The agency lead is staring at 400 rows. Discovery happened. Someone said "send something." The proposal has been in "sent" for eleven days. The CRM still says the pipeline is healthy because nobody closed it out. Two account managers can "see" the record. Nobody owns it. He's shopping for AI on top of the graveyard — a copilot that will "nurture," another field called AI Score.
More fields will not save him. A definition of next action will.
Until a lead means "a person with a date, an owner, and one next step that could make money," you do not have a CRM. You have a museum. AI will only make the museum searchable. I've seen the same pile in multi-location retail: a year of "leads" from a giveaway, a Facebook form, and three store iPads. Nobody can tell you who is supposed to do what tomorrow morning.
The three doors, applied to a pile of names
Effectiveness: follow-up that produces booked work, a signed SOW, or a store visit with a basket — not activity for the dashboard. If the sequence doesn't ask for money or a meeting, it isn't sales follow-up. It's content. Efficiency: reps and store managers stop living in their heads and their personal inboxes. The next step is on the record. Handoffs don't require an archaeology project. Quality: the human on the other end doesn't get four contradictory emails from two locations and a "just bumping this" that sounds like a robot. You don't auto-send anything that could be a promise, a legal claim, or a medical claim. For agencies and retail, that still matters: don't promise a deliverable or a price the model invented.
If a CRM vendor cannot tell you how a record turns into a next action this week, they're selling storage. Don't add a model to storage and call it a strategy. AI CRM for small business is a next-action machine or it is a searchable graveyard. Those are the options.
What a lead is (write this on a wall)
Pick one definition. Publish it. Fire the old ones.
For an agency: a lead is a company that asked for a thing you sell, has a name and a way to contact a human, and has not said no. It is not a PDF from a conference in 2024. It is not a LinkedIn connection you liked. It is not a referral you never called. Statuses that mean nothing — "warm," "nurture," "working," "old" — get killed or defined in one sentence each.
For multi-location retail: a lead is a person who requested a quote, a special order, a local service (delivery, hold, appointment), or a B2B account — with a location attached. It is not every email on the receipt list. It is not a giveaway entry with no permission to sell. Location is part of the definition. A lead for store two is not a trophy for store one.
Then the only field that matters besides name and channel: next action. Not notes. Not a novel. One verb, one owner, one date. Call. Send the proposal. Text the hold notice. Book the walkthrough. Close-lost with a reason. If you cannot fill that field, the record is dead. Mark it dead. Dead is a gift. It stops the theater.
The model will do what you ask. "Draft follow-up for all open leads." You will get 400 polite paragraphs. Half of them will be to people who hired someone else, moved, or were never a fit. You will look busy. You will make no money. You will train the staff that "AI follow-up" means spam. Worse: if you auto-send, you will hit someone who already told you no, or a retail customer who asked you to stop, or a prospect whose last email was a complaint. Quality dies. Effectiveness dies with it. So we do not start with a blast. We start with a definition, a purge, and a tiny queue of live next actions.
Agency: follow-up that makes money
You already know the leak. The discovery call happened. Someone said "send something." A proposal sits in "sent." Day four, silence. Day eleven, the thread is awkward. Day thirty, it's a ghost.
What I actually install: statuses that force a date. New, waiting on us, waiting on them, proposal out, verbal yes, closed-won, closed-lost. Each one requires a next-action date or it cannot be saved. That's software configuration, not a new platform. Sometimes you don't need a new tool. You need the definition and a required field. A first-draft bump, not a campaign. When a proposal is three days old, the helper drafts one short note in your voice, using the actual deliverable names from the proposal — not "just circling back." The salesperson edits and sends. Human sends. If they want to add a new promise, they add it, not the model. Lost reasons you can count. Price, timing, not a fit, went in-house, ghost. Ghost after two human attempts is allowed. Ten "circling back" emails is not a strategy. One owner. If two account managers can "see" the record, nobody owns it. AI cannot fix a socialized pipeline.
Hypothetical (not a client, not a promise): one recovered $6,000 monthly project that was sitting on "proposal sent" pays for a lot of process work. I will not pretend that's your average. Look at last quarter's "sent" column. Count how many had a next step dated this week. That number is the diagnosis. Effectiveness is recovered proposals and faster yes/no. Efficiency is the salesperson not re-reading a 19-message thread to remember the ask. Quality is a bump that sounds like the person they met, not a sequence from a blog.
Multi-location retail: the list is not a pipeline
You have three stores. Each manager has a phone. Corporate (you) bought a CRM because the iPads were a mess. Now the mess has stages.
Typical graveyard: special-order deposits that never got a "it's in" text. B2B accounts that filled out a form and got a generic coupon. Local influencers and schools in a list nobody should have been blasting. Review-ask candidates mixed in with people who returned a product angry. What I'd do before any model: split by location, always. A next action with no store is how store one steals store two's customer or nobody texts at all. Split by job: special order, hold, local delivery, B2B quote, hiring (hiring is not a lead). Different SLAs. Permission. If you don't know you can text them, you don't. The helper does not get to "find a way."
Draft the two messages that actually make money: "your order is in at [store]" and "we can quote the [thing] — here's who to talk to at [store]." Human send on anything priced. Auto-send is a maybe only on a narrow, boring transactional notice you already approved word-for-word — and even then I'd rather a human hit the button until you trust it. AI is useful for turning a messy note ("called, wants the gray one, maybe Friday, store??") into a clean next action. It is not useful for inventing a nurture sequence that pretends a one-time buyer is now in a funnel. Quality in retail is the customer recognizing the store they walked into. If location B's follow-up sounds like a corporate newsletter, you trained them to ignore you.
Sometimes the answer is no new tool
If your team lives in Gmail and a spreadsheet, and the spreadsheet has no next-action column, adding a big CRM plus a copilot is how you spend 90 days migrating ghosts.
Try this first: one list of live work. Agency: proposals out and discovery scheduled. Retail: open special orders and open B2B quotes. Required columns: owner, next action, date, location (retail), dollar range if you know it. Weekly 20-minute pass. Anything with a date in the past is either done, redated, or dead. Helper drafts from that list only. Never from the museum. If that already works, we might still connect a real CRM later because you need reporting across locations or a team of five. That's a software decision after the definition exists. If we flip it, you will implement nothing and blame AI.
Allowed drafts: the bump, the "order is in," the "can we book 15 minutes," the internal summary of a call turned into a next action. Not allowed to send alone: pricing that isn't copied from your offer file, claims about results, anything that sounds like a contract, anything to a person who asked you to stop, anything that merges two locations or two clients because the names were similar. Agencies: your offer file (what you sell, what you never promise) sits next to the helper. Retail: store hours, hold policy, and who can discount sit there too. Without those files, the drafts will sound like every other shop and occasionally lie. Human still sends customer-facing mail until the shop has a reason to trust the narrow cases.
Paste a screen of live records, not the museum:
Turn this CRM/spreadsheet mess into next actions. Do not send. Do not invent prices.
Definition of a lead: [paste]
Statuses allowed: [list]
Offer file / never-promise list: [paste]
Paste of LIVE records only (owner, last note, last date):
For each row: keep / dead, owner, one next action, date, draft bump if "waiting on them."
Kill rows that are PDFs, giveaways, or no permission to contact.
Never merge two companies or two stores because the names look similar.
This is the $1,500 job, not a free poke around your Salesforce. You send a screen recording of the CRM or the spreadsheet — seven minutes of how a lead actually moves, including the part where it doesn't. I score whether you have a definition of a lead in writing; whether every live record has an owner, a next action, and a date; where last month's money actually came from; which follow-ups would hit Effectiveness without wrecking Quality; and whether this is a configuration problem on the tool you already pay for, or a behavior problem no tool will fix.
You get a short map. First workflow. What we would not touch. Math labeled as a model, using your counts, not a borrowed case study. If the honest card is "no new tool, just the definition and a weekly pass," that's the card. I'd rather you be mad at the simplicity than six months into a migration.
FAQ
Will AI CRM software fix my small business pipeline? Not until a lead has an owner, a date, and one next action. A model on a graveyard writes 400 polite paragraphs. That's spam with a dashboard.
Do I need a new CRM to start? Usually no. Required columns on the live list you already have: owner, next action, date, location if you have stores. Migrate ghosts later, if ever.
Can the model send the follow-up sequence? No. It drafts a bump from the live queue. A human sends. Auto-send is how you email someone who already said no.
You already know the museum is a museum. The assessment is $1,500 for a 45–60 minute owner interview and seven-business-day analysis of your graveyard: what to kill, what to draft, what to send, and who owns the date. If the honest card is "no new tool," that's the card.