Skip to content
Nahl TechnologiesNahl Technologies

Drowning in Check Calls: AI Document Automation for an Indianapolis Freight Brokerage

By Samia Zaman

Reading time: 11 minutes. The third problem family in our engagement series: not slow quotes (Kestrel), not missed calls (Redbud) — paperwork. The email-shaped work that eats a logistics office alive.

The company

Limestone Freight Co. is a freight brokerage on the west side of Indianapolis, near the Plainfield warehouse corridor — 22 employees moving full-truckload and LTL freight through the Crossroads of America, where I-65, I-70, I-69, and I-74 meet. Roughly $9M gross revenue; like every brokerage, the real business lives in the ~15% net margin between what shippers pay and what carriers charge.

Their constraint isn't freight. It's that every load generates twenty emails, and humans process all of them: quote requests arriving as free-text emails, rate confirmations, carrier documents (insurance certs, rate cons, BOLs, PODs), and the industry's great time thief — check calls, the "where's my truck?" status requests that arrive all day, every day, from every shipper.

Industry estimates put ops-rep time on status communication at a quarter to a third of the day. At Limestone, five ops people were doing knowledge work maybe two hours a day. The rest was inbox.

Phase 1 — Observe (week 1)

For a brokerage, the process map is an inbox map. We categorized two weeks of the shared ops mailbox — every message, tagged:

Composition of Limestone's ops inbox over two weeksOf roughly 340 emails a day, 31 percent are status requests, 22 percent are quote requests, and 19 percent are carrier documents such as rate confirmations, proofs of delivery and insurance certificates. Together those three categories are 72 percent of the inbox and are mechanical work. The remaining 28 percent is judgment work that stays with people.Ops inbox — about 340 emails a day31%Status requests“where's load #4471?”22%Quote requestsfree-text: lane, weight, date19%Carrier documentsrate cons, PODs, insurance certs28%Everything elsereal judgment work72% mechanical — filled barsreading a document, looking it up, transcribing the answer back out
Two weeks of Limestone's shared ops mailbox, categorised message by message.

72% of the inbox was mechanical — reading a document or a question, looking something up in the TMS, and transcribing the answer back out. None of it requires a human decision. All of it was consuming humans.

Phase 2 — Analyze & Quantify (week 2–3)

Every input labeled, as always:

InputValueLabel
Mechanical share of ops inbox72% of ~340/dayClient-confirmed from the two-week categorization
Ops time consumed by it~19 hrs/day across 5 repsClient-confirmed — timed sampling during the audit
Automatable share of the mechanical work60%Conservative model — edge cases, angry customers, and unusual documents stay human
Loaded ops cost$58,000/yr per repBenchmark — ops/coordinator pay of roughly $45,000–$52,000 base (Salary.com, ZipRecruiter), loaded
Value of freed capacitymeasured in loads/rep, not layoffsModel — brokerages grow by covering more loads per rep; nobody in this model loses a job

The math: 60% of 19 hours is ~11.4 recovered hours a day — roughly 2,850 hours a year, worth about $79,000 in loaded cost. But the labor number understates it, and the model says so: a brokerage's growth limit is loads-per-rep. Freeing a third of the ops day raises that ceiling without a single new hire — which is why the sensitivity grid models capacity, not just cost:

If mechanical share is really…And automatable share is…Year-1 netVerdict
45% (lighter inbox than sampled)40%−$6,900Don't build — buy a TMS plugin and move on
72%40%+$11,200Marginal — automate status-only, defer the rest
72%60%+$41,300Build
72%70% + one added load/rep/week+$78,000+The growth case

Year-1 cost in the model: $37,700 ($16,100 build + $1,800/month operation). Payback at the confirmed baseline: about eleven months on labor alone; under six if the capacity case holds — and the dashboard in Phase 4 measures which one is true.

The first row, as always, is the audit earning its fee: a lighter inbox means the honest answer is an off-the-shelf TMS plugin, and the written verdict names which one.

Phase 3 — Build (weeks 4–12)

Limestone's document automation, and where work returns to peopleAn inbound email or document reaches an AI classifier that decides whether it is a status request, a quote, a document, or something else. Status requests go to a bot that reads the transport management system and replies with location and estimated arrival at any hour. Quote requests are extracted into a TMS record with rate history attached, but pricing stays with a human. Documents are identified, filed, and expired insurance is flagged by a deterministic rule. Everything else, plus anything the classifier scores low-confidence, goes to a human queue with full context attached.Inbound email or documentClassifierstatus? quote? document? other?[AI: reading unstructured email iswhat modern models are for]Status botreads the TMS tracking record, replieswith location and ETA — including 2 AM[AI drafts from live data — it canonly say what the TMS says]Quote intakeextracts lane, weight, dates; creates theTMS record; pulls rate history[AI extracts; pricing stays HUMAN —margin is judgment]Doc processoridentifies rate cons, PODs, insurance certs;files them; flags expired insurance HARD[extraction is AI; the expiry block isa deterministic rule]Human queuethe 28%, plus anything scoredlow-confidence, with full context[low confidence routes to people,loudly, rather than guessing quietly]
The three-part document system, and where work is handed back to people.

The house rules, third time, same rules: AI where it earns its place — reading emails and documents is the strongest legitimate use of these models that exists; judgment stays human — no model prices a load or waives an insurance flag; and low confidence routes to people, loudly, rather than guessing quietly.

Scope discipline note: we did not touch carrier sourcing, dispatch, or accounting. The inbox was the leak; the inbox is the build. That's why this is a $16,100 system and not a $100,000 "digital transformation."

Phase 4 — Measure (from week one)

  • Auto-resolved status requests: count and rate (target: >80% of status emails answered without a human, median reply under 2 minutes vs. the 47-minute baseline)
  • Quote-intake accuracy: extraction correctness sampled weekly — the honesty metric on the AI
  • Docs filed without touch: rate, plus every insurance flag audited (the compliance metric, zero tolerance)
  • Loads per rep per week: the growth number — the one the owner actually cares about, tracked against the pre-build baseline
  • The 90-day question: which sensitivity row is coming true, on the dashboard, in the client's own TMS numbers

What this maps to on our published pricing

StagePrice
Free scan — including the inbox categorization that starts everything$0
The audit — the labeled model and written verdict above$2,500, fully credited
The build — this scenario$16,100 (our automation builds start at $7,500 and typically land $10,000–$18,000; three integrated subsystems put this near the top of the band)
Operation — monitoring, extraction-accuracy review, model updatesfrom $1,800/month

Same exit as always: a don't build verdict, in writing, costs you the audit and saves you the build.

Questions logistics operators ask us about this

Does this work with our TMS?

The pattern integrates with any system that exposes its data — modern TMS platforms do, and even legacy ones usually have a path. Whether yours does is precisely what the audit's technical review answers before you commit a dollar to the build.

Is AI reliable enough for freight documents?

For reading and extracting — yes, demonstrably, and we measure the accuracy weekly and publish it to your dashboard. For decisions — pricing a load, accepting a carrier, waiving a flag — we deliberately don't use it. Extraction is AI work; judgment is your team's.

We're a warehouse / carrier / dispatcher, not a brokerage. Same idea?

Same idea. Anywhere in Indiana's logistics corridor — or anywhere in the country — that documents and status requests eat operations time, the model applies: categorize the inbox, price the mechanical share, automate the fraction that's safe, measure it.

Fictional client, real method. To run the inbox categorization on your actual operation: book the free scan. The method behind every number above is published in full; the companion engagements cover a manufacturer's quoting problem and an HVAC company's after-hours leak.

Related reading