By Sagar Shankaran, Founder of CallSphere
How A/E firms hand SF-330 assembly to Claude Cowork or ChatGPT Work: what it drafts, the worked hours math, and the profiles a licensed human must still read.
Key takeaways
Pull the marketing charge code out of Deltek Vantagepoint for one qualifications pursuit — the state DOT on-call for bridge load ratings, say — and add up every hour your marketing coordinator, two project managers and a principal booked against it. Between hunting project profiles, rewriting resumes into the government's boxes, chasing sub-consultant letters and fighting the page limit, a single SF-330 lands somewhere north of thirty hours. Call it thirty-four, and check your own timesheets before you believe me.
Thirty-four unbillable hours, for a package with maybe a one-in-four chance of a shortlist. Under the Brooks Act and its state equivalents you cannot buy your way in with a low fee — you win on qualifications, so the packet is the sales call. That job is the clearest candidate in an A/E office for being handed over as a goal rather than a task list.
An SF-330 is not a brochure. Section E wants a resume per key person: role, years with this firm, years elsewhere, registration by discipline and state, relevant projects with dates. Section F wants up to ten project profiles, each with the owner's contact, construction cost, completion year, and a description proving the work is like the work advertised. Section G is the matrix tying every person in E to every project in F. Section H is the approach narrative — the only part a principal genuinely needs to write.
None of it lives in one place. Project data is in Vantagepoint or Unanet or BQE CORE. Resumes sit in a folder where three versions of the same structural engineer's bio disagree about whether he is licensed in Tennessee. The last twelve submittals are PDFs nobody indexed. Half the sub-consultant material arrives by email the afternoon before it is due, because the MBE/WBE partner's coordinator is doing the same scramble.
An agent-run pursuit desk is software you hand a goal to — "assemble the SF-330 for this RFQ" — that then works across your project database, your resume library and your past submittals on its own for hours, and returns a drafted packet for a principal to correct and sign. It is not a chat box you paste a question into. That distinction is the whole point of this post.
Anthropic released Claude Cowork on 12 January 2026 and extended it to mobile and web in July. OpenAI released ChatGPT Work on 9 July 2026, running on GPT-5.6. Both were built for people who do not write code: you state the outcome, connect it to the files and applications the work lives in, and it breaks the job into steps, grinds away for a couple of hours, and hands back finished work — a filled spreadsheet, a priced packet, a set of completed forms.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent in your browser — 60 seconds, no signup.
The 2024 and 2025 version of this idea wrote paragraphs when you asked. You still supplied every step. The 2026 version accepts the outcome and works out the steps, then keeps going while you are in a design review. For a coordinator who spends a pursuit week retrieving and reformatting rather than writing, that is the difference between a tool and a second set of hands.
The cost side moved too. Running capable models is down roughly tenfold from 2025, which matters because a qualifications packet is a long job with a lot of reading. Two years ago, having software read eleven old submittals and forty project records to build one Section F was an argument with your CFO. It now costs less than the coordinator's parking that week.
flowchart TD
A["RFQ posted on the state DOT vendor portal"] --> B["Principal marks it a go at Monday pursuit review"]
B --> C["Coordinator states the goal and attaches the RFQ"]
C --> D["Agent pulls project records from Vantagepoint and bios from the resume folder"]
D --> E["Agent drafts Sections E, F and G into the SF-330 boxes"]
E --> F["Agent flags gaps: missing construction cost, expired seal, no owner contact"]
F --> G{"Principal review: is every fee, role and date correct?"}
G -->|Corrections needed| H["Coordinator fixes and logs the miss on the go-back list"]
G -->|Clean| I["Section H written by the principal, packet submitted before 2:00 p.m."]
The RFQ for a municipal water main replacement program drops Monday afternoon; responses are due in eleven days. At Tuesday's pursuit review the principal-in-charge says go. The coordinator opens Claude Cowork, points it at the RFQ file, the firm's project database and the resume folder, and states the goal: build the SF-330 for this solicitation, pull the ten most relevant water and sewer projects completed in the last seven years within 150 miles, and build Section G against the six people named in the go decision. Then she goes to a different pursuit.
By early afternoon there is a draft. Section F has nine profiles, not ten, with a note that the tenth candidate — the 2021 lift station rehab — has no final construction cost in the accounting record and a retired owner contact. Two resumes carry a registration that expired in a state the firm no longer works in. Section G is built, and it marks two people as thin: claimed on projects where their charged hours were under forty.
That last part is the useful bit. It read the pursuit like a nervous reviewer and told you what a scoring committee will notice. The project manager spends ninety minutes fixing real problems instead of two days on copy-and-paste.
Assumptions, all of which you should replace with your own accounting numbers: 40 qualifications packages a year; 34 hours of internal time each at a burdened $95 an hour; a 25% shortlist rate and a 40% win rate on shortlists; average net fee of $185,000 on a win. Suppose the agent cuts assembly time by 55% — not by writing better prose, but by removing retrieval and reformatting — and the freed hours buy four more submittals a year.
| Line | Today | With the packet drafted for you |
|---|---|---|
| Packets submitted per year | 40 | 44 |
| Internal hours per packet | 34 | 15 |
| Total unbillable pursuit hours | 1,360 | 660 |
| Cost of those hours at $95 | $129,200 | $62,700 |
| Shortlists at 25% | 10 | 11 |
| Wins at 40% of shortlists | 4.0 | 4.4 |
| Net fee won at $185,000 each | $740,000 | $814,000 |
That is $66,500 of recovered labor cost and roughly $74,000 of additional fee, against software costing a few hundred dollars a month per seat. Honest caveat: 0.4 of a win is a fraction, and fractions do not show up in one year. What you can prove inside a quarter is the hours line, because it is already on the timesheet under your marketing charge code.
Everything on an SF-330 is a representation to a public owner. If a profile overstates your role — you did the utility relocation, not the whole roadway — that is not a typo, it is a false statement on a federal form. The agent will happily write "led design of" because a marketing summary in your folder said so in 2019. Someone with a license and a memory reads every profile.
Still reading? Stop comparing — try CallSphere live.
CallSphere ships complete AI voice agents per industry — 14 tools for healthcare, 10 agents for real estate, 4 specialists for salons. See how it actually handles a call before you book a demo.
Three other things stay human. The go/no-go call, because software cannot tell you the selection committee chair worked with your competitor for nine years. Section H, because that is where you say something about this specific corridor and this community's last two failed bond referendums — and generic approach language is how firms fall out of the top three. And the seal: no agent signs, seals or dates anything, and both your state board and your liability carrier have opinions about that.
One more limit: the output is only as good as your project records. A firm whose accounting system carries no construction values, completion dates or owner contacts will get a draft full of blanks. That is not a software failure — it is a filing cabinet that was already broken and is now visible.
Both, and private work is often easier because there is no fixed form. Start with the SF-330 because its structure is rigid, so the draft is easy to grade — a box is either filled correctly or it is not. Once your project records are clean enough to feed Section F, developer RFPs and higher-education proposals come almost free.
You connect it to the applications and folders you choose. Most firms start narrow: export the project list and resume folder to a shared drive the agent can see, run a few pursuits that way, then widen access once you have watched what it does. Never let it write back into your accounting system.
Nothing in the SF-330 instructions or in state QBS rules asks how the document was prepared, and the responsible party is still the licensed professional who signs it. What gets you rejected is an inaccurate project profile or a missed page limit — review failures, not drafting failures. Keep the review step.
Small firms feel it more: in an eleven-person office the person assembling the packet is usually a project manager whose hours are billable, so every hour that comes off a pursuit goes straight back onto a project. The math is better for you than for the 300-person firm with a marketing department.
Take the last qualifications package your firm submitted — one whose outcome you already know — and have the agent rebuild it against the same RFQ. Compare it section by section with what your coordinator actually sent. You will learn three things in an afternoon: how much project data is missing, which parts of the packet the software genuinely handles, and which parts your coordinator was adding judgment to that nobody documented. Then run the next small on-call pursuit for real, with the coordinator reviewing every box.
More submittals means more inbound calls — committee staff confirming receipt, an owner's representative with a question before the interview, sub-consultants chasing a slot on the next pursuit — and they arrive while your principals are in the field. CallSphere builds AI voice and chat agents that answer the firm's line around the clock, capture the caller's project and question, book the interview on a principal's calendar, and pass along a transcript. It does not write your SF-330. It makes sure the shortlist call at 4:50 on a Friday is answered by something other than voicemail.

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.
The monthly IEEE 1366 reliability close takes 64 hours across three people. What goal-driven agents change, the arithmetic, and what stays with the engineer.
How pest control service managers hand the monthly food-account trend packet to a 2026 work agent as a goal - and what has to change about assigning work.
The phased plan, insurance estimate, predetermination narrative and financing page, finished before the patient leaves. What the owner has to change to get it.
Why co-pack quotes take six days, and how 2026 agents that return finished work rebuild the packet — costed formula, freight, spec sheet — in two hours.
A 1/1 commercial submission packet costs an account manager nine hours, eight of them gathering. In 2026 you hand over the goal and review the finished packet.
How a design firm builds an RFI test set from closed jobs, scores citation and routing accuracy, and widens an agent's authority only when the numbers earn it.
© 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