Skip to content
Back
Customer Success Account Prioritization: Outlit's AI Rubric

Customer Success Account Prioritization: Outlit's AI Rubric

Author: Josh Earle

Josh Earle

TL;dr

Customer success account prioritization changes when an AI reads every account every morning: finding problems gets easy, and keeping the list short becomes the hard part. Outlit's AI asks one question about each account: how important is it that the owner acts in the next two business days? It answers on five levels, from None to Critical, and each owner gets a short to-do list each morning, usually five to ten tasks. Three ideas set it apart from a human prioritization framework. A customer waiting on us comes before our own plans. The conversation itself (who wrote last, what we haven't answered, what we promised) is the strongest signal. And most of the rules exist to stop the AI from inventing urgency. The rules are our opinions, learned from real customer success work and tested in our evals. The full rubric and a copy you can use are below.

Why does an AI need a different prioritization rubric?

Human frameworks exist because nobody can read everything, so teams sort by account size, red health scores, and upcoming renewals. An AI reads every email and ticket before the owner logs on, so it fails the other way: it finds a plausible reason to act on most accounts, which is no list at all. The hard part is getting it to say "nothing to do" for most of the book. (If most of your accounts never get a human look today, start with scaled customer success.)

So the list is short on purpose. Each owner gets five to ten tasks a morning, one per account, with the five most important at the top. It's easy to drown a CSM in alerts, and an alert they learn to ignore is worse than none. Most of our effort goes into making sure every task on the list is worth a look, and we keep adjusting the rules as we grade real lists.

What question does the AI ask, and what are the five levels?

For each account, the AI answers one question: how important is it that this owner acts for this customer in the next two business days?

LevelMeaning
None Nothing to do for this customer now
Low Can wait a week or more
Medium Worth doing this week, but not in the next two days
High Act within the next two business days
Critical Act today

The question is about one person and a short window. So a risky account its owner can't help this week scores Low, and a healthy account with a pricing question unanswered for two business days scores High.

The AI only picks the level. It doesn't write the reason the owner sees. That reason is built from the emails and records themselves, with dates, such as "Asked for a renewal quote on Fri Nov 6. No reply from us since." An AI asked to explain its choice will sound convincing whichever level it picked. A reason built from the record can only say what happened.

What does the AI weigh?

These rules are opinions. They come from what we've seen in real customer success work, and they're the version that holds up best in our evals, which reward a list that stays quiet unless something is worth a look. The AI reads them before it scores each account:

  • A customer waiting on us comes first. A direct question unanswered for two business days is High, however small the account. Teams tend to rank their own quarterly review prep above an unanswered email the health score never sees.
  • New beats old. Something that changed in the last two days can be High. An account that has been red for months with nothing new is Low.
  • A renewal date alone isn't a reason. A renewal counts when something is still open: a price question, a competitor, a trial ending, or a deadline. In the last two weeks before renewal, only what can still close it counts, like a signature or a direct question. That stops a pile of check-in emails in the last two weeks.
  • A missed renewal stays on the list. If the renewal date passed and nobody recorded the outcome, it's High until someone does.
  • Our promises count. A commitment due in the next three days is High. A date the AI picked for a step in its own plan is only an estimate. Without that rule, every step in the AI's plan becomes a deadline it set for itself, then flagged itself for missing.
  • Keeping a customer beats growing one, but a customer asking to buy more is High.
  • Things the owner can't change stay Low, such as FYIs, routine usage, and anything only the customer can do.
  • Account size breaks ties. It doesn't decide who gets attention, so a small customer who is waiting on us outranks a large one with nothing new.

How does the AI read the conversation?

The signals that matter most come from the back-and-forth with the customer, which most teams track from memory. The AI can't act on "waiting on us" until someone defines it, so Outlit calculates each pattern below before the AI scores. The numbers are our current defaults. Back-test them on your own accounts.

Real contact means both sides talked. That's a call, or an email thread where the customer replied. Six unanswered emails from the owner count the same as none. A CRM "last activity" field that counts your own sends is one reason a CRM struggles with customer success. It also takes resolved customer context to know that the caller, the reply from a personal address, and the person who filed the ticket all belong to one customer.

A reply we owe is a thread where the customer sent the last message, asked something, and nobody on our side has answered. After two business days it's High. Thank-you notes and automated mail don't count, and a call afterwards ends the wait. A support ticket becomes the CSM's to answer when the CSM is on the thread, or when nobody on our side has replied for two business days.

Our own unanswered emails become a follow-up after three business days, but only if the email asked something or proposed a next step. A pointless nudge costs more than a missed one.

Going quiet before a renewal is a warning. No real contact for 60 days with renewal within 45 days lifts the account to at least High. On its own, silence only takes a spot on the list when one is free. The same silence 90 days out doesn't count. This close to renewal, silence can mean they're deciding without you.

Promises we made are specific commitments from our side, such as sending an export by Friday. "We'll look into it" and vague roadmap talk don't count. A promise due within three days is High.

What customers say is easy to over-read, so each reading spells out the near miss:

The customer…CountsDoesn't count
Pushes back on price Asks for a discount or last year's price Says it's expensive without asking for anything
Looks at alternatives Runs a trial, gets a quote from a competitor, plans a migration Asks "how do you compare with X?"
Signals they may leave Gives notice, or asks to turn off auto-renewal Asks to turn off auto-renewal because that's their policy for every vendor
Is frustrated with us Says so in their own words Files a neutral bug report

"Unless the price comes down, we won't renew" means the customer is in doubt. It doesn't mean they're leaving yet.

How does the AI avoid inventing urgency?

The AI can ignore a written rule, so the rules that matter most also run in Outlit after the AI scores:

  • A date the AI set for its own plan reads "planned for" and never goes overdue. Only a date a person set, a meeting, a contract deadline, or a promise to the customer is firm.
  • Routine activity alone keeps an account no higher than Low.
  • A renewal date alone doesn't create a task.
  • A task marked done can't come back on the same evidence. A new message from the customer can bring it back.
  • Only specific facts, such as price pushback in an unanswered email, can take a spot on a full list. A high score can't.
  • In the other direction, a customer who gives notice or says they're leaving is never below High, and going quiet before a renewal lifts an account as described above.

The short list is a guard too. Only five tasks are recommended each morning, one per account and normally Medium or above. The rest wait below in "Up next". No more than two of the five can be replies we owe, and no more than two can be follow-up nudges, so promises and risks keep their spots. When the list is full, an urgent task replaces the weakest one. An account shown before stays off until something new happens, except for replies we owe, promises, risk warnings, cancellations, and questions the AI needs the owner to answer.

How does the rubric decide? Four examples

These made-up examples show how the rules apply.

Account AAccount BResultWhy
Small, with a direct question unanswered for 3 business days Large, nothing new this week A is High, B is Low A customer waiting on us comes first, and size only breaks ties
Renewal in 10 days, asked four days ago for last year's price Renewal in 40 days, routine activity only A is High, B is Low B's renewal has nothing open
No real contact for 60+ days, renewal in 30 Same silence, renewal in 90 A is at least High, B isn't lifted Silence matters only close to renewal
A step the AI dated for Thursday A step the CSM dated for Thursday Only B can go overdue The AI's dates are estimates

What to do once an account says it's leaving is covered in helping at-risk customers.

Evidence check

An illustrative, anonymized example. Today is Tuesday 2026-11-10.

  • What the AI sees: renewal on 2026-12-08, 28 days away, from HubSpot, on an account below the owner's median revenue. The last real contact was a Gong call on 2026-10-06. In a Gmail thread, the customer's finance lead wrote last, on Friday 2026-11-06.
  • What changed: on 2026-11-06 the finance lead asked for the renewal quote at last year's per-seat price. That's a reply we owe and price pushback in one email.
  • What doesn't change it: on 2026-11-09 the owner sent setup docs in a different thread and asked nothing, so that isn't a reply. Thirty-five days without real contact is short of the 60-day line. Yesterday the question was one business day old, so on this smaller account it didn't count as waiting yet. Today it does.
  • Level: High. A customer question has waited two business days, and it's about price before a renewal. Because the email pushes back on price, the task can replace the weakest one if the list is full.
  • What the owner sees: "Asked for the renewal quote at last year's per-seat price on Fri Nov 6, 2 business days ago. No reply from us since. Renewal Dec 8." The task closes when Outlit sees the reply, and a draft is ready to edit and send.
  • Next check: tomorrow morning. If there's still no reply, the task stays on the list.

Copy the rubric

Paste this into your agent's instructions, and have code fill in the customer details.

You rank one account owner's customers each morning. For each customer,
answer one question with one level. Don't explain your answer.

QUESTION
How important is it that {owner} acts for {customer} in the next two
business days? Use only this customer's details. Today is {today}.

LEVELS
None: nothing to do for this customer now
Low: can wait a week or more
Medium: worth doing this week, not in the next two business days
High: act within the next two business days
Critical: act today

RULES
1. A customer waiting on us comes first. A direct question unanswered
   for 2+ business days is High, whatever the account size.
2. New beats old. A change in the last two days can be High. Nothing
   new since last week is Low.
3. A renewal counts only when something is open: a price question, a
   competitor, a trial ending, a deadline. In the last two weeks, only
   what can still close the renewal counts.
4. A renewal date that passed with no recorded outcome is High.
5. Our promises due in the next 3 days are High. A date an AI agent
   picked for its own plan is an estimate, not a promise.
6. Keeping a customer beats growing one, but a request to buy more is High.
7. FYIs, routine usage, things already handled, and things only the
   customer can do are Low.
8. Account size breaks ties. It doesn't decide who gets attention.

CUSTOMER DETAILS (filled in by code)
renewal: {date}, {N} days away | none in the next 120 days
annual revenue: {amount} | unknown
last real contact: {date}, {call | email} | none in 180 days
still open: one line each (price question, competitor, trial ending...)
waiting on us: {subject}, customer's message unanswered {N} business days
open steps: {step}, due {date}, set by {person | customer | agent}
recent evidence: up to 5 lines | nothing new in the last 7 days

Build the conversation patterns (including going quiet before a renewal) and the guards from the sections above in code, along with the tie-break (nearer renewal, then larger revenue). A prompt that says "never treat the agent's own dates as urgent" is asking the model to police itself.

The Claude Code guide covers wiring up the agent.

How do you tune a prioritization rubric?

  • Version every change, and treat a reworded question as a new question.
  • Grade real morning lists without knowing which version made them, before and after each change. That's how we tune ours.
  • Treat done, won't do, and wrong as labels for your next grading round. In Outlit they close or hold the task. They don't retrain anything.
  • Back-test the numbers. For each customer that churned, find the first morning it would have reached High. For each one that renewed, count the mornings it sat on the list with nothing to do.

Where does Outlit fit?

Outlit runs this rubric every day to decide what lands in each owner's inbox. Before scoring, it matches calls, emails, and tickets from your call recorder, inbox, and help desk to the right customer. When something needs a reply, Outlit drafts it, and you choose which customer groups it sends for on its own and which wait for a person. Start with the copied rubric and a back-test. Running it on your whole book every day is the part Outlit automates.

Frequently asked questions

Does account size matter at all?

Yes, in a few small ways. The AI sees revenue as background. It breaks ties: nearer renewal first, then larger revenue. And on accounts below the median revenue of the owner's book, a question doesn't count as waiting until it's two business days old. Beyond that, a large account with nothing new shouldn't outrank a small one where someone is waiting.

Why doesn't sentiment drive the ranking?

We take an overall sentiment reading, but the AI doesn't see it when it ranks. A "sustained negative sentiment" signal was moved to background context after back-tests. Frustration counts only when the customer puts it in their own words. Slack isn't part of the ranking today.

When is a short daily list the wrong design?

When the owner already reads every thread. A CSM with a dozen accounts who talks to each one weekly already knows who's waiting on a reply.