Real builds, measured results

The office answered the phone from 8am to 5pm on weekdays. Everything else went to voicemail. Instead of guessing what that cost, we pulled six months of call records: roughly 90 calls a month arrived outside office hours, about 22 of them genuine new-job inquiries. Next-morning callbacks recovered around five. The rest had already reached the next contractor on Google.

At the company's average job value and margin, that gap was worth roughly $18,400 a month in gross profit — going to voicemail.

A voice agent that answers every out-of-hours call, qualifies the job, checks live technician availability, offers two appointment slots, confirms by SMS, and writes the job into the field service system before anyone wakes up. It was scripted from transcripts of the office's own best coordinator, so it sounds like the company — not like a chatbot.

• Any mention of a gas odor, flooding, or no heat with an infant or elderly resident routes straight to the on-call technician's phone.
• It never quotes a price beyond the published diagnostic call-out fee.
• Addresses outside the service area get a polite decline and a texted referral.

Before launch, the agent replayed thirty real recorded calls and its booking decisions were compared against what the office actually did — then the client's own staff spent a round of live calls actively trying to break it. It ran a week in shadow mode with every transcript emailed to the owner, then weekends only, then all out-of-hours. Nothing was trusted before it was measured.

Two workflows consumed the working day. Brokers phoned and emailed carriers three to five times per load for status updates — about 600 hours a month spent chasing status. Meanwhile, inbound spot-quote requests took an average of 45 minutes to answer, in a market where the first credible quote usually wins the load.

A track-and-trace agent that contacts carriers at scheduled checkpoints — SMS first, a voice call if there's no reply — parses whatever comes back, updates the transport management system, and involves a broker only when something is off-plan. And a quoting agent that reads inbound requests, extracts the lane and requirements, prices against live market data and the firm's own margin rules, and returns a quote within minutes. The language model extracts and drafts; deterministic code does the pricing.

• A load projected more than an hour late, silence after two attempts, or any mention of a breakdown alerts the broker immediately — with full context.
• It never commits capacity that hasn't been booked.
• Quotes outside the approved margin band, for a new customer, or above a value threshold always go to a human for one-click approval.

Five hundred historical quote emails were replayed and extraction was measured field by field before a single live quote went out. A month of past loads ran through the tracking logic to confirm it caught every exception the brokers had caught. Tracking launched on one broker's book first; quoting ran in approval-required mode for a month before any threshold was loosened.

Every month, staff emailed and called clients for bank statements, receipts and missing invoices — roughly 190 hours a month, and the work everyone deferred longest. Behind it, around 40,000 transactions a month needed categorizing, with about 15% requiring a manual touch or a query back to the client. Month-end close averaged eleven business days.

A document chaser that knows precisely what is outstanding for each client, chases on an escalating schedule across email and SMS, accepts uploads by reply — photographed receipts included — validates them and files them to the right client and period. And a reconciliation agent that categorizes transactions against each client's own history, never a generic chart of accounts, auto-posts only above a measured confidence threshold, and drafts plain-English queries for anything genuinely ambiguous.

• Chasing stops entirely after the fourth touch and hands off to a human — no client relationship is worth a deadline.
• It never posts to a tax-sensitive account without human review, and never alters a reconciled period.
• No client query is sent without staff approval.

Categorization was replayed against a full quarter of already-reconciled transactions, and the auto-post threshold was set where measured accuracy exceeded 98% — with a senior accountant reviewing 200 sample categorizations and every drafted query. The chaser launched on thirty clients with staff approving every message for the first week; reconciliation ran a full close cycle in review-everything mode before any threshold moved.

Paid advertising delivered roughly 400 inquiries a month. About 45 were viable. Paralegals spent around 90 minutes screening each one — approximately 530 hours a month, most of it spent ruling out the 355 cases the firm would never take. Worse, viable leads waited an average of four hours for a response, in a practice area where the firm that responds first usually signs the client.

An intake screening agent that engages every inquiry within sixty seconds by SMS, gathers the facts conversationally, scores viability against the firm's own criteria — with a written reason for every factor, not a bare number — and books qualified prospects directly into an attorney's calendar. Alongside it, a case summary agent that turns the intake record and uploaded documents into a one-page brief before the consultation, with every factual claim cited back to its source document. No citation, no claim.

• It never gives legal advice, never says whether someone has a case, and never estimates value — every conversation opens by stating that no attorney-client relationship is formed.
• Filing deadlines are computed by deterministic code; the language model is never permitted to do that arithmetic.
• An expiring deadline, a near-threshold score, or a distressed caller escalates to a human immediately.

Two hundred historical inquiries — one hundred accepted, one hundred declined, each with the recorded reason — became the gold-standard test set. The only score that truly mattered was false negatives: viable cases wrongly declined. The target was zero, and the threshold is deliberately tuned to over-refer to humans. The managing attorney reviewed and signed off every line of the script, two attorneys verified every citation in twenty sample briefs, and the agent ran two weeks in shadow mode with human review of every decision before going live on a single advertising source, then all inbound.

Three structural drains. Account managers spent about fourteen hours per client per month assembling reports — pulling numbers from analytics, paid social, search and the CRM, then writing commentary — around 630 hours a month, performed by the most expensive client-facing people in the building. The fourteen-hour figure came from observation, not estimate. Meanwhile, client sales teams took hours to days to respond to the leads the agency generated — and renewals were judged on cost per acquisition, a step the agency didn't control. And approved creative briefs queued three to five days for a first draft.

A reporting agent that pulls platform data nightly, compares it against prior periods, cross-references the campaign changes logged in the project system, and drafts commentary explaining why numbers moved — producing a review-ready report for the account manager. A speed-to-lead agent deployed into client funnels that responds to form submissions within sixty seconds, qualifies, and books warm prospects for the client's sales team. And a brief-to-draft agent grounded in each client's approved past copy, so brand voice is inherited rather than invented.

• Every figure is retrieved by an API tool — the model writes commentary but never produces a number, and unexplained movement is flagged, not explained away.
• Nothing is sent or published without human review, and the lead agent never negotiates or quotes.
• Messaging consent, quiet hours, opt-outs, and each client's own AI-disclosure policy are verified per client before launch — asked about, never assumed.

The last three months of reports were regenerated for ten clients and put side by side with what was actually sent; account managers scored the commentary for accuracy, and anywhere the agent asserted a cause without evidence, the instructions were tightened. Three senior writers blind-reviewed drafts against a brief for voice consistency. Reporting launched on ten clients with full review, speed-to-lead with two willing clients, and drafting stayed internal until the writers trusted the output.