By Sagar Shankaran, Founder of CallSphere
Seasonal seat counts break per-seat AI pricing for general contractors. A worked comparison against hosting an open model for bid-season document reading.
Key takeaways
Pull last year's invoices for whatever AI tool your office standardized on and put the seat count next to the month. On a commercial GC of about $85 million a year, it looks roughly like this: 34 seats in January and February, 62 through the spring bid run, 96 from June through September when the summer school work and the second shift are going, and back to 58 by Thanksgiving. At $40 a seat that is 812 seat-months, or $32,480 a year, for a tool nobody in your company thinks about.
Now look at who those seats are. Six estimators and the preconstruction manager live in it four hours a day from February through May. Nine project managers use it maybe an hour a day. And then there are sixty-odd foremen and field engineers who open it for six minutes to clean up a Raken daily report, plus the warranty coordinator, the AP clerk keying sub invoices in Sage 300 CRE, and the safety director building toolbox talks. Per-seat pricing charges the same for the foreman with six minutes as for the chief estimator in the middle of a $40 million hard bid.
That mismatch is old news. What is new is that as of 2026 you have a genuine alternative, and it is worth knowing exactly where the line falls before a vendor draws it for you.
Seat pricing was designed for offices where everybody sits at a desk all year and uses the same tool at roughly the same rate. A general contractor is the opposite in three specific ways.
Headcount swings with the season. In most of the country you staff up between April and October, and public school work compresses the whole calendar into the ten weeks between Memorial Day and Labor Day. You are paying for peak seats during the months when your field people have the least time to sit down and type.
Second, the heavy work is lumpy and it is not per-person at all. During bid season your estimating department will chew through fifty to seventy project manuals, each 1,500 to 2,000 pages, hunting for allowances, alternates, unit prices, liquidated damages language, the wage determination, and the scope splits between Division 03 and Division 05. That is volume work that wants to run overnight, not a person clicking.
Third, the biggest population of people who could use it is not on your payroll at all. You have four hundred subcontractor and supplier users touching your Procore account. You will never buy them seats.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for home services in your browser — 60 seconds, no signup.
An open-weight model is one whose maker publishes the model itself, so you can run it on hardware you rent or own and pay for the machine rather than paying by the person, by the month.
Through 2024 and 2025 the honest advice was that the free-to-run models were a step behind and you would feel it on hard reading. That gap closed a lot this year. Moonshot AI released Kimi K3, a 2.8-trillion-scale sparse mixture-of-experts model — the largest open model in the world, and built so that only the slice of it relevant to your question does the work, which is why it does not require a data center to answer. The rest of the open tier moved up behind it.
The practical translation for a contractor: reading a project manual and producing a scope sheet is now firmly inside what a model you host yourself can do well. So the question in front of you is no longer capability. It is buy-per-seat versus run-your-own, and that is a spreadsheet question with a real answer.
flowchart TD
A["Invitation to bid arrives, 1,800-page project manual"] --> B["Manual loaded on the in-house machine at 6pm"]
B --> C["Pass 1: allowances, alternates, unit prices"]
B --> D["Pass 2: wage determination and LD clause"]
B --> E["Pass 3: scope sheet for each bid package"]
C --> F["Draft bid tab in the estimator's own format by 6am"]
D --> F
E --> F
F --> G["Chief estimator edits it and sends the invitations"]
Here is the job that makes the case. Wednesday, 6 p.m., three invitations to bid land from the same public owner with a June letting. Each one is a project manual, a drawing set, and an addenda folder. The estimating department has eleven working days and four other bids in front of it.
The machine reads all three overnight. By 6 a.m. Thursday the chief estimator has, for each bid: every allowance and alternate with its page reference, the unit price schedule, the liquidated damages number and the milestone it attaches to, the Davis-Bacon wage determination and effective date, the bond and insurance requirements, the DBE participation goal, the substantial completion date, and a first-pass scope sheet for each bid package flagging the two or three items that always fall between trades — fire caulking, wall backing, roof blocking, temporary power.
None of that is a decision. Every line of it is a lookup, and every one of those lookups is currently done by a $78-an-hour estimator with a highlighter at 9 p.m. That work has no natural seat count, which is exactly why paying by the seat prices it wrong.
This is where most articles stop being useful, so here is the honest version. You do not have an IT department. You have an IT manager who also fixes the plotter, and a managed services provider on a monthly retainer. Running your own model adds three obligations to that: somebody keeps the machine current, somebody decides who can see which job's documents, and somebody answers the phone at 7 a.m. during bid week when it will not respond.
Budget four hours a week of your IT manager and an add-on to the MSP retainer. If your answer to "who owns this" is "we will figure it out," buy seats and stop reading. A tool that is down during the second week of April costs more than it saves.
Assumptions: 812 seat-months a year at $40, as above. Rented machine time for the overnight reading at $1,150 a month. MSP add-on $750 a month. IT manager at $56 an hour burdened, four hours a week. Twenty-eight people genuinely use it daily and keep commercial seats.
Still reading? Stop comparing — try CallSphere live.
See the home services AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
| Line | Math | Amount |
|---|---|---|
| Today: everybody on seats | 812 seat-months × $40 | $32,480 |
| Keep seats for the 28 daily users | 336 seat-months × $40 | $13,440 |
| Rented machine, in-house tool for everyone else | $1,150 × 12 | $13,800 |
| MSP add-on | $750 × 12 | $9,000 |
| IT manager time | 208 hrs × $56 | $11,648 |
| New total | $47,888 | |
| Difference | worse by | −$15,408 |
On seats alone, running your own loses. Say that out loud, because a lot of people will tell you otherwise. It only turns when you count the overnight reading, which per-seat pricing was never going to give you: 60 bids a season, six hours of estimator time each on lookups the machine can do, at $78 an hour burdened, is $28,080 of estimating capacity handed back. Against the $15,408 gap you are ahead by about $12,672 — and, more to the point, your chief estimator gets to bid four or five more jobs in a season without you hiring a second estimator at $118,000 burdened.
The lesson is not "self-host everything." It is: pay per seat for people, and run your own for volume. If you have no volume work, per-seat is the right answer and always was.
It will not decide scope. The two or three items that fall between trades are exactly the ones that get argued about at the bid table, and the reason they get argued about is that the documents genuinely do not say. Your chief estimator earns his money on that judgment, not on finding page 1,411.
It will not read a drawing the way a takeoff person does. On-Screen Takeoff and your quantity work are a separate discipline. What is being described here is the written word — the manual, the addenda, the front-end documents.
And it does not know your history with a subcontractor, which is half of preconstruction. The machine can tell you the DBE goal is 12 percent. It cannot tell you which of the three qualifying subs will actually staff the job in August.
No, and most contractors should not. You can rent the machine by the month or by the hour from a hosting company, which is what the numbers above assume. Buying hardware only makes sense if you also have documents that are not allowed to leave your building.
Some of it, eventually, and priced per seat. That is fine for the daily users. The overnight reading of fifty project manuals is not on anybody's roadmap as a seat feature, because it is not a seat-shaped job.
Give it to one person with a name and a deadline, run it against one bid week, and compare the scope sheets against what your estimator produced by hand. If it does not hold up on three real bids, kill it. That is a four-week test, not a program.
Start there. Before any of this, audit the seat list against actual use for the last ninety days. Most GCs find fifteen to twenty seats belonging to people who used it twice. That audit is free and it usually pays for the experiment.
A closing note on a related cost most contractors never put on this spreadsheet: during bid season, the office line rings all day with subs asking whether an addendum dropped, whether you are bidding a job, and whether they are on the invite list — and every one of those calls lands on the person you least want interrupted. CallSphere builds AI voice and chat agents that answer the phone and website chat, take the caller's details, book meetings and pass along the lead around the clock, so the estimating room gets its afternoons back. It answers calls; it does not price your bid.
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.
When per-seat AI licensing stops paying for a benefits agency, why the claims files decide it, and the three-year arithmetic on owning a machine instead.
Open models closed the gap in 2026. A worked buy-versus-rent comparison for a 58-employee credit union, including the staffing cost nobody budgets for.
400 estoppels and 1,900 inspection photos a month change the math on AI licensing for community association managers. A worked breakeven, plus the limits.
How commercial GCs search restricted SSI and CUI drawing sets using AI that runs on hardware in their own server room, with a worked cost example for one job.
When per-seat AI pricing stops making sense for a host agency or tour operator, what running an open model actually costs in people, and the data case for it.
How general contractors keep an AI assistant inside the conformed set, ASIs and approved submittals, with citations a super can check before inspection.
© 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