The question
It's September. Your queue is forty actions deep, and most of them were supposed to be awarded last month. Someone in the office has already asked whether ChatGPT can help.
The answer is yes, for maybe a third of what's in front of you. The risk is not that AI handles that third badly. The risk is that under this much pressure you stop sorting which third it is — and one afternoon a package that shouldn't have gone into a commercial tool goes into a commercial tool, because the queue was long and the tool was open.
That sorting is the skill. Not prompt writing. Not picking between ChatGPT and Claude. Sorting.
This issue is about how to sort a year-end queue in the two weeks before the thirtieth, and how to do it without putting something into a system you cannot take back.
Field notes: triage, not tools
Take the queue and sort it three ways. Not by dollar value and not by customer — by what a commercial AI tool can do with the work.
Real hours. Summarizing, restructuring, gap-checking, and turning something long into something short for a program manager who was never going to read the long version. A twenty-page draft PWS becomes a one-page plain-language summary for the requiring official in about four minutes, and the four minutes are real, because the alternative was you writing it at seven that evening. This bucket is where the time is. It is also the least glamorous bucket, which is why most AI training skips past it.
Minutes saved, minutes owed back. Anything where the output has to be checked line by line before you can use it. AI will produce a market research summary in ninety seconds, and then you will spend twenty minutes confirming that the vehicles it named exist and the pricing it cited is current. That is not a bad trade on a slow week. In the last two weeks of September it usually is, because verification is the part that gets skipped when the clock is the constraint, and an unverified output is worse than no output.
No business at all. Determinations, findings, and anything carrying your signature as a considered judgment. Anything touching a proposal you have already received. Anything where the input itself is the problem rather than the output.
The decision rule underneath all three:
Year-end doesn't change the rule. It changes the ratio — more work in bucket one, less time to police the line between two and three. Sort the queue once, in writing, at the start of the week. A list you made when you were calm is a better control than a judgment you make at 6:40 on a Thursday.
The prompt: the requirements package gap-check
This is a bucket-one task. Run it on an anonymized requirements package before it reaches your desk for action. It finds what is missing, which is the failure mode that costs a week later in the cycle.
You are reviewing a draft requirements package for completeness only.
Do not rewrite it, do not improve the language, and do not add requirements.
Using only the text I have pasted below, list what is missing or unmeasurable
in these five categories:
1. Deliverables that are named but never defined
2. Acceptance criteria that cannot be objectively measured
3. Missing period of performance, place of performance, or delivery schedule
4. Requirements written as a design specification where a performance
outcome appears to be what is wanted
5. Internal contradictions between sections
For each finding, quote the exact line from my text that prompted it, and
state what a reviewer would need to see instead. If a category has no
findings, say "none found" rather than producing one.
Do not tell me what any regulation requires. I will check that myself.
TEXT:
[paste here]The guardrail. Anonymize before you paste. Strip the program name, the office, the vendor names, the dollar figures, and any internal point of contact. And never paste a document a vendor submitted to you — not a proposal, not a capability statement, not a quote. Vendor-submitted documents can carry instructions hidden in white text, alt text, or metadata that a model will read and you will not.
The verification step. Every finding quotes a line from your text. Go check that the quoted lines appear in your document, word for word. A quotation that isn't there is a fabrication, and one fabricated quote means you discard the whole output rather than salvage the parts you liked. That check takes ninety seconds and it is not optional.
What it returned
We built a deliberately flawed sample package for course use — not a live document — and ran the prompt against it. The relevant lines:
C.2 The Contractor shall provide comprehensive preventive maintenance
for all HVAC equipment.
C.3.2 The Contractor shall respond to emergency service calls in a
timely manner.
C.3.4 The Contractor shall install Trane Model XR14 replacement units
where compressors have failed.
C.4 Deliverables: Monthly Status Report. Annual Maintenance Plan.
Quality Control Plan.
C.5 All work shall be performed in a professional and workmanlike
manner to the satisfaction of the COR.
C.6 The Contractor shall provide 24/7 coverage. Normal hours of work
are 7:00 a.m. to 4:00 p.m., Monday through Friday.It caught what you would want a second reviewer to catch. Three deliverables named in C.4 with no content, format, or due date attached to any of them. "In a timely manner" and "to the satisfaction of the COR" flagged as unmeasurable, with the note that a response time and an objective standard belong there instead. No period of performance anywhere in the document. And C.3.4 called out as a brand-name design specification sitting inside a package that otherwise reads as performance-based.
That is a useful ten minutes. It is not a substitute for reading the package.
What it missed
The expensive one. C.2 says "all HVAC equipment" and the package contains no equipment inventory and references no attachment. There is no basis to price the work. That is the most costly omission in the document and the prompt walked past it, because the five categories cue completeness of language rather than completeness of scope. If you use this prompt, add a sixth category: quantities or inventories referenced but not provided.
The false positive. It flagged C.6 as an internal contradiction — 24/7 coverage against 7-to-4 duty hours. Those are reconcilable, and routinely are; what the package needs is an after-hours response and rate provision, not a rewrite. A reviewer acting on that finding goes back to the requiring official over nothing, which is exactly the credibility you cannot spend twice in September.
Both failures are the same lesson. The tool is a checklist that reads fast, not a reviewer that understands what the work costs.
What could go wrong
Year-end pressure is exactly when data-safety shortcuts happen, and the shortcut is almost never a decision. It's a paste.
Three buckets, and the sorting takes a second once you've done it a few times.
Public — paste freely. A posted solicitation. A published regulation. Your agency's own public AI policy. Anything already on a public website is already public.
Anonymize first. A draft statement of work. A market research summary. An internal memo. Strip the identifiers and the numbers, and what's left is usually structure — which is the part AI is useful on anyway.
Never. Source-selection-sensitive information. Controlled Unclassified Information. Anything a vendor submitted. Evaluation notes, scores, and consensus documents. Pre-award pricing. There is no anonymization that makes these safe, because the sensitivity is in the substance, not the labels.
The mistake to watch for is treating "I removed the names" as sufficient. A redacted evaluation narrative is still an evaluation narrative.
Your jurisdiction
Federal. CUI and source-selection-sensitive material are the two categories that end the conversation — no anonymization, no exceptions, no matter what the queue looks like on the twenty-ninth. The end-of-year surge is a well-known pattern, not a rule that relaxes anything. Check your agency's own AI guidance before you rely on any of this, because it may be narrower than the general position.
State. Your fiscal year started in July, so this is planning season rather than close, which makes it the better month to set your own rules before you need them. What counts as public varies enormously by state, and that variation is the whole game here. Your state's public records act is the starting point, not your comfort level.
Local. No surge staff and often no policy at all, which makes the triage rule matter more rather than less. Write your three buckets down once and keep them where you can see them. A one-page rule you wrote yourself beats a policy you don't have.
Contractor. A buyer under year-end pressure needs your submission to be findable, complete, and boring to evaluate. Assume the government reviewer is working from a summary of your document rather than the document, and write the first page accordingly. Also assume your proposal is not going into anyone's chatbot — and if you have reason to think otherwise, that's a conversation with the contracting officer, not a complaint after award.
What changed
August 21 — NARA issued AC 11.2026, guidance on applying the Federal Records Act to AI materials. The headline is that using an AI platform does not by itself create a federal record; the substance is that subsequent use of AI-generated material in agency business can. Worth reading in full before you decide what to keep.
September 21 — SBA's proposed size standards comment window closes. The August 20 proposed rule would consolidate 995 standards into 338 and raise thresholds broadly, including in management consulting. If you compete for or award set-asides, the competitive pool changes. Docket SBA-2026-0199.
Still pending — GSA's draft GSAR clause on safeguarding data in large language model systems. Comments closed August 3. As of this writing we have found no final clause or deviation issued, so contractors and buyers are still working without that text.
From the course
This issue is Module 1 of Commercial GenAI for Public Sector Procurement — specifically Video 1.2, "Data Safety and What You Can (and Cannot) Input," which covers the three-bucket classification, practical anonymization technique, and the prompt injection rule behind the guardrail above. Module 1 is available now; Modules 2 through 5 release as they finish. Roughly 3.5 CLPs for Module 1, subject to your agency's approval. Purchase orders and government purchase cards are accepted.
Your turn
Here's the one I'd like an answer to: what task did you decide not to use AI on this month, and what made you stop?
Reply to this email. The answers shape what goes in the next issue, and the ones that change my thinking get written up — anonymized, always.
About The Wolverine Group
The Wolverine Group is a Woman-Owned Small Business in Washington, D.C., delivering information operations, open-source intelligence, and acquisition solutions to the Department of Defense, the Intelligence Community, and federal civilian agencies. For twenty-five years we have turned complex, multi-source data into decisions leaders can act on. Our acquisition practice supports federal buyers directly and trains the workforce that does the buying — which is where this brief comes from. UEI VK81K1DMC6L5 · CAGE 6BUU5 · Top Secret facility clearance · CMMC Level 2.