AI for Freight Brokers: The 2026 Operator's Guide

2026-07-29 · Tommaso Maria Ricci

AI for freight brokers: the margin math nobody wants to look at

Here is the number that should end every debate about whether AI for freight brokers is worth the trouble. Gross margins in truckload brokerage sit around 15 percent of the freight bill, and once you subtract salaries, technology, insurance, and bad debt, most mid-sized brokerages operate at 85 to 95 percent of net revenue. That means for every dollar of net revenue you keep, 85 to 95 cents walks straight back out. Your entire business is a spread, and the spread is thin. In a business like that, a 10 percent reduction in cost per load does not improve your operation, it changes what your operation is worth.

I write this as a founder, not as a theoretical consultant. I have spent twenty years taking real companies from point A to point B using process, automation, and in recent years applied AI systems. I do not sell a transportation management system and I have no software to place with you. What I have is a method for working out where a technology investment actually moves a number in the profit and loss statement, and where it just produces another dashboard nobody opens. This article applies that method to freight brokerage, including the parts that vendors do not put on the slide.

Why brokerage is the most automatable business in freight

Freight brokerage has a structural feature that almost no other business has: it is entirely made of information work. You do not own trucks. You do not touch cargo. You do not run warehouses. What you own is a coordination function between shippers who have loads and carriers who have capacity, and every single step of that function is data moving between people.

Think about what a brokerage day actually consists of. Reading a load tender from an email or a portal. Pricing it against a lane you may or may not have run recently. Finding a carrier. Checking whether that carrier is legitimate and insured. Negotiating a rate over the phone. Building the load in the system. Sending a rate confirmation. Calling for a pickup update. Calling for a delivery update. Updating the shipper. Collecting the paperwork. Approving the invoice. Chasing payment.

Count how many of those steps involve producing new value versus moving information between two parties who both already have it. In most brokerages the ratio is brutal. The average broker spends the majority of the day on communication and data entry, and a small minority on the two things that actually generate profit: winning shippers and building carrier relationships that give you capacity when the market turns.

This is what makes brokerage different from asset-based trucking. A carrier that automates its back office still has to move a truck down a highway, and the truck is the constraint. A brokerage that automates its back office has removed the constraint entirely, because the constraint was always human throughput. That is why the operators who get this right do not just cut cost, they raise loads per employee, which is the only metric in this industry that compounds. The same principle applies across the sector, and I have laid it out more broadly in the guide to AI for logistics companies.

Load matching and carrier selection: the end of the load board scroll

Start with the core activity. A load comes in, and somebody has to cover it. The traditional process is a combination of memory, a carrier database, and a load board, executed by a person who is simultaneously handling six other loads.

The problems with that process are well known inside the industry and rarely stated plainly:

It is biased toward recency, not fit. Brokers call the carriers they called last week, because those are the ones they remember.
It ignores most of the carrier network. Every brokerage has thousands of carriers in the system and works with a few hundred of them. The rest are dead weight in a database.
It optimizes for speed of coverage, not margin. Under pressure, a broker takes the first acceptable rate, not the best available one.
It loses the knowledge when the person leaves. Carrier relationships live in a rep's head and phone, which is why broker turnover is so expensive.

A well-built matching system attacks all four. It ranks carriers for a specific load based on things a human cannot hold in memory: lanes that carrier has actually run for you, their historical acceptance rate at different rate levels, their on-time performance on similar lanes, equipment type, current known position based on recent activity, and even the time of day they typically accept.

The output is not a magic answer. It is a shortlist of eight carriers ranked by probability of acceptance at a target rate, instead of forty phone calls in descending order of desperation. The measurable effect is on two numbers: time to cover, and margin retained on covered loads. Both are trackable from day one, which is exactly what makes this the right place to start.

The backhaul and continuous move problem

There is a second-order use case here that most brokerages never touch because it requires holding too much in mind at once. When you cover a load from Atlanta to Dallas, that carrier is now going to be empty in Dallas. If you have a load out of Dallas in the next 48 hours, you have leverage: you can offer the carrier a two-load package and capture better economics on both.

Doing this manually is nearly impossible at scale, because it requires cross-referencing every carrier's future position against your entire open board in real time. Doing it algorithmically is straightforward. Brokerages that build this capability do not just improve margin, they become materially more attractive to carriers, which is the currency that matters when capacity tightens.

Pricing and quoting: where brokers guess and lose

Ask a broker how they priced a lane and you will get some combination of the market rate index, what they got last time, and instinct. All three are legitimate inputs. None of them is a pricing strategy.

The structural problem in brokerage pricing is that quotes are made under time pressure with incomplete information, and the cost of being wrong is asymmetric. Quote too high and you lose the load, which costs you nothing visible, so nobody counts it. Quote too low and you cover it at a loss or scramble, which is painful but at least you know it happened. That asymmetry pushes the whole industry toward quoting low, because losses are visible and missed opportunities are not.

An AI pricing layer changes the inputs available at the moment of the quote:

1. Your own historical performance on that specific lane, including what you actually paid carriers, not what the index said the market was. 2. Current tightness signals: how many carriers accepted at what rate in that region in the last few days, how long comparable loads took to cover. 3. Seasonality and known demand events for that origin and commodity type. 4. Shipper-specific behavior: which accounts accept quotes above index, which ones only ever take the lowest bid, and what your realized margin on that account has actually been over twelve months. 5. Win probability at each price point, which is the input almost nobody has and the one that changes decisions.

That last item is the whole game. A broker who knows that quoting 2,450 dollars gives a 70 percent win probability at 14 percent margin, while 2,600 gives a 40 percent win probability at 19 percent margin, is making a business decision. A broker without that information is guessing and calling it experience.

The number to track is realized margin per load by account, compared against quoted margin. In most brokerages the gap between the two is significant and unexamined, and closing part of it is the single fastest path to profit improvement. I have covered the general framework for quantifying this kind of intervention in the guide on AI ROI for business.

The case: 30 percent sales growth from applied AI

A concrete example from outside freight, because the mechanism transfers directly. With WSB Sport we increased sales by 30 percent using AI applied to marketing and contact management. Not through one brilliant campaign, but by building a system that captured demand, responded fast, and never let a contact go cold.

For a brokerage the equivalent is response time on inbound quote requests. Shippers, especially mid-market ones, frequently award the load to whoever responds first with a credible number. If your quote turnaround is four hours because a person has to research the lane, you are structurally losing loads to brokers who answer in eight minutes. That is not a sales skill problem, it is a process latency problem, and process latency is exactly what automation removes.

Track and trace: killing the check call

Every brokerage runs on check calls. Where is the truck. Did it load. What is the ETA. It is the single largest consumer of operational hours in the industry and it produces zero value: the information already exists, it is just trapped in a driver's head or in a telematics system nobody has connected.

The automation path here is mature and unusually clean:

Automated location capture through carrier telematics integrations, tracking apps, or, at minimum, structured text-based updates from drivers.
Exception-based management, meaning the system flags only the loads that are actually deviating from plan, instead of a dispatcher checking all 200.
Predictive ETA, which uses transit history, traffic, and hours-of-service constraints to say the load will be late four hours before the driver knows it will be late.
Automatic shipper notification on those exceptions, which converts a bad-news phone call into a proactive update and materially changes how the account perceives you.

The last point is worth pausing on. In brokerage, service quality is mostly perceived through communication. Two brokers can have identical late-delivery rates, and the one that flags problems early keeps the account while the other loses it. Exception management is not just a cost saving, it is a retention strategy.

The number to watch is check calls per load and the percentage of loads with automated tracking. Both are easy to baseline and both move fast. The same operational logic applies to the carrier side of the market, which I have covered in the guide to AI for trucking companies.

Fraud and carrier vetting: the existential problem

This is the section that matters most in the current market, and the one most brokerages are underinvested in relative to the actual risk.

Cargo theft has shifted from a physical crime to an information crime. According to CargoNet's 2025 theft trend analysis, estimated losses climbed to roughly 725 million dollars, up around 60 percent from the prior year, with the average value per theft rising to approximately 274,000 dollars. The volume of incidents stayed relatively stable while the value per incident jumped, which tells you exactly what changed: this is no longer opportunistic theft from truck stops, it is organized groups selecting high-value shipments and obtaining them through fraud.

Double brokering is the primary mechanism. A criminal obtains the rights to a shipment using a stolen or fabricated carrier identity, re-brokers it to a legitimate carrier who has no idea they are participating in a crime, and the freight is diverted before it reaches the consignee. Industry estimates of annual losses from double brokering alone run into the hundreds of millions of dollars.

The cost lands disproportionately on intermediaries. ATRI research on the cost of cargo theft found that motor carriers average more than 520,000 dollars in annual theft losses, while logistics service providers, the category that includes brokers, average more than 1.84 million dollars, with an annualized cost to the industry as high as 6.6 billion dollars. Read that gap again: the party that never touches the freight carries more than three times the average loss of the party that hauls it.

For a broker this is not just a cargo claim. It is a liability question, an insurance question, and frequently an account-ending event. And it is happening at a scale that manual vetting was never designed to handle.

Where AI genuinely helps on fraud

Fraud detection is a pattern recognition problem, which is the thing machine learning is unambiguously good at. The practical applications:

Identity consistency checks at onboarding: matching the carrier's authority, insurance certificate, contact details, banking information, and physical address against each other and against known-fraud databases. Fraudulent setups almost always contain internal inconsistencies that a rushed human misses.
Behavioral anomaly detection: a carrier that has run flatbed in the Southeast for two years suddenly bidding on high-value electronics in California is a signal, not a coincidence.
Communication pattern analysis: newly registered domains, email addresses that differ by one character from a legitimate carrier's, phone numbers that appear across multiple unrelated carrier profiles.
Banking change monitoring: a request to change remittance details mid-load is one of the highest-signal fraud indicators that exists, and it is trivially automatable as a hard stop.
Cross-load correlation: the same equipment number or driver appearing on loads for two different carriers at overlapping times.

None of this eliminates fraud. Criminal operations adapt, and any broker who tells you their system stopped the problem has not been targeted yet. What it does is raise the cost of attacking you relative to attacking the brokerage down the street, and it catches the majority of attempts that rely on volume rather than sophistication.

What AI cannot do here

Be clear about the limit. Fraud prevention is fundamentally a policy problem with a technology assist. If your team is authorized to override a flag because the load has to move today and the customer is on the phone, the system will be overridden, and it will be overridden precisely on the loads where the pressure is highest, which are the loads criminals target.

The controls that actually work are procedural: no banking changes without out-of-band verification, no first-time carrier on high-value freight without a secondary approval, mandatory hold on any load where the system raises a specific class of flag. AI makes those controls cheap to enforce. It does not make the decision to enforce them.

Back office: invoicing, settlement, and claims

The back office is where brokerages quietly lose margin they already earned. The work is document-heavy, exception-heavy, and almost entirely manual in most operations: matching bills of lading and proof of delivery against loads, reconciling carrier invoices with rate confirmations, handling accessorial disputes, chasing missing paperwork, processing quick pay, managing claims.

The automation is straightforward and the return is fast:

1. Document capture and extraction: rate confirmations, bills of lading, proof of delivery, and carrier invoices are read automatically and the data lands in the system without transcription. 2. Three-way matching between the rate confirmation, the delivered load, and the carrier invoice, with only mismatches escalated to a human. 3. Accessorial validation against the contracted schedule, which catches detention and layover charges that should have been rejected and routinely are not. 4. Automated document chasing for missing paperwork, which is the primary cause of delayed invoicing to shippers and therefore of stretched working capital. 5. Claims triage with automatic assembly of the supporting file, which shortens the cycle on the disputes that get resolved and speeds up the decision to concede the ones that will not.

The financial impact runs through days sales outstanding. Brokerage is a working-capital business: you typically pay carriers faster than shippers pay you, so every day of delay in invoicing is a day of financed float. Cutting invoicing lag by a few days on meaningful volume frees real cash, permanently, without selling a single additional load. The broader pattern behind this kind of process work is in the guide to AI workflow automation for business.

Sales: shipper prospecting and lane intelligence

The commercial side is where most brokerages are weakest, and it is not for lack of effort. It is because brokerage sales is traditionally a volume activity: call lists, cold outreach, persistence. It works, badly, at enormous cost per acquisition.

AI changes the targeting rather than the effort. The useful applications:

Lane-fit prospecting: instead of calling every manufacturer in a metro area, identify shippers whose likely freight profile matches lanes where you already have reliable, well-priced capacity. Selling into your strength is a fundamentally different conversation from selling into an empty network.
Signal-based timing: new facility openings, distribution center announcements, expansion news, hiring patterns in logistics roles. These are public and they indicate when a shipper's freight needs are about to change, which is the only moment an incumbent broker is genuinely vulnerable.
Existing account expansion: analysis of which lanes you already serve for a customer versus which lanes they almost certainly run based on their facility footprint. This is the cheapest revenue in the entire business and almost nobody works it systematically.
Churn signals: an account whose volume is drifting down, whose tender acceptance is falling, or whose load mix is changing is telling you something months before they tell you directly.

The metric here is not activity, it is revenue per account and win rate by segment. A brokerage that raises win rate from 5 percent to 8 percent on targeted outreach has effectively cut its cost of acquisition by nearly 40 percent without hiring anyone. The general playbook for this kind of commercial system is in the guide to AI for sales.

The case: from 9 to 10 million without adding structure

A useful parallel from a different industry. I worked with a hospitality operation running roughly 9 million in revenue that had the classic problem of perishable capacity, demand estimated by feel, and pricing set by habit. By working on demand forecasting and commercial response speed, revenue passed 10 million without adding rooms, staff, or invested capital. A million in additional revenue built on process, not investment.

For a brokerage the translation is precise. Your capacity is not trucks, it is the number of loads your people can handle well. If your team covers 300 loads a week and the same team could cover 380 with the manual work removed, you just created revenue capacity at unchanged fixed cost. That is the difference between growing with margin and growing by hiring, which in this industry is the difference between a business worth selling and a business that merely gets bigger.

What AI will not fix in your brokerage

I would rather lose your interest here than have you spend money badly, so here is the honest counterweight.

It will not fix a bad book of business. If your account portfolio is concentrated in low-margin, high-service accounts that only buy on price, better software makes you slightly more efficient at losing money. Portfolio quality is a commercial decision, not a technology one.

It will not replace carrier relationships. In a tight market, capacity goes to the brokers that carriers want to work with: the ones who pay fast, load accurately, and do not waste a driver's day. No algorithm generates that goodwill, and the brokerages that treated carriers as interchangeable during soft markets discovered exactly what that cost them when the market turned.

It will not survive dirty data. Brokerage systems are typically full of duplicate carrier records, inconsistent lane naming, accessorials entered as free text, and equipment types that mean different things to different reps. Every model built on that produces confident nonsense. Data cleanup is the unglamorous prerequisite and there is no way around it.

It will not compensate for having no differentiation. If a shipper cannot articulate why they use you rather than the three other brokers quoting the same lane, efficiency gains just let you race to the bottom faster.

Self-assessment: how ready is your brokerage

Before spending a dollar, measure. Score yourself 0 to 2 on each of these six questions, then add it up. This is the same diagnostic I run at the start of any project, because without knowing where you start you cannot know what you can gain.

1. Data quality. What state is your carrier and lane data in? - 0: duplicates everywhere, free-text fields, nobody owns it. - 1: reasonably clean in the core system, messy at the edges. - 2: deduplicated, standardized, with at least two years of reliable history.

2. Pricing. How do you set a quote? - 0: index plus instinct, no record of quotes that were lost. - 1: index plus internal history, tracked inconsistently. - 2: structured pricing with quoted versus realized margin measured by account.

3. Tracking. How do you know where a load is? - 0: check calls, all day, on every load. - 1: some automated tracking, mostly on larger accounts. - 2: majority automated with exception-based management.

4. Carrier vetting. What happens when a new carrier signs up? - 0: authority and insurance check, and we move on. - 1: a documented checklist executed manually. - 2: automated multi-source verification with hard stops on defined risk flags.

5. Back office. How does an invoice get processed? - 0: manual matching, paper chasing, frequent disputes. - 1: partly automated with heavy manual exception handling. - 2: automated capture and matching, humans only on exceptions.

6. Commercial. How do you decide who to prospect? - 0: lists and volume, disconnected from our actual network strength. - 1: some targeting, mostly based on rep preference. - 2: prospecting driven by lane fit and demand signals, with win rate tracked.

Reading your score

0 to 4: manual operation. You are leaving margin on the table on every load, but the upside is that early interventions produce fast, obvious returns. Start with data cleanup and back office automation. Do not touch predictive pricing yet, it will not work on your data.
5 to 8: in transition. You have process discipline but it depends on individuals. This is where AI converts episodic competence into a system. The best return comes from pricing intelligence and automated tracking.
9 to 12: mature. Your foundations are solid. AI does not rescue you, it multiplies you: predictive matching, win-probability pricing, and advanced fraud detection are where you gain a durable edge that competitors cannot copy quickly.

Whatever the score, the number itself is not the point. The point is knowing which process to attack first based on expected return and your cash position. That is exactly the conversation worth having with someone who has already taken companies from where you are to the next level, rather than paying tuition to find out which sequence works.

Roadmap: the first 30, 60, and 90 days

Failed transformations share one trait: they start with everything at once. The method that works is the opposite. One intervention at a time, each tied to a number.

First 30 days: baseline and cleanup

Establish the baseline: loads per employee per week, average time to cover, quoted versus realized margin by account, check calls per load, days sales outstanding, percentage of invoices requiring manual intervention. Without a baseline there is no return on investment, only opinion.
Clean the carrier and lane data. Deduplicate carriers, standardize lane definitions, structure accessorial codes. It is the least interesting work in this entire article and it determines whether everything after it functions.
Automate document capture on the back office, which is the fastest return with the lowest risk and no operational disruption.
Month goal: cut invoice processing time in half and have numbers you actually trust.

Days 31 to 60: pricing and tracking

Deploy tracking automation on your highest-volume accounts first, and convert dispatch from all-load monitoring to exception management.
Build the pricing baseline: analyze twelve months of quotes against wins and realized margins, by lane and by account. Most brokerages find at least one account that has been unprofitable for a year.
Introduce win probability into the quoting workflow, initially as guidance rather than as a control.
Goal: measurable movement in realized margin per load and a real reduction in check call volume.

Days 61 to 90: matching, fraud, and consolidation

Deploy carrier matching on a defined subset of lanes, with margin tracked against a control group so you can prove the effect rather than assert it.
Harden fraud controls: automated onboarding verification, banking change hard stops, high-value load approval rules.
Work the existing book with lane expansion analysis on your top 50 accounts.
Compare against the baseline and decide what to extend, fix, or kill.

The guiding principle is that each phase closes with a number that moved, not with a tool that was adopted. If at day 90 you cannot say which metric changed and by how much, the project failed regardless of how good the interface looks.

The mistakes that cost the most

After twenty years building and fixing companies, the AI mistakes are almost always the same ones. Listing them saves months and a lot of money.

Buying technology before defining the number. If you cannot say which metric must improve and by how much, you are buying reassurance, not results.
Starting with the most impressive capability. Predictive pricing demos beautifully and fails immediately on dirty data. Boring work first.
Automating a broken process. If your quoting workflow is wrong, automating it means being wrong faster and at higher volume.
Ignoring the reps. If brokers believe the system exists to measure them rather than to help them, they will work around it, and they will be very good at it. Explain what it does for their paycheck, not how advanced it is.
Treating fraud tooling as a substitute for policy. Systems flag, humans decide. If overrides are unrestricted, the flags are decorative.
Underestimating change management with carriers. Automated communication that feels robotic to a driver damages exactly the relationships that determine whether you get capacity in a tight market.
Not measuring. Without a baseline and a comparison, every AI investment becomes an act of faith. Acts of faith do not show up in the profit and loss statement.

What it costs and when it pays back

The question I always get is what it costs. The honest answer is that it depends on what you attack and where you start, but the right way to reason about it is against the size of the lines you are moving.

Run the arithmetic on a brokerage doing 50 million in gross revenue at a 15 percent gross margin, so 7.5 million in net revenue:

A single point of realized margin recovered, through better pricing discipline and eliminating unprofitable accounts, is worth 500,000 dollars in gross margin, and most of it falls to the bottom line because your cost base does not change.
Back office automation typically removes the manual handling on the large majority of clean invoices, which frees the equivalent of one to three full-time positions that you redeploy into sales rather than replace.
Days sales outstanding. Cutting five days of invoicing lag on 50 million in revenue frees roughly 685,000 dollars of working capital, permanently.
Loads per employee. If exception-based tracking and automated documentation raise throughput by 20 percent, you absorb a year of growth without adding headcount, which in a business with 85 to 95 percent operating ratios is the difference between profitable growth and expensive growth.
One prevented fraud incident. With average theft values around 274,000 dollars, a single avoided event can pay for the entire program.

These are conservative figures and none of them requires selling more freight. Against returns of this order, the investment in systems and process reorganization typically pays back in months rather than years, because the lines involved are enormous relative to the margin.

The hard part is not finding the money. It is choosing the right first intervention, the one that pays back fastest and funds the next. Getting the sequence wrong is the surest way to turn AI into a cost rather than an investment, and in a business where working capital is everything, that is an expensive mistake to make twice.

Where this ends up: the structural bet

I will close with the uncomfortable strategic read, because it determines how urgently you should act.

Freight brokerage exists because matching supply and demand in trucking is hard, fragmented, and information-intensive. That has been a durable moat for decades: the market has hundreds of thousands of carriers, most of them very small, and coordinating them requires human effort at scale. The digital brokers who promised to eliminate the human entirely learned that the last mile of this problem is relationship work, not software.

But the moat is not made of the parts that are being automated. The moat is carrier relationships, shipper trust, and pricing judgment on complicated freight. Everything else, which is to say the majority of what a brokerage does today by headcount, is coordination overhead that is becoming steadily cheaper.

The strategic implication is straightforward. Cost per load is going to fall across the industry, and competitive pressure will pass most of that saving through to shippers. Brokers who automate will end up with roughly the same margins they have now on materially more volume per person. Brokers who do not will find themselves with the same cost base competing against operators whose economics are structurally better. This is not a technology adoption decision, it is a decision about what your cost structure looks like in three years relative to the people you bid against. The broader version of that argument, applied across industries, is in the guide on why every CEO needs an AI strategy.

FAQ

What does AI for freight brokers actually do day to day?

It works in six areas: matching loads to the carriers most likely to accept at a target rate, pricing quotes with win probability rather than instinct, automating track and trace so dispatch manages exceptions instead of making check calls, detecting fraud and double brokering during carrier onboarding and load execution, automating document capture and invoice matching in the back office, and targeting shipper prospecting based on lane fit and demand signals. The highest-return starting point is usually not the most impressive one: document automation and data cleanup produce results within weeks and create the foundation everything else depends on.

Will AI replace freight brokers?

No, and any vendor claiming otherwise has not covered a load in a tight market. What gets automated is coordination overhead: check calls, data entry, document chasing, invoice matching, initial carrier screening. What does not get automated is negotiating with a carrier who has three other options, keeping a shipper through a service failure, and pricing complicated freight where history is thin. The realistic outcome is fewer people doing coordination and the same or more people doing commercial work, with loads per employee rising significantly. Brokerages that use this to expand rather than to cut generally end up ahead.

How does AI help with double brokering and freight fraud?

It helps in two specific ways. At onboarding it cross-checks authority, insurance, banking details, addresses, contact information, and domain history against each other and against known-fraud data, catching the internal inconsistencies that a rushed human misses. During execution it flags behavioral anomalies: mid-load banking change requests, carriers bidding far outside their historical lane and equipment profile, duplicate equipment or driver identifiers across unrelated carriers. Given that cargo theft losses reached an estimated 725 million dollars in 2025 with average incident values near 274,000 dollars, this is now a financial control, not an IT project. It does not eliminate fraud, and it only works if your team is not permitted to override flags under commercial pressure.

How long before we see measurable results?

Back office automation shows results within the first month, because it removes manual hours immediately. Tracking automation and check call reduction typically show up within 60 days. Pricing intelligence takes longer, usually 60 to 90 days, because it needs a clean history of quotes, wins, losses, and realized margins to be useful. Predictive carrier matching needs the most data and the most patience. The factor that stretches timelines is almost never the technology, it is the state of your carrier and lane data at the start.

Do we need to replace our transportation management system?

Usually not, and replacing it first is a common and expensive mistake. Most of the value in the early phases comes from layers that sit alongside the existing system: document processing, data cleanup, pricing analysis, tracking integration. A full system replacement is a twelve to eighteen month project that consumes the organizational attention you need for the actual improvements. The sensible order is to fix the data, add capability around the system, prove the returns, and only then evaluate whether the underlying platform is genuinely the constraint.

Is this only for large brokerages?

No. The economics arguably favor smaller operations, because a brokerage with fifteen people has more of its total cost tied up in coordination work than one with a dedicated operations organization. Document automation and data cleanup are accessible at almost any size and do not require a technology team. What smaller brokerages should avoid is starting with predictive pricing or matching, which need data volume to work and will produce unreliable output on a thin history. Start with the manual work, build the data asset, then move to the models.

What is the single biggest mistake brokerages make with AI?

Automating a broken process. If your quoting workflow produces unprofitable quotes, automating it means producing unprofitable quotes faster and in greater volume. The order that works is: measure what you actually do today, fix the process logic, clean the data it depends on, and only then automate. The second biggest mistake is buying tools before defining which number has to move, which guarantees you cannot tell afterward whether the investment worked.