Agency client selection

Agency client fit scorecards prevent random cleanup pitches.

Choose the contractor client whose follow-up question you can actually investigate, with a reviewer who can act on the findings. Start there, even when another account has a bigger advertising budget.

Last updated: October 2, 2026

By George King, AI Cleanup Doctor

An agency client fit scorecard helps choose where to begin. Several clients may have plausible reasons to go first. Perhaps one has old estimates waiting for a decision, while another has just changed its website form and wants to check the new route. A score can sort that list. It cannot supply the missing records or appoint someone to review them.

This guide builds on the existing Agency Client Fit Scorecard. The tool adds points from selected checkboxes. The five evidence checks below are a manual step you take alongside it; the tool does not verify the evidence or enforce these checks for you.

Choose a question before you choose a score

A general complaint such as "our leads are expensive" leaves too much room for interpretation. The agency may mean cost per form submission. The contractor may mean the cost of a booked job. The office may be thinking about calls outside the service area. Those are different questions, and a cleanup review should not pretend to settle them all.

A workable starting point is narrower: "For recent requests from the repair form, can we identify who accepted each request and whether a next action was recorded?" That question has a source, an observable handoff and a clear output. The review can identify missing evidence without deciding whether the advertising campaign caused a sale.

Choose a service line and a review period before looking at the results. Keep the time zone and the meaning of each source label with the sample. If one system records a form submission and another records a callback, do not treat those timestamps as the same event. The agency lead outcome notes guide explains how to keep a tracked action separate from a supported business outcome.

The five evidence checks

These checks are a working method for this review, not a validated prediction model. Use "unknown" where the evidence is missing. A score built from assumptions will be harder to defend than a short note explaining what you still need.

Separate activity from a proven follow-up problem

Recent inbound activity gives you something to inspect. It does not establish that every inquiry is qualified, unique or ready to buy. Keep the original source category before adding a normalized label. If a record was marked "call," preserve that label; add "callback state unknown" as a separate observation rather than changing it to "lost opportunity."

Google Ads describes conversions as activities that an advertiser chooses to measure, including purchases, sign-ups and phone calls. For this review, the practical implication is to name the measured activity. A recorded call is not, by itself, proof of a booked job or collected revenue.

Then locate the specific question in the evidence. A form log may show that a notification was generated, while the working queue has no accepted owner. That supports the statement "owner acceptance is not shown in this sample." It does not prove that nobody replied through another channel. Ask the reviewer where that reply would be recorded before concluding that the handoff failed.

Keep conflicting records together. If the CRM says "closed" and the estimate sheet says "waiting," preserve both labels and their timestamps. The reviewer should resolve the disagreement. An AI summary should not silently pick the more convenient version.

Find the person who can review the result

The reviewer might be an owner, office manager, dispatcher lead or another authorized role. A name in the agency's contact list is not enough. Confirm that the person will review this particular question and what they may decide afterward.

Write the boundary in ordinary language: "The office manager will review unmatched form records and identify the correct owner. This exercise will not send customer messages, edit estimates or change ad settings." Record the reviewer's decision next to the evidence it relies on, so someone returning to the note can check the reasoning.

NIST's AI Risk Management Framework is voluntary guidance for considering trustworthiness across the design, use and evaluation of AI systems. Our narrower practice here is to keep AI-assisted organization inside a defined human review. NIST does not endorse this scorecard, and using it does not establish compliance or certification.

If nobody can accept the result, hold that client and investigate another eligible account. Do not solve a missing reviewer by granting the agency broader authority or by letting software decide what to send.

Reduce the sample before sharing anything

For an owner-handoff question, useful fields may include a redacted record ID, source, service category, broad area, received time, recorded owner role, response state and next action. Keep enough context to understand the event, but remove information that does not help answer the question.

The first review should not depend on names, complete phone numbers, email addresses, street addresses, payment details, recordings, passwords or unrelated message bodies. Replacing a name with an ID does not automatically make the remaining record anonymous. An unusual service description or a precise location may still identify someone.

Keep the original source under the client's existing controls. Share only the authorized sample through the agreed route, and agree who can access it and what happens to it after the review. The purpose-first redaction guide offers a field-by-field starting point. If the question cannot be answered without material outside the agreed scope, stop that sample and ask for a revised scope.

A fictional comparison between two client candidates

The following accounts and records are invented to demonstrate the method. They are not customers, testimonials, measured results or performance benchmarks.

A-12 goes first because the question can be checked. The proposed output is small: for each of the two records, identify supporting evidence for owner acceptance or leave the state unresolved with a named next step. Finding a reply in another system would change the interpretation. The review must allow that result.

B-07 may become suitable later. Ask for an authorized reviewer and a narrow estimate question, such as whether a defined group of estimates has a documented next action. Do not ask for the entire customer database just to give the account a higher score.

Use the number without letting it overrule the evidence

The interactive scorecard currently groups checkbox signals around leakage, demand timing and agency fit, then shows a directional fit result. Record the selected signals and the reason for each selection. Two accounts are not comparable if one score uses verified records while the other uses guesses.

A score of 70 is not a 70 percent probability of a sale, retention or recovery. Nor can positive demand signals cancel missing authorization. Apply the manual checks separately and record the outcome as ready for the specified internal review, needs clarification, or hold.

Revisit the result when there is new evidence or a changed review scope. Do not raise the score because the agency wants a proposal to move faster. If the original question changes, retain the earlier note so a later reader can see why the decision changed.

Finish with a review packet someone can use

A short packet is enough when it tells the reviewer where to look and what to decide. Include the source and period, the exact question, the fields retained and excluded, the reason this client comes first, and the person responsible for the next step. Keep references to the evidence rather than copying unnecessary private records into the note.

For the fictional A-12 example, the packet could read:

The finish line is an accepted review output, not a promise to recover a job. If the evidence remains unavailable, say so and record the owner of the next check. If a record still cannot be resolved, keep it unresolved and say what is missing.

How to introduce the review to the client

When you already work with the client, describe the observed question directly: "We have a small sample of repair-form requests where owner acceptance is unclear. Would you like an internal review of that handoff before deciding whether to change the process?" Only say you have that sample when you do, and obtain the necessary approval before using it.

An ads agency may start with what happens after a tracked call. A website agency may inspect the destination of one form. A CRM consultant may investigate conflicting states. An SEO agency can separate a search-generated inquiry from the later handoff. These are possible review angles, not claims that those channels are failing.

Describe the proposed output and its limits. Avoid client-retention promises, guaranteed revenue or referral-fee claims. Any commercial or referral terms need their own explicit agreement; a fit score does not create them.

Frequently asked questions

Which agency client should get the first cleanup review?

Choose a client with a specific review question, inspectable demand records, an authorized human reviewer and a sample you can minimize. A large budget alone is not enough. Resolve authority or sensitive-data problems before starting.

Does a high score mean the client will buy or recover revenue?

No. The score is an internal prioritization aid, not a probability or a forecast. It does not predict retention, rankings, leads, bookings, sales, referral income or revenue.

Does the scorecard automatically enforce these five checks?

No. The current interactive tool adds points from selected checkboxes. The five checks in this guide are a separate manual review. Keep a written reason and evidence reference for each decision.

Can an agency send follow-up messages after scoring a client?

Scoring does not grant permission to contact customers, access a CRM or change a campaign. Those actions need their own appropriate authorization and review. Keep the first exercise internal.