Estimates go cold for a boring reason. There was no second contact. Not a badly written second contact, no second contact at all, because the person who sent the quote moved on to the next site visit and the record sat in the pipeline with nobody's name against it. That is the problem AI can genuinely solve here, and it is worth being precise about what solving it looks like, because the version most people buy is a reminder sequence and a reminder sequence is not follow-up.

What AI is actually good at in this workflow

One thing, and it does it well: turning the estimate record into a draft that already knows the job.

A generic reminder says "just checking in on the quote we sent." A useful one says which roof, which slope, the two options you priced, the thing the customer said they were worried about on the day, and what happens to the schedule if they decide this week. The second version takes a salesperson eight minutes to write from scratch and thirty seconds to approve if something else drafted it. That difference is the entire return on this workflow, and it compounds because the follow-up that actually gets sent is the one that was already written.

The second thing it can do is watch the pipeline for you. Estimates that passed their decision date, jobs with no logged contact since the visit, quotes that were opened four times and never answered. A human sales manager does this on a good week. Software does it every day.

What it must never do

Draw the line at anything commercial, and draw it in writing before the build.

  • Price and scope. No discounts, no "we could probably do it for", no reworked line items. If the customer pushes on price, that is a human conversation and the system's job is to flag it fast.
  • Availability and dates. A drafted message must not promise a start date the schedule cannot support. This is the single most common way an automated follow-up creates a problem the office has to clean up.
  • Urgency it invented. No fake deadlines, no "prices increase next week" unless prices genuinely increase next week and somebody authorised saying so.
  • Warranty, guarantee, or contract language. Approved wording only, pulled from a document, never generated.
  • Facts about the job it was not told. If the estimate record does not contain the answer, the draft says the office will confirm. It does not guess at the material, the permit, or the lead time.

Write these as a list your team can see, not as an instruction buried in a prompt. The list is the specification. The prompt is just how it gets enforced this month.

Decide the cadence before you choose a tool

Cadence is a commercial decision and it belongs to the business, not the platform's default template.

What we would run for most trades: a same-day message confirming what was quoted and what the next step is, a genuine check three or four days later that asks one question rather than repeating the quote, and a decision-forcing contact a week or so after that which makes it easy to say no. Then stop, and move the record to a slow list. Three good contacts beat eight nagging ones, and the third one is where most of the value sits because almost nobody sends it.

Long-cycle work needs a different shape. A homeowner comparing three roof replacements is not on the same clock as somebody who needs a water heater this week. If your business sells both, build two sequences and route by job type or value. Running one sequence across everything is what makes follow-up feel like spam to half the list and feel absent to the other half.

The stop rule matters more than the sequence

Name the conditions that take a record out of the automation and hand it to a person. A reply of any kind. A question about price. A complaint. A request for a callback. An out-of-office or a bereavement. Anything the classifier is not confident about.

The rule most businesses forget is the one for a job that was already won or already lost by another route. If the office booked it over the phone and nobody updated the record, the sequence keeps chasing a customer who has paid a deposit. That is a small technical detail and a large reputational one.

Draft for review, then earn the right to send

Start every implementation at draft-for-review. The salesperson sees a prepared message and approves, edits, or discards it. You find out what the system gets wrong while the cost of being wrong is one click.

Move to automatic sending only for the narrow, low-consequence part of the sequence, and only once you have a month of drafts where the approval rate is high and the edits are cosmetic. High-value estimates should probably stay on review permanently. The three permission levels worth arguing about, read, draft, and execute, are set out in our piece on approval workflows for AI agents, and estimate follow-up is the clearest case on the whole list for stopping at draft.

The prerequisite nobody wants to hear

This workflow is only as good as the estimate record behind it. If the quote lives as a PDF attachment and the context lives in a salesperson's head, there is nothing for the system to draft from and it will produce exactly the generic reminder you were trying to replace.

Before building anything, check that the record holds the fields the message needs: what was quoted, the options and their differences, the customer's stated concern, the decision date they gave you, and who owns the relationship. If those fields are not being filled in reliably, fix that first. It improves your follow-up on its own, with or without AI, and it is the difference between a useful implementation and an expensive one. Our AI readiness checklist walks through the same test for any workflow.

Measure accepted estimates and lost reasons

Reply rate and open rate will both go up. Neither one is the point.

Track acceptance rate on quoted work, time from estimate to decision, and the share of estimates that get a decision at all rather than fading. Then track the thing that pays for the project twice: the lost reason. A follow-up sequence that reliably extracts "we went with someone cheaper" or "we postponed until spring" or "we never got the revised figure" is producing pricing and process intelligence you cannot buy anywhere else. Most businesses have no idea why they lose, because nobody asks the question at the only moment it gets answered honestly.

Take the baseline before you build. Our measurement plan guide covers the smallest version worth keeping.

What a first version looks like

One job type, one sequence, drafts only, one owner reviewing them, and a before number for acceptance rate. Run it for a month against the estimates you would otherwise have let go quiet, and compare. If it works you will know, because the pipeline will be shorter and the reasons will be written down.

That scope is what AI workflow automation is for: one repeatable workflow with visible review points, rather than an assistant pointed at the whole business. The estimates you have already sent are the cheapest pipeline you own, and they are usually the least worked.

FAQs

Can AI negotiate quotes?

No. It can summarize objections and draft responses, but price and scope decisions need authorized people.

Should messages send automatically?

Often the safer first version is draft-for-review, especially for high-value estimates.

What is a good success measure?

Accepted estimates and clearer decisions matter more than the number of reminders sent.