The Army's AI Call-Monitoring RFQ: What a Real DoD AI Requirement Looks Like

Most federal AI requirements are written by people who are not writing for an audience. They are writing to buy something by a date. That is what makes a small Army solicitation posted on July 31, 2026 more instructive than another vendor keynote: an operating unit of the government says, in public and in its own words, what it wants an AI system to do.

DefenseScoop reported the notice on August 3. The underlying document is solicitation W91QF4-26-Q-A021 on SAM.gov, titled “ACB Inmate Call Monitoring Software.” Everything below comes from that posting unless otherwise attributed.

What was posted, and by whom

The issuing office is W6QM MICC-Fort Leavenworth, an office of the Mission and Installation Contracting Command inside Army Contracting Command. Place of performance is Fort Leavenworth, Kansas. The notice type is a combined synopsis/solicitation, meaning the synopsis and the solicitation are published as one document rather than a market survey followed weeks later by a request. Posted July 31, responses due August 13 at 2:00 p.m. Central, automatic archive August 28.

The background section states the scope of the operation plainly: the Military Correctional Complex “encompassing the USDB, MWJRCF, and the NWJRCF” holds “more than 650 inmates,” and the United States Disciplinary Barracks “is the only maximum-security facility within the Department of Defense.” It then draws the line that governs everything that follows: “All non-privileged inmate communications are monitored to prevent criminal and unauthorized activity.” The word doing the work there is non-privileged, which places attorney-client and similar protected communications outside the scope of what is monitored.

The stated purpose: “Having the ability to utilize A.I. software and systems to assist in monitoring communications will increase staff and inmate safety as well as provide the ability to detect, deter, and investigate criminal and unauthorized activities.”

What the requirement actually asks for

The scope paragraph names the integration point. The product “shall provide an A.I. software that is AWS Cloud GOV certified and interfaces with the Military Corrections Command existing ViaPath Phone system.” ViaPath Technologies, formerly Global Tel*Link, is one of the large incumbent providers of correctional telephone systems. That single clause tells a vendor more than the rest of the document: this is not a greenfield build, it is an analytics layer bolted onto somebody else's call platform, and the audio, metadata and access model all belong to that platform.

The salient characteristics are where the AI requirements live. Condensed, the notice asks for:

• A website-accessible platform using “both semantic and agentic A.I.” for call and message monitoring.
• Digital transcripts of all non-privileged calls “in near real time,” scanned for searchable keywords.
• Continuous monitoring that “sends alerts to correctional facility staff when a keyword is found in a transcript.”
• Detection of vocabulary drift: the tool “uses the transcripts to determine when new words are being used to replace old words for criminal activities, I.E. ‘murder’ and drug names.”
• “Analysis of 100% of the volume in near real-time, high-accuracy rate of transcription, immediate translation from Spanish to English.”
• Multi-word combination search, 24/7 operation, “extensive audit functions and link analysis availability.”
• Daily law enforcement assistance reports on “incidents, issues, and criminal trends,” plus a named Account Manager and System Specialist, and staff training delivered over Microsoft Teams.

Neither “semantic” nor “agentic” is defined. Read against the rest of the list, semantic plausibly means matching on meaning rather than literal strings, and agentic plausibly means the platform takes follow-on actions, generating the daily reports and firing alerts, without a person driving each step. A vendor that guesses wrong on those two words builds the wrong product, which is why the question period on a solicitation is worth more than the deadline.

Reading the notice metadata

The fields around the description carry information most readers skip. The solicitation number itself decomposes under FAR 4.1603: positions one through six are the activity address code of the issuing office (W91QF4), positions seven and eight are the fiscal year (26), and position nine is the instrument type. Here that letter is Q, which FAR 4.1603 assigns to requests for quotations. So this is an RFQ, not an RFP, which signals simplified acquisition procedures, a price-driven evaluation, and a short fuse. Thirteen calendar days from posting to close is consistent with that reading.

The NAICS code, 513210 (Software Publishers), fits. The product service code sits askew. It is 7B21, which the official PSC manual defines as “IT and Telecom - Compute: Mainframe (Hardware and Perpetual License Software),” covering “computing hardware and software products for traditional mainframe computers and operations running legacy operating systems.” The same manual carries a separate family of codes for subscription software, including DA10 for business application software delivered as a service. This requirement describes “deployment of a cloud-native platform” with “continuous platform improvements delivered as part of the subscription.”

Why care about a mismatched four-character code? Because that code is how vendors find the work. Saved searches on SAM.gov are built on PSC and NAICS, and a speech-analytics company filtering on software-as-a-service codes would never have seen this notice. Classification is discovery, so search the requirement language too, not only the taxonomy.

The stack hiding behind four bullets

The following is our analysis, not reported fact. Four short bullets in this notice describe a system that is harder to build than it reads.

Transcription of 100% of volume, near real time. Correctional phone audio is narrowband, compressed, noisy, frequently code-switched, and spoken by people who know they are recorded. Word error rates measured on clean benchmark audio do not survive contact with it. An honest bid states the error rate on this audio, which is the discipline behind how to evaluate an AI vendor's accuracy claim.

Keyword alerting. Trivial to demo, hard to run. The operating variable is not whether a keyword fires but how many false alarms a small staff will absorb before they stop reading alerts. That is a precision-and-recall tradeoff with a human cost on both sides, and it belongs in the requirement as a number. Our primer on precision, recall and false-extraction rate covers how to set that dial deliberately.

Detecting substituted vocabulary. This is the research-grade item on the list. Finding that a new word has taken the semantic position of an old one is lexical-change detection: track how a term's neighbors in embedding space shift over time, or watch which unfamiliar tokens start appearing in the contexts a known term used to occupy. It works, and it also produces a stream of plausible-looking false discoveries that need human adjudication before anyone acts on them.

Agentic behavior. Automatic daily reports on “incidents, issues, and criminal trends” means a model summarizing model output, two inference layers away from the audio, with transcription error compounding into summary error. Our guide to when agentic AI genuinely helps makes the same point in a lower-stakes setting: autonomy is cheap to add and expensive to evaluate.

What the document does not say

The gaps are the most useful part of the notice, because they generalize. There is no numeric accuracy threshold, no false-positive budget, no described acceptance test, no data retention or records schedule, and no stated language coverage beyond Spanish to English. “AWS Cloud GOV certified” is not a defined term of art; the defined terms are FedRAMP authorization and the DoD impact levels, and a serious response names both along with its authority to operate path. The requirement for “customer support by POST certified experts” is left undefined in the posting as well.

One more line is worth flagging for anyone who writes statements of work. The opening sentence calls this “a non-personal services contract.” Later, the same document asks the contractor to “collaborate with MCC investigators” and staff an Account Manager and System Specialist “to support Agency priorities.” FAR 37.104(a) defines a personal services contract by “the employer-employee relationship it creates between the Government and the contractor's personnel,” and 37.104(b) bars agencies from awarding one without statutory authority. Nothing here says the line was crossed. It does say the line is thin, and thin lines get tested during performance.

Why it matters

Three things travel beyond this one buy.

First, the salient-characteristics list is the acceptance test whether or not it was drafted as one. Every bullet becomes a thing the government can point at later. If you are the vendor, the highest-value move is not to echo those bullets back in your quote but to supply the missing measurements yourself: state the word error rate you will hold on their audio, the alert precision you will hold at a stated recall, and the method used to compute both. That converts unstated expectations into a shared definition of working, which is the whole subject of reading a solicitation without missing the important parts.

Second, the governance question is open and worth asking out loud. OMB memo M-25-21, the standing federal policy on agency AI use, sets minimum risk-management practices, including pre-deployment testing and ongoing monitoring, for AI designated high-impact, a category whose listed examples include law enforcement applications. Whether this system carries that designation, and who makes the call, is not answered in a SAM.gov posting. Federal buyers of AI are working under rules that keep moving; see our read on the OMB unbiased-AI procurement rule and on what changes when AI touches regulated data.

Third, this is what most defense AI actually looks like. The headline contracts run to billions; the volume runs through documents like this one, with a two-week fuse, an incumbent integration constraint, undefined terms of art, and technical claims nobody has been asked to quantify. Reading them closely is a skill, and the primary source is free.

Sources: SAM.gov contract opportunity — “ACB Inmate Call Monitoring Software W91QF4-26-Q-A021,” MICC-Fort Leavenworth, posted July 31, 2026 (primary source; notice ID 6cdb70ff186f4236a4986141982cff8f, NAICS 513210, PSC 7B21, responses due August 13, 2026); DefenseScoop, August 3, 2026; FAR 4.1603, PIID numbering; FAR 37.104, personal services contracts; FPDS Product and Service Codes Manual (April 2024, PDF); U.S. Census Bureau NAICS, code 513210. Analysis and framing by Precision AI Academy.

Common questions

Has anything been awarded? Not as of this writing. The posting is a solicitation with a response deadline of August 13, 2026, and no award notice was attached to it when we read the record.

What is the difference between an RFQ and an RFP here? The ninth character of the solicitation number is Q, which FAR 4.1603 assigns to requests for quotations. A quote is not an offer the government can accept into a binding contract the way a proposal is; the government issues a purchase order in response. In practice, a faster and more price-sensitive buy.

Does the system monitor attorney calls? The notice scopes the requirement to non-privileged communications and states that all non-privileged inmate communications are already monitored. How privileged calls are identified and excluded is a function of the underlying phone system rather than something the posted description specifies.

Where would a small software company start on something like this? With the questions, not the quote. On a thirteen-day RFQ the undefined terms are the entire risk, and asking the contracting office in writing is both allowed and expected.