Every business I've worked with has a graveyard of half-finished transactions. Contracts sent but not signed. Invoices delivered but not paid. Approval requests sitting in someone's inbox behind 400 other emails. The work is 90% done. The revenue and the outcome are stuck in the last 10%.

Follow-up is the most expensive form of nagging in business. It's expensive because someone senior usually has to do it (a partner, an office manager, the founder), and because the cost of not doing it shows up as cash flow gaps, stalled projects, and quiet churn. It's also the kind of work AI agents are unusually good at, because it's structured, repetitive, and driven by clear state changes.

Here's how we actually build follow-up agents that work, and where they earn their keep.

Why follow-up breaks down in the first place

The failure mode is almost always the same. Someone sends a document, an invoice, or a request. They mean to check on it in a few days. Then a customer escalation happens, or payroll needs running, or a new patient walks in, and the follow-up slips. A week later the sender has moved on mentally. The recipient has forgotten. Both sides now need to reconstruct context to move forward, which raises the friction of finishing the task.

Most CRMs and billing systems have some kind of reminder feature, but they tend to send the same email on day 7, 14, and 30 regardless of what's actually happening. They don't read replies. They don't adjust tone. They don't know that the client already called and said the check is in the mail. So people turn them off, or they get filtered as spam, and the follow-up problem returns.

What a follow-up agent actually does

When we build a follow-up agent for a client, it's usually a small system with four moving parts:

  • A source of truth. The agent connects to wherever the "state" of the thing lives: your e-signature tool, your accounting system, your PM tool, your intake platform. It pulls status on a schedule (say every hour) and identifies items that are stuck.
  • A definition of stuck. Not everything overdue needs a nudge. An invoice at day 3 past due for a 20-year client is different from an invoice at day 15 for a new one. The agent uses rules you set: aging thresholds, dollar amounts, customer tier, prior payment behavior.
  • A communication loop. The agent drafts and sends the follow-up through the right channel (email, SMS, sometimes a voice call), tracks replies, and understands what came back. If a customer replies "paying Friday," the agent notes that and pauses further nudges until Friday plus a grace day.
  • An escalation path. When the agent can't resolve something, it hands off cleanly to a human with full context: what was sent, when, replies received, what it recommends next.

That last piece is what separates a working agent from a spam machine. Follow-up agents should reduce human work by handling the 80% of nudges that are routine, and route the tricky 20% to a person with everything they need to decide in 30 seconds.

Three workflows we see most often

Unsigned documents

Contracts, engagement letters, NDAs, patient consent forms, employment paperwork. The agent watches your e-signature platform, and when a document sits unsigned past a threshold (24 hours for hot deals, 72 for standard), it sends a short, human-sounding nudge. Then it checks whether the recipient opened the document. If they opened it but didn't sign, the message is different than if they never opened it at all. If someone replies with a question ("can we change the payment terms?"), the agent doesn't try to answer. It flags the deal owner and shares the exact question.

For a services firm we worked with, this one workflow cut average time-to-signature from 11 days to 4, mostly because deals stopped falling into the "I meant to sign that" void.

Unpaid invoices

This is where the ROI math is easiest. If you have $200K in aging receivables and you shorten collection by two weeks, you free up real working capital. The agent segments invoices by age and customer, sends progressively firmer reminders, handles the common replies ("send me a copy," "which PO was this against," "we paid this already"), and pulls the invoice PDF or the payment record without a human touching it.

For dental and medical practices, the same pattern applies to patient balances after insurance. The agent sends a reminder, offers a payment link, answers questions about what the balance is for at a general level, and escalates disputes to the office manager. It stays strictly on the billing and operational side.

Pending approvals

Purchase orders waiting on a manager. Design proofs waiting on a client. Prior-authorization requests sitting with a payer. Change orders stuck between accounts. These are usually the least tracked and the most damaging, because they hold up downstream work that's already been scheduled.

An approval agent watches the queue, understands who owes what, and nudges accordingly. For internal approvals it might ping in Slack or Teams. For external ones it emails and, in higher-stakes cases, follows up with a scheduled call from a voice agent to confirm the request was received. The agent tracks aging by approver, so you can finally see that Steve in finance is your bottleneck, not the vendors.

The details that make or break these agents

Building the happy path is easy. What makes a follow-up agent worth using is how it handles the messy middle:

  • Reply understanding. The agent needs to read replies and classify them accurately: paid, disputing, needs more info, out of office, unsubscribe request, angry. Getting this wrong is how you end up emailing a grieving spouse about a $47 balance.
  • Tone control. First nudge is friendly. Second is direct. Third references consequences (late fees, hold on services). The agent should never sound robotic or threatening, and it should always leave the door open to a human conversation.
  • Suppression rules. No follow-ups on weekends or after hours. No follow-ups to accounts flagged as sensitive. No follow-ups during known service outages when your invoices are wrong. These rules prevent the small disasters that erode trust.
  • Auditability. Every message the agent sends should be logged and reviewable. For healthcare specifically, this means HIPAA-aware channels for anything patient-related, and no PHI going through tools that aren't covered under a BAA. Confirm the specifics with your own counsel and your vendors.

What to measure

If you're going to invest in follow-up automation, measure it against real numbers, not activity. The ones that matter:

  • Days sales outstanding (DSO) for invoices.
  • Time from send to signature for contracts.
  • Approval cycle time by stage.
  • Percent of follow-ups that resolve without human touch.
  • Escalation quality: when a human does get pulled in, how often does the agent's summary contain what they need?

We generally target 70 to 85% of routine follow-ups closing without human involvement within 90 days of go-live. Under that, the agent's rules or its reply handling need work. Over that, you should be raising the bar and expanding scope.

Where to start

Pick the follow-up workflow that's costing you the most right now. For most services businesses, it's AR. For deal-driven teams, it's contracts. For healthcare front offices, it's often insurance follow-up and patient balances. Start with one workflow, one system of record, one channel, and get it working cleanly before layering on the next.

The teams that get the most out of these agents treat them as coworkers with a narrow scope, not magic buttons. You define the playbook. The agent runs it every hour, every day, without forgetting or getting tired. Your people spend their time on the exceptions and the relationships. If you want to map out where follow-up is bleeding your team dry and what an agent could take off their plate, talk to our team and we'll walk through it with you.

Frequently Asked Questions

How is this different from the reminder features in my existing software?

Built-in reminders send the same message on a fixed schedule and don't read replies. A follow-up agent adjusts based on whether the document was opened, what the customer said in their last reply, payment history, and the value of the item. It also hands off cleanly to a human with context when it hits something it shouldn't handle alone.

Will customers know they're talking to an AI?

Our recommendation is to keep the agent's messages honest and short, and to make it easy to reach a real person. In practice, the first few nudges look like a professional email from your business, and when a customer needs a real conversation, they get one. For voice, we're clear about disclosure and follow whatever your jurisdiction requires.

Is this safe for healthcare front-office use?

Yes, when built correctly. That means using HIPAA-aware infrastructure, signed BAAs with any vendor touching PHI, minimum necessary data in messages, and audit logs for everything sent. The agent stays strictly on operational tasks: appointment reminders, balance follow-up, insurance status, intake completion. Clinical questions get routed to your staff. Confirm the compliance specifics with your own counsel.

How long does it take to get one of these running?

For a single workflow (say, AR follow-up on one accounting system), we typically go from kickoff to live in three to five weeks. The bottleneck is almost never the AI. It's getting clean access to your systems, agreeing on the rules, and writing message templates your team is proud to send.

What happens when the agent gets something wrong?

You build the review muscle from day one. Every escalation and every reply gets logged. In the first month we review a sample daily with the client and tune. After that, most teams shift to weekly. Mistakes still happen, but they get caught early and the rules get sharper over time.