By Sagar Shankaran, Founder of CallSphere
Deloitte found 84% report positive AI ROI. The brokerage process to measure first, four baseline numbers to count before you start, and the weekly scoreboard.
Key takeaways
Eighty-four percent. That is the share of organisations investing in AI that report positive return, according to Deloitte's State of AI in the Enterprise 2026. It is also the number your operations manager will put in front of you this quarter, and the one to be suspicious of — not because it is wrong, but because it says nothing about whether your brokerage will be in it.
What is useful there is not the headline but the pattern underneath. The companies reporting a return did the same three things in the same order: they picked one messy process rather than a strategy, kept a human reviewing the output, and measured a baseline before switching anything on. The ones with nothing to show skipped the third step and then argued about feelings for a year.
So the question is which process. Not the most annoying one — the one you can already count today, where errors have a dollar figure, and where volume is high enough that ninety days produces a real sample.
In a freight brokerage, that process is carrier onboarding and vetting. It runs several hundred times a year, it is fully documented by regulation, it produces countable errors, and when it goes wrong it goes wrong in five figures.
Here is the discipline in one sentence, and it is worth quoting to whoever is proposing the project: you do not get to claim a return on an AI project unless you wrote down the baseline — the hours, the error count and the dollars — before it was switched on.
Think about how a carrier gets set up at 4:15 on a Friday. A carrier sales rep has a reload out of Joliet that must move tonight and finds a truck on DAT One. The carrier is new to you. The packet goes out through MyCarrierPackets or RMIS and comes back in eleven minutes with a W-9, a signed broker-carrier agreement and an ACORD 25 certificate of insurance. Somebody is supposed to check the FMCSA SAFER snapshot, confirm the operating authority is active and not three weeks old, confirm the insurance meets your minimums and names you correctly, check the safety scores, and confirm the phone and email on the packet match the FMCSA record rather than a lookalike.
At 4:15 on a Friday, with a load that must cover, some of that gets a glance instead of a check. Every broker knows it. That gap is where double brokering and strategic cargo theft walk in — a stolen identity of a legitimate carrier, a recently transferred authority, a dispatcher's phone number that is one digit off the registered one.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for logistics in your browser — 60 seconds, no signup.
The consequences have a hard floor: you carry a $75,000 BMC-84 bond, your customers' contracts put the cargo loss on you, and your insurer will ask what your vetting procedure was.
Spend one week counting these. Not estimating — counting, out of your TMS and your onboarding system.
flowchart TD
A["Rep finds a truck on DAT for tonight's reload"] --> B["Agent pulls SAFER snapshot, authority date, BOC-3, ACORD 25"]
B --> C["Agent calls the phone number on the FMCSA record"]
C --> D{"All seven checks clear?"}
D -->|Yes| E["Carrier activated, rate confirmation released"]
D -->|No| F["Held for the compliance specialist"]
E --> G["Weekly scoreboard: unverified dispatches this week"]
F --> G
Nothing exotic, which is the point. When a packet comes back it reads the ACORD 25, checks auto liability and cargo limits against your minimums, checks your company is named correctly as certificate holder, and checks the expiry dates. It pulls the SAFER snapshot and reads authority status, grant date, whether authority was recently reinstated or transferred, and the safety scores. It checks the address and notices when it resolves to a mailbox store. It compares the phone and email against the FMCSA record and flags a mismatch. It checks the name against your do-not-use list and your claims history.
Then it does one thing your people rarely have time for on a Friday: it places a short verification call to the number on the FMCSA record, not the one on the packet, and confirms the dispatcher and the truck. The whole set takes ninety seconds. A compliance specialist takes twenty minutes on a good day and four on a bad one.
Everything it flags goes to a person. That is not a limitation, it is the design — the pattern behind the 84 percent is human review kept in place, not removed.
Assumptions, all replaceable with yours: 900 new carriers onboarded a year, 12,000 loads a year, compliance staff loaded cost $28 an hour.
| Measure | Baseline, prior 12 months | 90-day pilot, annualised |
|---|---|---|
| New carriers onboarded | 900 | 900 |
| Median minutes, packet to active | 41 | 9 |
| Packets approved after 6 p.m. with checks skipped | 22% | 3% |
| Loads dispatched with unverified insurance or authority | 148 | 6 |
| Active carriers with an expired certificate today | 7.4% | 1.1% |
| Confirmed double-brokered or stolen loads | 6 | 1 |
The labour saving is easy: 900 carriers × 32 minutes saved = 480 hours a year × $28 = $13,440. Real, but not the argument.
The exposure is the argument, and it has to be modelled honestly. Suppose historically 4 percent of loads dispatched without verification ended in a costly incident, at an average net cost after insurance of $18,500. Baseline: 148 × 4% = 5.9 incidents, about $109,000. Pilot: 6 × 4% = 0.24 incidents, about $4,400. Modelled avoidance: roughly $105,000 a year. Say the word "modelled" out loud when you present it, because six incidents a year is far too few to prove anything in ninety days, and anyone who tells you otherwise is selling.
Rare, expensive events cannot be measured in a quarter. Leading indicators can. So the single number to put on the wall is this: loads dispatched this week to a carrier that had not cleared all seven checks. Count it every Friday. The target is zero.
Still reading? Stop comparing — try CallSphere live.
See the logistics AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
It works for three reasons. It is countable from your TMS without anyone's opinion. It moves within days rather than quarters. And it sits upstream of every cargo claim you have ever paid, so improving it is not a proxy for what you care about — it is a cause of it.
Pair it with two supporting counts: median minutes to active, and share of active carriers with an expired certificate. If those three move in ninety days and nothing else in your process changed, you have the closest thing to proof that a business of your size is ever going to get. If they do not move, stop. That is what a pilot is for.
The judgement calls stay human, and there are more of them than a vendor will admit. A carrier whose authority is four months old is not automatically a fraud — plenty of good owner-operators started this year, and if you refuse everyone under six months you will not cover Laredo in produce season. Somebody has to weigh that, and it is a person.
Insurance edge cases stay human too: a certificate with a reefer breakdown exclusion, a cargo limit fine for consumer goods and nowhere near enough for the electronics load you are about to tender, a policy that expires mid-transit. So does anything with a claims history at your own company. And the call that decides a close case — to the insurance agent listed on the certificate, not the carrier — is still a person's job, because the point is to hear how the answer sounds.
One more limit worth stating: this reduces the odds, it does not eliminate them. Sophisticated identity theft of a legitimate carrier can pass every document check, because the documents are real. Verification calls to the registered number, tracking that matches the tendered truck, and a rep who notices that the driver's story does not match the paperwork are what catch those.
Ninety days, and set the decision rule in writing on day one: which three numbers must move, by how much, and what happens if they do not. Deciding the rule afterwards is how people talk themselves into a renewal.
It probably will be. Every brokerage that counts unverified dispatches for the first time finds a number nobody wanted to see. That is the value of counting it — and it is much better found by you in a spreadsheet than by your insurer during a claim investigation.
No. Those are sources. The work being handed over is the assembling, comparing and chasing across all of them plus your own TMS and claims history, which is what actually takes the twenty minutes and what gets skipped at 4:15 on a Friday.
They are right that anything adding friction at 4:15 will be resisted, which is why median minutes sits on the scoreboard next to the compliance number. If setup time goes up, the pilot has failed on its own terms. Ninety seconds of checks against forty-one minutes of waiting is an argument you win with data instead of authority.
Onboarding generates phone traffic all day — carriers chasing setup status, dispatchers calling to check whether a packet came through, drivers ringing after hours about a load they were told to take. CallSphere builds AI voice and chat agents that answer those calls and web chats 24/7, capture the MC number and callback details, and route or book what needs a person. It handles the conversation, not the vetting decision — that stays with your compliance desk, where the scoreboard says it belongs.
Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
Map one messy process and prove it: spray ticket records, the five baseline numbers to capture before you start, and the error rate that ends the debate.
The past-due report is the gym process to baseline before buying AI: five numbers to capture, a worked example on a 1,400-member club, and the honest limits.
Carrier shortlists are built by software before dawn. What makes a 38-truck fleet invisible: stale MCS-150, no lane page, and a phone that stops at 6pm.
Deloitte found 84% of AI investors report positive returns. For P&C carriers the provable process is FNOL intake - here are the four numbers to baseline first.
How a security guard company proves AI paid for itself: map the open-shift callout, take a 30-day baseline, and track overtime as a share of billed hours.
Pull twelve months of factor deductions, convert to chargeback dollars per $100,000 shipped, and you have a baseline that settles the AI argument in 90 days.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI