A 58-Person Credit Union Pays Per Seat for AI Its Tellers Touch Twice a Week. Kimi K3 Made Owning the Box Cheaper.

Open the general ledger and find the line your CFO created last spring called "AI software." At a 58-employee, $240 million credit union it started around two hundred dollars a month — the marketing coordinator and the CEO. Then compliance wanted it. Then member solutions. Then the three branch managers asked why they were the only ones without it. By the time the June board packet printed, that line was a five-figure annual commitment billed per person per month, for a tool most of those people open twice a week for eleven minutes.

Nobody is being wasteful. The pricing simply does not match the shape of a credit union — and as of 2026, it no longer has to.

Per-seat pricing was built for a company where everyone types all day

Look at your org chart. Of 58 people, maybe nine are member service representatives on the phone queue, fourteen are tellers and member advocates in branches, six are in lending support, three are in member solutions chasing 30-day delinquencies, one is your BSA officer, one runs IT and the core, and the rest are accounting, managers and executives. Your operating expense to average assets ratio is the number your examiner circles and your board argues about in November. Every recurring per-head cost lands directly in it.

Per-seat software assumes each seat gets roughly equal value from the tool. True at a software company. Not true at a credit union, where the work is spiky and concentrated: the phone queue peaks between 11:30 and 1:30 and again on the Fridays when share drafts post, tax-refund season buries the loan desk from late January through March, and December brings the skip-a-pay flood. Your usage is not 58 people all day. It is a few people doing a lot of repetitive lookups in a few concentrated hours.

An open-weight model is one whose maker publishes the finished model itself, so you can run it on hardware you own — a server in your own building or at your data processing partner — instead of renting access to it per employee per month. That distinction used to be academic, because the free versions were noticeably worse. In 2026 it stopped being academic.

What Moonshot AI shipped, and why a $240M shop should care

Moonshot AI released Kimi K3, now the largest open model in the world at 2.8 trillion in size. It is built as a "mixture of experts," which in plain terms means only a fraction of it wakes up for any one question — so it runs on far less hardware than that number suggests. More important than K3 specifically: the open tier as a whole closed most of the gap with the paid frontier tools during the first half of 2026. For the work a credit union actually does with these tools — read this policy and tell me what it says, summarize this member's call history, explain this Reg CC hold in language a member will accept — the open models are now good enough that the difference does not show up in your member's experience.

So the question is no longer "is the free one usable." It is a purchasing question you already know how to answer, because you answer it every time you weigh buying an ITM against keeping a branch lease: rent per head, or buy the asset and depreciate it.

flowchart TD
  A["Member calls about a returned share draft"] --> B["MSR types the question in plain English"]
  B --> C{"Does the answer need account data?"}
  C -->|No| D["In-house model answers from your policy manual"]
  C -->|Yes| E["MSR pulls the record in the core and pastes the detail"]
  E --> D
  D --> F["MSR reads it back and notes the call in the core"]
  F --> G["VP of Operations reviews flagged answers each Friday"]

Tuesday, 11:40 a.m., in the member solutions center

A member calls about a $32 courtesy pay fee on a share draft that cleared two days after a direct deposit she thought had landed. The MSR is 40 seconds into the call and does not remember whether your board-approved overdraft policy allows a second courtesy waiver inside twelve months, or whether the December policy update changed the waiver limit from two to one.

Today she puts the member on hold, opens the shared drive, finds a PDF named OD_Policy_FINAL_v3_rev2.pdf, and skims. Ninety seconds, sometimes three minutes, and about one time in eight she is reading a superseded version. With a model running on your own hardware, she types the question, gets the answer with the policy section quoted back, and the member never hears hold music. No per-seat license was consumed, because there are no seats — the box does not care whether nine people or twenty-nine ask it something today.

That is the whole argument. Per-seat cost grows with headcount. Owned cost grows with how hard you work the box. Credit unions have large frontline headcount and low per-person usage, which is precisely the shape where owning wins.

The arithmetic, with the assumptions written down

Assumptions, all illustrative: 58 employees, 42 of whom would realistically be licensed; a per-seat price of $32 per person per month; a single server capable of running a strong open model for a shop this size at $34,000, depreciated over three years; 1.2 kW average draw at $0.16 per kWh; your core administrator's loaded cost at $58 an hour.

Year one is a wash. Year four is not close. And the divergence gets worse for the rent option every time you grow: absorb a small credit union with twelve staff and per-seat adds $4,608 a year, while the owned box adds nothing until you actually saturate it. Run the same table with 22 licensed users instead of 42 and renting wins outright — which is the honest answer for a credit union under about $80 million in assets.

The people cost nobody puts in the board packet

Running your own model is not free labor. Somebody has to patch the machine, restore it when it falls over on a Saturday, and keep a written record of all of it for your examiner. At a shop your size, that somebody is the person who already administers your core — and that person is already the single point of failure for your whole operation.

You have three realistic routes. One: your existing core administrator adds it, budget six to eight hours a month and accept that you now have a second thing only one person understands. Two: your CUSO or data processing partner hosts it for several credit unions and bills you a flat monthly fee — usually the right answer, because it converts a staffing risk into a contract. Three: your state league or a shared-services group runs it collectively. If none of those three is available to you, do not buy the box.

Where renting per seat is still the right call

If your usage is genuinely light — a handful of managers writing member letters and board minutes — the math never gets there, and the operational risk of a server you half-understand is not worth $3,000 a year. If you have no second person who could restore that machine, do not do it. And if what you want is the finished-work behavior of tools like Claude Cowork or ChatGPT Work, where you hand over a goal and it comes back with a completed spreadsheet, that is not what a self-hosted model does. Owning gives you a very capable reader and writer that answers questions cheaply and privately. It does not hand you a finished 5300 Call Report worksheet.

What to do on Monday

Do not start with a hardware quote. Start with a usage count. Ask your vendor for last quarter's active-user report and sort it by how many days each licensed person actually used the tool. At most credit unions that report is brutal: a third of the seats show fewer than four active days a month. Cancel those at renewal first — that alone often saves more than the whole hosting project. Then, with the real number of heavy users in hand, ask your CUSO or data processing partner one question: "Do you host an open model for member credit unions, and what would you charge us monthly?" If the answer is a flat fee under what you are paying in seats, you have your decision without buying a server at all.

Frequently asked questions

Will our examiner have a problem with us running our own model?

Not with the concept. What they will ask about is governance: a board-approved acceptable use policy, a record of who can access it, evidence that answers touching member accounts are reviewed by staff, and change control on updates. Running it in-house actually simplifies the vendor due diligence conversation, because there is no third party receiving member data. Write the policy before the box arrives, not after.

Is the free model actually as good, or is that a sales pitch?

For reading your own documents, drafting member correspondence, and summarizing call notes, the gap has effectively closed in 2026. For long, complicated reasoning work — say, a tricky asset-liability scenario or a messy general ledger reconciliation — the paid frontier tools are still meaningfully better. Many credit unions end up with both: an owned box for volume work, and three or four paid seats for the CFO, the compliance officer and the CEO.

What happens when the server dies on a Saturday during tax season?

Your staff falls back to how they worked in 2024 — not a catastrophe, but a real service dip. Ask whoever hosts it what their restore time is, get it in writing, and keep a small number of paid seats alive as a backup. That combination is still cheaper than licensing everybody.

Do we need to hire anyone?

If you go through a CUSO or your data processing partner, no. If you self-host, you are not hiring a person but you are committing several hours a month of your most irreplaceable technical employee. Price that honestly in the board packet, because if you leave it out and it later shows up as overtime, you will lose credibility on the next technology request.

One related note: a lot of what your member service representatives look up all day starts as a ringing phone. CallSphere builds AI voice and chat agents that answer the member line, handle the routine questions — branch hours, routing numbers, card activation, "did my deposit post" — book appointments with a loan officer, and pass the rest to a human with the context already attached. It pairs naturally with an in-house model doing the reading and writing behind the counter, and it keeps the 11:30-to-1:30 queue from stacking up while your MSRs are on longer calls.