PAN Lab example

Illinois DCFS Augintel

The record that reads back: a case-note mining overlay

Caseworkers write the notes. An overlay reads every note ever written and hands search results, case overviews and safety alerts back to them. The same extractions travel upward as practice-fidelity measures, compliance data and statewide trends about the people who wrote them, and the vendor's own analysts sit in that loop. Modeled on the statewide case-note mining deployment at Illinois' child-welfare agency: its shape, not the real platform. There is no risk score anywhere on this board, because the same agency shut a predictive-scoring pilot down in 2017 and built the successor deliberately score-free. What is left is quieter. One layer nobody has evaluated is the narrator of the whole record, and almost every number anyone cites about it belongs to the vendor, including the ones spoken in Congress. Before you pick a target level: this board cannot be won under Service and Safety Targets or All Governance Targets. With every tool the Lab currently offers, no affordable combination brings this system inside the win condition at those settings. That is a measurement of the deployment this network is derived from, not a puzzle waiting to be cracked. Explore and Service Targets Only can be won.

Stylized model of a documented deploymentCaseworker documentation & copilots

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the Case-note-mining-class overlay read by two audiences network: 8 components and 16 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 10 assumed · 3 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

assumed
This models the case-note-mining pattern documented in the Illinois DCFS Augintel case file - one legacy narrative record read by one overlay that serves a copilot audience, a supervisory audience and a management audience at once - and not a reconstruction of the actual platform.
assumed
Every quantitative outcome figure in this case is vendor-originated, including the figures delivered in November 2025 congressional testimony and repeated by agency officials: about five hours saved per worker per week, a 20 percent administrative-time estimate, an up-to-30-percent compliance increase, 621 detected expectant mothers against fewer than 100 known before, tripled kin networks, an 80 percent unstructured-data share, and a training corpus claimed at more than 200 million case notes. A hearing room is a channel, not a check; the figures are labelled at each use and none of them is treated as measured.
assumed
The heavy workload against very limited manual capacity is derived, not defaulted. Demand comes from the hearing record's own baseline of over four hours per day of caseworker administrative work across a statewide caseload with roughly 7,500 provisioned users. The low capacity is the counterfactual the overlay actually replaces: retrieving and synthesizing a full narrative corpus by hand out of a legacy green-screen system. The record's most concrete statement of what that manual capability saw is that the agency knew of fewer than 100 expectant mothers in care where the tool later surfaced 621 more - a vendor figure, and the reason capacity is drawn very low rather than at the corpus default.
baseline
The whole-corpus read edge is drawn at the top of the scale on the strength of the agency's own federal filing, which states that all notes in the case-management system can be mined, and of the vendor release describing the safety feature scanning current, prior and connected case notes at case open or transfer. It is the only maximal edge in this network that rests on a primary agency document rather than a vendor statement.
assumed
The training-corpus inflow is drawn at a substantial level rather than full strength because the deployed behaviour the record describes comes from the state-tuned model rather than the raw corpus, and the corpus size is a vendor claim no independent source verifies. It is deliberately not marked privacy-sensitive: no source located describes what the corpus contains, whose records it draws on, or whether it is de-identified, so the conservative reading is to record the silence rather than assert an exposure.
assumed
The source-text output constraint is drawn as a bounded automated check on the copilot channel, at a low level. It is drawn at all because first-year coverage records the agency and vendor stating the tool identifies information by extraction and abstractive summarization and does not compose new content, and because the product ships click-through to the underlying note. It is drawn no higher because abstractive summarization is a rewrite and no published evaluation bounds how far a summary can move from its source. It is placed on the frontline channel because that is the audience the statement was made about.
assumed
The supervisory tier is a reviewer with a real inbound pathway rather than an ornament: the launch materials name supervision-meeting preparation and engagement monitoring as first-class uses and the role breakdown lists supervisors as provisioned users. Its review pathway is drawn faint because the practice is named and never measured - no coverage, cadence or completion figure exists in any source located.
assumed
Leadership, quality assurance, data stewards and the vendor's own analytics group are drawn as a second group of staff because the sources give them a different relation to the same record than the frontline class has: they read extractions about the workers and the families and decide where programs are established or ended, and the vendor's staff sit inside that loop. Their only inbound pathway on this network is the extraction channel - no source documents them reading the raw record directly, and that narrowness is drawn rather than filled in.
assumed
The management-to-frontline peer edge carries a documented channel - fidelity measurement, compliance data gathered in the background from the notes, and program decisions reaching the people whose writing produced them. What that channel does to documentation behaviour is a structural inference from the documented uses, flagged as such in the case file and never asserted here; the only documented effect on documentation is a vendor-relayed claim that quality and motivation improved. This is also the only operator-to-operator reinforcing pathway on this network: no peer diffusion between caseworkers is documented for this deployment, so none is drawn.
baseline
Three pathways are drawn without flow because the sources document them as absent rather than merely unobserved: machine text entering the permanent record (the agency describes a read-side tool that does not compose new content), the published usage-and-trend assessment (promised by the agency in 2023, with nothing published through the November 2025 record and no audit of this deployment located), and an independent accuracy review (the public figures are vendor detection yields, never precision or recall). Each is drawn so the gap is visible and a lever can reach it.
assumed
The data flow out to the vendor's platform is drawn present rather than inactive because the vendor's position over the extraction is the deployment's design, not a pressure that might open: the vendor built the layer that reads the corpus, tuned it over about six months into a state-specific model, and its own analytics group reads the extractions inside the agency's management loop, with this deployment the anchor customer of a platform the vendor also runs for other jurisdictions. Where Illinois notes are processed or held is stated in no source located, and nothing here asserts a hosting arrangement; the crossing is drawn on the vendor's documented operating position instead. It is held at a substantial level rather than full because the flow stays inside the child-welfare purpose the notes were written for and no replication of the record into another jurisdiction's view is documented. The Illinois contract value and procurement route appear in no public source located, so nothing here characterises the data terms.
baseline
This agency's oversight record is historical, not current. The joint inspector-general finding and press scrutiny that ended its predecessor predictive-analytics pilot in December 2017 is the only oversight event at this node documented to have changed system behaviour, and it shaped the current score-free design. The November 2025 congressional hearing was promotional in posture and its majority summary records no AI-risk scrutiny, so no claim about present oversight rests on it.
assumed
Served children and families are not in the dynamics. This deployment carries no demographic scoring and no risk score of any kind, and no notice or correction channel is documented for families whose notes generate a label. No civil-liberties or family-advocacy critique targeting this deployment specifically was located, which is an absence of evidence rather than a clearance. Any differential harm is recorded in the case file and measured outside any diagram like this one.

What this example does not show

Vendor figures remain vendor figures even when spoken in Congress. Every quantitative outcome in this case is vendor-originated, including figures delivered in the November 2025 testimony and repeated by agency officials: about five hours saved per worker per week, a 20 percent administrative-time estimate, an up-to-30-percent compliance increase, 621 detected expectant mothers against fewer than 100 known before, tripled kin networks, an 80 percent unstructured-data share and a training corpus claimed at more than 200 million case notes. The hearing was a promotional channel, not oversight scrutiny, and no independent evaluation of detection accuracy, time saved or outcomes has been published.
Served children and families are not modelled here. This Lab reads institutional propagation only, and the people whose case notes are being mined are boundary-only. They have no documented notice or correction channel when a note about them generates a label, and any harm to them is recorded in the case file and measured outside any diagram like this one.
The write-side loop, workers adapting what they document because the notes are mined for fidelity, compliance and early-warning signals, is a structural inference from the documented uses and is flagged as one. The only documented effect on documentation behaviour is a vendor-relayed claim that quality and motivation improved.
Claims that oversight changed behaviour at this agency rest only on the December 2017 shutdown of its predecessor predictive-analytics pilot. No audit of the current deployment was located, and nothing here treats the 2025 hearing as scrutiny of it.
The Illinois contract value and procurement route are not public in any source located, so no cost claim is made here beyond the vendor's own comparison to a full case-management-system replacement.
No civil-liberties or family-advocacy critique targeting this deployment specifically was located. That is an absence of evidence rather than a clearance, and it should not be read as one, given the family-surveillance and worker-surveillance surfaces the record documents.
The primary record ends in November 2025. The deployment is ongoing as of 2026 with no post-2025 primary documents located, so nothing here describes its current configuration.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

Illinois DCFS rolled out a vendor natural-language-processing layer statewide over its legacy SACWIS case-management system beginning in 2023, provisioning roughly 7,500 users with a DCFS-estimated 20 percent administrative-time saving; the agency's federal Annual Progress and Services Report states that all notes in SACWIS can be mined, rendered as per-case Risks and Strengths overviews with click-through to source notes. By the vendor CEO's November 2025 congressional testimony, more than 6,000 Illinois staff use the tool and it saves each social worker about five hours per week — vendor figures delivered in a hearing whose majority summary records no AI-risk scrutiny — and no independent accuracy evaluation, published usage metric, or audit of the deployment has been located.
empirical
- Government Elisco, Written Testimony of Marty Elisco, CEO and Co-Founder, Augintel, Hearing: Leaving the Sticky Notes Behind (US House Committee on Ways and Means, Work and Welfare Subcommittee, 2025) https://waysandmeans.house.gov/wp-content/uploads/2025/11/Marty-Elisco_Written-Testimony.pdf
- Government Illinois Department of Children and Family Services, 2024 Illinois Annual Progress and Services Report (APSR) (2024) https://dcfs.illinois.gov/content/dam/soi/en/web/dcfs/documents/about-us/reports-and-statistics/documents/apsr-fy24.pdf
- Trade press Government Technology, Illinois Adopts Natural Language Processing Tech for Child Welfare (2023) https://www.govtech.com/computing/illinois-adopts-natural-language-processing-tech-for-child-welfare
- Government House Ways and Means Committee, Four Key Moments: Hearing on Harnessing Innovation and New Technology to Help America's Foster Youth Succeed (2025) https://waysandmeans.house.gov/2025/11/20/four-key-moments-hearing-on-harnessing-innovation-new-technology-to-help-americas-foster-youth-succeed/
From the 2023 announcement onward the same platform carried a second, management-facing channel reading the same note store: early-warning signs, caseworker safety, practice-model fidelity, and statewide trends per the launch release; compliance-related data collected in the background from the notes workers already document, service-outcome mining tied to funding accountability, and vendor-analytics deep dives informing where programs should be established or eliminated, per the vendor CEO's 2025 testimony; and a named Chapin Hall partnership using machine learning over case notes to measure practitioners' use of Motivational Interviewing. The associated documentation-behavior effect on the workers whose notes are mined is a structural inference, not a documented finding.
empirical
- Vendor Augintel, Chicago Start-up Augintel Unlocks Key Insights in Illinois DCFS Case Notes Using Natural Language Processing (NLP) Software (PR Newswire, 2023) https://www.prnewswire.com/news-releases/chicago-start-up-augintel-unlocks-key-insights-in-illinois-dcfs-case-notes-using-natural-language-processing-nlp-software-301881319.html
- Government Elisco, Written Testimony of Marty Elisco, CEO and Co-Founder, Augintel, Hearing: Leaving the Sticky Notes Behind (US House Committee on Ways and Means, Work and Welfare Subcommittee, 2025) https://waysandmeans.house.gov/wp-content/uploads/2025/11/Marty-Elisco_Written-Testimony.pdf
- Academic Chapin Hall at the University of Chicago, Brian Chor (staff page listing 2025 presentations: Chapin Hall / Illinois DCFS / Augintel machine learning partnership on Motivational Interviewing use) (2025) https://www.chapinhall.org/person/brian-chor/
In December 2017 the same agency, Illinois DCFS, terminated its Eckerd Rapid Safety Feedback predictive-analytics pilot — a 366,000-dollar sole-source arrangement that scored more than 4,100 children at 90-percent-plus probability of death or serious injury while assigning low risk to two children who died — after a joint OEIG/DCFS-OIG report found the arrangement had been misclassified as a grant rather than a no-bid contract. That shutdown is the only oversight event at this agency documented to have changed system behavior, and it preceded the current deployment's deliberately score-free extraction design.
empirical
- Investigative The Imprint, Illinois Drops Rapid Safety Feedback, A Predictive Analytics Tool (2017) https://imprintnews.org/politics/stateline-illinois-drops-rapid-safety-feedback-predictive-analytics-tool/28913
- Trade press Government Technology, Illinois Ends Child Abuse Prediction Program (2017) https://www.govtech.com/health/illinois-ends-child-abuse-prediction-program.html

Where this connects

Institutional pressures in this domain

Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
Deadline pressure — Statutory or managerial timeliness rules reward fast approval of machine output over slow disagreement.
Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
Staff turnover — Experienced skepticism leaves; new staff calibrate their trust on the tool itself.
Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.

All of them in context on the Caseworker documentation & copilots domain page.

Levers available here and the patterns behind them

Vet connections — Connection authorization
Store less data — Data minimization
Peer sharing rules — Peer-edge governance
Mark AI-written records — Provenance labeling
Keep skills sharp — Deskilling-arrest mandate
Review on schedule — Oversight cadence & retrospectives
Check with a second model — Cross-model verification
Gate record entries — Human-in-the-loop write gating
Gate vendor updates — Vendor quality gate
Pause AI on alarms — Deployment circuit-breaker
Upgrade model — Improve the model