A marketing agency can use AI to review account health by scoring every retainer monthly against a short scorecard of signals you already have, such as results, responsiveness, payment and relationship strength, then asking the AI to summarise the evidence and flag accounts that dropped, so you call the at-risk client before they raise a concern.
It does not need a prediction model. A ten- or twenty-client agency has too few clients for statistics to mean anything, and the real problem is not a lack of cleverness but a lack of time: the warning signs are scattered across inboxes, meeting notes and invoices, and nobody collects them until the client says they are leaving. The AI's job is to collect and summarise. Yours is to decide and act.
What an at-risk retainer looks like from the inside
By the time a client gives notice, the decision was made weeks earlier. Looking back at a lost retainer, the same patterns tend to appear:
- Replies that used to take an hour now take three days, or come from a different person.
- A monthly review that used to have the owner now has an assistant, or is rescheduled twice.
- Questions about the contract: notice period, scope, "what are we actually getting each month?"
- Results flat or down for two months running, with no conversation about why.
- An invoice paid late for the first time.
- A new marketing hire at the client who asks for "a fresh look at everything".
- Silence: no complaints, no praise, no requests. It is the easiest signal to miss, because nothing arrives to be noticed.
None proves anything alone. The method below turns them into something you can count, so a pattern is visible across the month and across the whole book.
The scorecard: six signals, each scored 0 to 2
Keep it small enough that account managers will fill it in. Six signals, three possible scores each, 12 points maximum.
| Signal | 0 (healthy) | 1 (watch) | 2 (at risk) |
|---|---|---|---|
| Results against agreed goal | On or above target | Under target for one month | Under target two or more months, no plan agreed |
| Responsiveness | Replies within a day | Two to four days, or slower than before | No reply for a week, or contact has changed |
| Meeting attendance | Decision-maker attends | Delegate attends | Cancelled or skipped twice |
| Payment | On time | Up to 14 days late | More than 14 days late, or disputed |
| Scope and contract | No questions | Asks about scope or extras | Asks about notice, termination or reduction |
| Relationship | Warm, proactive | Neutral, only reactive | Terse, or new stakeholder with no history |
Total score 0 to 3 is green, 4 to 6 amber, 7 and above red. A single 2 on payment or contract questions moves an account to red by rule, whatever the total, because those two are the closest to a decision. Adjust the thresholds after a few months, when you can compare scores against what happened.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowA worked month: ten retainers
Illustrative case: a ten-person agency has ten retainers worth $52,000 a month in total, ranging from $2,500 to $9,000. On the first Monday, the account managers complete the scorecard for their clients in a shared sheet or Notion database. They add a one-line note per signal for anything scored above zero. The scores come out like this:
| Client (generic) | Monthly fee | Score | Colour | Main reason |
|---|---|---|---|---|
| Pest-control firm | $3,000 | 2 | Green | Paid late once, otherwise healthy |
| Car repair garage | $2,500 | 5 | Amber | Owner skipped last review; results flat |
| Furniture maker | $4,500 | 8 | Red | New marketing manager; asked about notice period |
| Online clothing store | $9,000 | 4 | Amber | Return on ad spend under target for a month; slow replies |
| Other six clients | $33,000 | 0 to 3 | Green | No concerns |
The red account, the furniture maker, is worth $4,500 a month, or $54,000 a year. The amber online store is worth $108,000 a year. Seeing them next to each other is the value of the exercise: a red furniture-maker account is urgent, but the amber clothing store is the bigger exposure if it slides, and the agency should talk to both this week. Neither would have been noticed in the normal rush of delivery.
Where AI fits: gather, summarise, rank
The scorecard is a manual habit. AI makes it faster and more complete by doing three jobs.
Gathering the evidence
Export, or paste, what you have for each account over the last 30 days: the meeting notes, the email subject lines and dates, the invoice and payment dates, and the monthly result figures. Remove names and anything personal. Then ask the assistant to extract the signals:
Below is 30 days of material about one agency client: meeting notes, a list of email dates and who sent them, invoice and payment dates, and monthly results against target.
Fill in this scorecard, using only the material: results vs goal, responsiveness, meeting attendance, payment, scope/contract questions, relationship tone.
For each signal give a score (0 healthy, 1 watch, 2 at risk) and quote the evidence in a short phrase. If the material has no evidence, write "no evidence" and score 0.
Then list any pattern you notice across signals, and any change since the previous month if the previous scorecard is below.
Do not guess at the client's intentions.
MATERIAL: [paste]
PREVIOUS SCORECARD: [paste]The instruction to quote the evidence is what lets the account manager check the score in ten seconds. The rule against guessing intentions stops confident speculation: "the client is probably unhappy" is not a finding.
Summarising the book
Paste the ten scorecards, minus names, into one prompt and ask for a one-page review: accounts that moved up or down in score, the three most at-risk by value, patterns across accounts (for example, three clients mentioning budget cuts), and one suggested action for each amber or red. The output will be a draft. The account lead edits it into the monthly review document, since she knows that the furniture maker's new marketing manager used to be a client of a rival agency.
Drafting the response
For any account that needs a conversation, the AI drafts the message. It is the most obviously useful step and the one that needs the most editing, because a message that sounds like a retention script does harm.
Drafting the call and the email for an at-risk client
For the furniture maker, the account lead's plan is a call, not an email, because the signal is the new manager. She asks the assistant for a call plan, giving it the scorecard evidence and the account history. A trimmed, illustrative output:
Aim: reintroduce value to the new marketing manager and learn what she wants. Open with congratulations on her appointment and ask what she'd like to change. Do not mention the notice period. Prepare a one-page summary of the last six months: spend, enquiries, cost per enquiry, and three things done that she may not know about. Ask: what does success look like for her in the next quarter? End with a proposed next step: a 45-minute planning session next week.
The lead keeps the structure and drops one suggestion, a discount offer, which would signal weakness before the client has said anything. She also adds a point from her own knowledge: that the manager's predecessor set the targets, so the targets themselves may need resetting. The AI could not know that.
If the client has already given notice, the email draft changes. A win-back message should be brief, honest and not defensive:
Subject: Before you go: a quick conversation?
Hi [first name],
Thanks for letting us know. I'm sorry we haven't delivered what you needed, and I'd like to understand where we fell short, whether or not you decide to stay.
Would you give me 20 minutes this week? I'll listen first, and if there's a change you'd want us to make, I'll tell you plainly whether we can make it.
[your name]Ask the assistant for three versions at different lengths, pick the most direct, and send it from a person. Avoid any version that offers a discount in the first message, since it teaches clients that threatening to leave earns one.
Scope drift: a quick sum that explains many exits
One cause of churn rarely appears on a results chart: the agency quietly doing more than the retainer pays for, resenting it, and the relationship cooling. Illustrative sum for the car repair garage client at $2,500 a month. The retainer was priced on 25 hours a month at $100. Over three months the logs show 31, 34 and 37 hours, because the client keeps asking for small extras: a new offer banner, a change to the opening-hours text, an extra report. That is 102 hours against 75 paid, 27 hours or $2,700 of unpaid work.
The AI can find this if you give it the time logs and the original scope: "List requests that fall outside the scope below, group them by type, and total the hours." The useful reaction is not a bill for the extras, which would sour things, but a conversation: "We've been adding about nine hours a month beyond the plan. Shall we make those part of the scope at $3,400 a month, which is 34 hours at the same rate, or trim back to the original 25?" Either answer is better than silent resentment, and the client sees the value they have been getting.
A quarterly review the client actually sees
The internal scorecard has an external twin: a short quarterly review that you hold with the client. Its purpose is to remove doubt before it grows. Ask the assistant to draft a one-page agenda and summary from your reporting data: the goal agreed at the start, what moved, what did not and why, three things planned for the next quarter, and one question for the client. Illustrative output for the online clothing store:
Goal: 400 online orders a month from email and paid social. Last quarter: 310, 355, 372. Cost per order down from $11.40 to $9.60. What did not move: repeat-purchase rate, flat at 18%. Next quarter: test a post-purchase email sequence; rebuild the two lowest-converting product pages. Question for you: is repeat purchase still the priority, or is it new customers?
The summary is useful because it ends on a question. A client who has to answer it is a client who is engaged. If they cannot be bothered, that itself scores a 1 on meeting attendance next month.
Payment signals: handle them without panic
A first late payment is more often an admin problem than a decision to leave, but it is also the one signal that appears in your own records without any effort. Add the payment date to the scorecard automatically if your invoicing tool allows it. When an invoice is overdue by a week, ask the AI for a short, friendly message that assumes an oversight, mentions the invoice number and offers to resend. Do not escalate on the first occasion. If two consecutive invoices are late, move the score to 2 and treat it as a conversation about the relationship, not about money. Chasing late payments with AI reminders that sound human has wording for each stage.
Win-back after the client has left
A departed client is not always a lost one. Diary a message for three months after they leave, which is long enough for the reason to have played out. Give the assistant the exit notes, which should contain the real reason, and ask for a brief, honest note that refers to it and offers something specific. For the furniture maker, whose exit reason was "we wanted to try someone cheaper", an illustrative draft:
Hi [first name],
I hope the last few months have gone well. I wanted to say there's no sales pitch here: if the new set-up is working, I'm glad.
If it would help, I'd be happy to look at your last quarter's numbers for free and tell you what I see, whatever agency you're with. Just reply if that's useful.
[your name]It offers a small, genuine favour and no pressure. Keep a list of departed clients in the tracker with their reason for leaving and a re-contact date, and review it each quarter. Not every client returns, but one who does has tried the alternative, so the conversation starts from what they missed.
Turning scores into actions
A score without an action is paperwork. Agree in advance what each colour triggers.
- Green: no action beyond the normal rhythm. Once a quarter, send an unprompted summary of wins to keep the relationship visible.
- Amber: the account manager makes a call or sends a personal note within five working days, asking one open question: "Is this still what you need most?" The score is rechecked in two weeks.
- Red: the account lead or founder calls within two working days. A recovery plan is agreed within a week and written down, with a dated review.
A rule such as "any payment more than 14 days late or any question about notice moves the account to red" is better than judgement on the day, because it prevents an optimistic account manager from talking themselves out of a signal.
Learning from the clients you lose
Every lost retainer is a free test of the scorecard. After a client leaves, go back over the previous six months of scores. Which signals appeared first? Which scored 0 right up to the end? If the final three months show a steady rise from 3 to 9, the scorecard worked and the problem was action; if the score was 2 the month before they left, a signal is missing. Ask the assistant to look at the last six scorecards and the exit conversation notes, and to propose one new signal or one changed threshold. Add it if it makes sense. For a broader method for the reasons customers cancel, why customers cancel and how to run a churn analysis with AI covers the analysis side.
A few signals tend to be missing in agencies. The first is "single point of contact": if the client's only advocate leaves, the account falls. Add a line, "do we know two people at the client?". The second is "scope drift", where the client has asked for extras for three months in a row without paying for them, and the team resents it. The third is the agency's own capacity: if the account manager is overloaded, responsiveness will drop on your side first, and the client will feel it.
Retention begins at onboarding and at the contract
Reviews catch trouble but do not create goodwill. The strongest retention measures are in the first month and in the contract. A client whose goals, approvers and reporting rhythm were agreed in the kick-off, as in agency client onboarding with AI, has fewer reasons to drift. A clear notice period and a defined scope, as covered in AI clauses for agency contracts, remove the vagueness that often triggers the "what are we getting?" questions. And a monthly report that links results to the goal agreed, not a wall of metrics, keeps the value visible.
Privacy, fairness and the limits of the method
This is internal analysis of your own client relationships, and it should stay proportionate. Do not paste client personal data, private messages from individuals or anything covered by confidentiality terms into an AI tool without checking your contracts and the tool's data settings. Summarise from your own notes and counts, not from raw email bodies where you can avoid it. And remember that a client with a low score is not a problem to be managed but a person with a reason. The scorecard's purpose is to prompt a conversation. If the conversation shows the client's needs have changed, the right result may be a different scope, or a graceful ending, and not a rescue.
Also avoid false precision. A score of 7 is not meaningfully different from 6; the colours are what matter. And the first three months of scores are for calibration, not decisions: you are learning what a typical month looks like for each client.
How to check the routine is working
- Every account has a score every month, filled in by the person who knows the client, not by the assistant alone.
- Every amber and red account has a logged action and a date.
- At least one conversation each quarter started because of the scorecard, not because the client raised something.
- The notice-to-score check: after any loss, the scorecard showed amber or red at least a month before notice. If not, change the scorecard.
- The review takes an hour or less. If it takes longer, the scorecard has too many signals or the evidence is too hard to gather, so cut the signals or automate the collection.
Over a year this gives you something more useful than any single save: a record of which signals matter for your clients, and a habit of calling before you have to.
Questions about reviewing agency account health with AI
What are the earliest warning signs a retainer is at risk?
Slower replies, fewer attendees at reviews, new people appearing who were not part of the original decision, questions about contract terms, late payment, and a client who stops disagreeing with you. Any one is weak evidence. Several in the same month are worth a call.
Can AI predict which clients will leave?
Not reliably with a small client book. With ten to thirty clients there is no data to train on. What AI does well is summarise your own notes and signals against a scorecard you define, so nothing slips, and draft the follow-up. The judgement stays with you.
Should I tell the client I score their account health?
The scorecard is an internal tool, like a pipeline review. You do not need to share the score, but you should act on it openly: a call to ask how things are going is honest and welcome. Avoid anything that feels like surveillance.
How often should account health be reviewed?
Monthly for the whole book, with a faster weekly look at any account already marked amber or red. Quarterly is too slow for monthly retainers, because the warning signs often show up and resolve, or not, within thirty to sixty days.
Further reads
- Ad Agency AI Operating System: A Lightweight Version in Notion — A Notion layer where the scorecard data can sit.
- What Is AI Churn Prediction and Can a Small Gym Use It? — How churn prediction works in a business with far more members.
- Automate Client Onboarding With AI: Forms, Contracts, and Emails — Forms, contracts and emails for a strong first month.
- AI Rollout Plan for a Ten-Person Marketing Agency — Plan AI use across the whole agency team.
- Gym Automation Mistakes That Increase Member Churn — Automation mistakes that push people to leave.
- AI Tools for Agencies: What a Five-Person Team Pays Each Month — Seven layers priced per seat for a five-person agency, with annual and monthly totals, a leaner option and a skip list.
- Marketing Agency Pricing After AI: Retainers, Packages and Deliverables — Reprice agency retainers line by line instead of cutting everything: a worked $4,000 retainer, three pricing models and how to answer an AI discount request.
- Website Maintenance Plan: Packages a Web Agency Can Add With AI — Three maintenance tiers with tasks and price bands, tool costs per site, and the minutes AI saves on reports and update checks.
- AI Scope of Work Generator for Agencies: SOWs With Change Control — A worked example from client call to signed SOW: counted deliverables, exclusions, change-control wording and a filled change request form.
- AI Social Media Monitoring: A Monthly Insight Report for Clients — Lead with the monthly insight report: a template, a filled example, alert thresholds, a sentiment spot-check and what listening tools cost for five clients.
- Content Automation With AI: A Weekly Pipeline for a Small Agency — A Monday-to-Friday content pipeline for a small agency: briefs, AI drafts, human edit, client approval gates and scheduling, with a quality checklist.
- AI Tools for Business Analysis: 6 Questions to Ask Your Sales Spreadsheet — Six questions for your sales sheet: top customers, margin, slow movers, seasonality, churn and price changes, each with a check.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: none from vendors; this tutorial describes a scorecard method. The agency, its clients, scores and figures are illustrative.