving the company or moving to a different role. |
Competitor mentions are a good example of why this matters. A customer might mention a competitor because they're evaluating it, because they used it before switching to you, or because another team uses it. Odie won't apply a marker just because a name comes up, so describe the situation you care about instead of listing names.
Choose between a rule and AI
Odie can use account data in AI instructions, but a rule is still the better choice for a hard cutoff. Rules check a field the same way every time and don't use tokens. AI is for signals that need judgment, especially ones that depend on both what customers say and where the account stands.
| Use a rule when | Use AI when |
|---|---|
| The answer is a value in a field, like a renewal date or plan | The answer depends on what customers say |
| You need the exact same result every time | The signal needs judgment, like how serious a concern sounds |
| You want to narrow the accounts before AI reads anything | You want to weigh a conversation against the account's situation |
For example, "Apply when a customer says their budget is shrinking and their product usage has dropped over the last few months" is a good AI instruction, because Odie has to weigh what the customer said against a trend.
When a signal has a hard cutoff, put the cutoff in a rule. For "enterprise accounts renewing soon that are unhappy," use an Enterprise Segment and a Rule-based Logic step for Renewal Window is Next 90 days, then an AI step for the unhappy part. The rules narrow the accounts first, so AI only reads conversations for accounts that pass them.
Use Additional Context for examples and terms
Keep your full answer to the four questions in Detection instructions. Use Additional Context for details that help Odie apply it:
- Terms your company uses, such as product names or what "QBR" means.
- Example phrases that count, and similar phrases that fall outside the signal.
- Background on situations you've seen, such as "General pricing questions during onboarding are routine. This marker is about pricing tied to replacing us."
Write branch paths that don't overlap
In a multi-condition branch, Odie checks the paths in order and the first match wins. Answer the four questions for each path.
- Describe one distinct situation per path.
- Put the most specific path first, so a broader one doesn't catch those accounts.
- Make sure the "What's happening?" answer is clearly different from path to path.
For example, if Path 1 is "customer mentions budget cuts" and Path 2 is "customer mentions cost concerns," almost every Path 2 account will also match Path 1. Make Path 2 about something different, such as evaluating another vendor.
Choose sources that match the signal
Evidence to analyze sets which conversations Odie reads. Choose only the sources where the signal actually shows up, since extra sources add noise. Odie uses what it knows about the account either way.
| Signal | Good sources |
|---|---|
| Renewal or budget plans | Calls, Emails |
| Product problems or bugs | Tickets |
| Evaluating another vendor | Calls, Emails |
| Your team's own observations | Notes, Commentary |
Check the results and refine
You can't test an AI marker before you save it yet, so start small.
- Save the marker for a small Segment.
- After it runs, open the marker on the Markers page and review the accounts that have it.
- If the wrong accounts have it, sharpen your "Where's the line?" answer or tighten "Who has to say it?"
- If accounts you expected are missing, check that the right sources are selected and that the conversation happened after you activated the marker. Then loosen "How clear does it need to be?"
- When the results look right, widen the Segment.
Before you save
- What's happening: Does it describe what a customer says or does, not a conclusion like "churn"?
- Who: Have you said whose words count?
- How clear: Would a passing mention or a name alone be enough? If so, raise the bar.
- Where's the line: Have you said what separates this signal from similar situations, written as what counts?
- Rules: Are hard cutoffs, like a renewal window or a plan, handled by rules or Segments?
FAQs
Why isn't my AI marker showing up on any accounts?
Odie leaves the marker off when the evidence is unclear. Check that the right sources are selected, that the conversations happened after you activated the marker, and that your "How clear" answer isn't so strict that real conversations rarely meet it.
Why is my AI marker showing up on accounts it shouldn't?
The instructions are probably describing a topic instead of a situation. Sharpen your "Where's the line?" answer, and say who has to say it.
Can I track several signals in one marker?
Use one signal per AI Detection marker. To sort accounts into different signals, build a multi-condition marker with a branch and give each signal its own path.
Do longer instructions work better?
Not by themselves. A few sentences that answer the four questions usually work better than a long, general description.