Fourteen Seats, 1,400 Certificates a Year: When an Agency Should Host Kimi K3 Instead of Buying Licenses

You already priced this and you did not blink. Fourteen people, thirty dollars a seat, five thousand and change a year — less than the agency spends on the E&O premium for one producer. So why would an eleven-to-twenty-person agency ever consider running its own model?

Because the seat count and the document count are two completely different businesses, and only one of them has people in it.

Count your people. Now count your documents.

The people number is easy: a principal, three producers, two commercial account managers, three personal lines CSRs, a claims person, a benefits person, a bookkeeper, an office manager. Fourteen. Fourteen seats is a rounding error on the P&L.

Now count the paper, on an agency with roughly 3,000 personal lines policies and 260 commercial accounts:

• Around 1,400 certificates of insurance a year, most of them ACORD 25 with the same three general contractors asking for slightly different additional insured wording.
• Direct bill commission statements from 23 carriers, monthly, in 23 different layouts, reconciled by one bookkeeper against the agency management system.
• Policy download transactions through Ivans running into the tens of thousands a year, with a steady trickle that fails to match and lands in a suspense queue.
• Renewal document sets — dec pages, endorsements, cancellation notices — arriving continuously and needing to be read, filed and, occasionally, acted on within a deadline.

Per-seat pricing prices a person sitting at a keyboard. None of the work above is a person sitting at a keyboard. It is volume, it arrives on its own schedule, and it spikes hard in the fourth quarter.

What changed on the open side in 2026

Two years ago the free-to-run models were a hobby. They could summarise an email; they could not read a scanned commission statement without inventing a column. That stopped being true this year. Moonshot AI released Kimi K3, now the largest openly available model in the world, and the open tier as a whole closed most of the distance to the paid frontier products on ordinary document work — reading, extracting, matching, summarising.

At the same time the paid side got roughly ten times cheaper than it was in 2025. So the decision is not "open is cheap, paid is expensive." Both got cheap. The decision is about shape.

Running your own means the agency licenses nothing per person — it buys or rents one computer, installs a model that costs nothing to license, and pays for electricity, upkeep and one person's attention instead of a monthly fee that grows with headcount or usage.

flowchart TD
  A["Pick one job: certificates, commission statements, or download suspense"] --> B{"Is the work done by a person at a desk?"}
  B -->|Yes| C["Buy seats for those people"]
  B -->|No, it is document volume| D{"Monthly metered bill above $1,500?"}
  D -->|No| E["Stay on metered use, revisit in two quarters"]
  D -->|Yes| F{"Is the volume steady or only a Q4 spike?"}
  F -->|Only Q4| G["Rent extra capacity for the season"]
  F -->|Steady| H["Run your own on one machine, name an owner"]

Three volume shapes that break per-seat pricing

Certificates. A general contractor's compliance portal rejects an ACORD 25 because the additional insured wording does not match the subcontract. That kicks off a read-compare-reissue loop that has nothing to do with how many CSRs you employ. It has to do with how many certificate holders your commercial book touches, which for a contractor-heavy agency is thousands.

Commission reconciliation. Twenty-three carriers, twenty-three formats, once a month, plus the ones that pay by lump sum with a spreadsheet attached. Your bookkeeper is one seat. The job is 23 documents times 12 months times however many line items each carries.

The fourth-quarter spike. October through the first week of January, document volume in a commercial agency roughly doubles. Under per-seat you either buy seats you do not need for eleven months, or you buy temporary help and give them access to client files, which brings its own security and licensing questions.

The crossover, worked honestly — and it is higher than the pitch decks say

Here is the part most articles skip. At normal agency volume, running your own loses. Assumptions are illustrative; substitute your own document counts.

Read that table twice. The break-even sits somewhere around $1,500 a month of metered spend for steady, repetitive document work. Below it, buying seats and paying per use is the correct answer and anyone telling you otherwise is selling hardware. Above it, and only if the volume is steady rather than seasonal, owning starts to win — and it wins by a lot, because your bill stops growing when your document count grows.

What running your own actually costs in people

The line item nobody puts on the slide. Owning a machine that does production work means:

One named owner inside the agency. Usually the office manager or the operations person, not a producer. Their job is not technical: it is knowing when the thing is not running, knowing who to call, and making sure someone checks the output weekly. Call it two to four hours a month.

An IT provider on retainer. Patching, backups, monitoring, and a callback commitment. If your agency currently calls a guy when the printer breaks, you are not ready to own a machine that certificates depend on.

A written fallback. When the machine goes down on December 29 in the middle of 1/1 season, what happens? The answer has to be "we switch that job back to the paid service for a week," and that switch has to have been tested in July, not discovered in December.

Re-testing when the model changes. Open models improve fast. Every swap means re-running a sample of real certificates and commission statements and comparing against known-good answers. Budget half a day per change, and do not let it happen the week before renewals.

For a fourteen-person agency, that is realistically a tenth of one person plus a modest retainer. Which is fine at 40,000 documents a month and absurd at 6,500.

What you do not run yourself, at any volume

Anything carrying a licensed judgment. Coverage recommendations, the decision to bind, surplus lines diligent effort, a claim advocacy position — those are producer decisions with an E&O file attached, and where the words came from is irrelevant to whether a licensed person signed off.

Also: do not make your first project the self-hosted one. Run the job on a paid service for a quarter. Get the accuracy where you want it, get the checks written down, get staff used to reviewing rather than typing. Only then ask whether the monthly bill justifies moving it in-house. Agencies that start with the hardware end up with a very expensive box in a closet and a job nobody trusts.

And if your reason for owning is data control rather than cost — keeping payroll schedules and claimant names off third-party systems for 23 NYCRR 500 or your state's version of the NAIC data security law — say that out loud. It is a legitimate reason and it changes the maths, because compliance work has a cost too. Just do not confuse it with a savings argument.

Frequently asked questions

Can we run this on the server we already have in the back office?

Almost certainly not the one running your file shares and your phone system. Document reading at volume wants its own machine and its own graphics hardware, and sharing it with anything the office depends on is how you end up with both broken on the same afternoon. Price a separate box.

Is an open model good enough to read a carrier commission statement?

For structured, repetitive documents that your agency sees hundreds of times — the same 23 carrier layouts every month — yes, and it gets better the more examples of your own corrected output you keep. For one-off, messy documents like a hand-annotated subcontract insurance requirements exhibit, the paid frontier products are still noticeably better. Split the work accordingly; there is no rule saying you pick one.

Our staff are already using the paid tools on their own. Does that change the answer?

It changes what you should do first, which is write a policy and buy the seats officially. Staff using personal accounts to read client documents is a third-party disclosure problem under your own security program and a very awkward question on the cyber renewal application. Sort that out this month; the hosting question can wait a quarter.

Where a per-call service still beats owning anything

One category never belongs on your own hardware, and it is the phone. Answering a live caller means uptime you cannot promise from a closet: nights, weekends, the Monday after a hailstorm when the personal lines line rings 200 times. CallSphere builds AI voice and chat agents that answer the agency line and website chat around the clock, capture the caller's details, and book the callback with the right producer or CSR. Buy that as a service, own the document work if the volume justifies it, and keep the two decisions separate.