Education AI agent · Child nutrition
Child Nutrition Claim & Verification AI Agent
Own the household application, the direct-certification match and the counts behind the claim, while the student record, the operative notice and the wellbeing referral stay on their own pages and a named official signs.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Official signs
- Never cuts off
- After launch
- Agent Care
What this agent does
Builds the claim pack, not the signature
In
- 01
A meal is served and counted at the point of service, and it enters the record with its site and its day.
- 02
A claim month closes, and the daily counts are read against eligible enrolment times the attendance factor.
Reason
- 03
A site drifts past its expected count, and the exception is raised with the day, the meal type and the register.
- 04
A cleared exception carries no reason, so it stays open — an unexplained clearance is worse evidence than none.
- 05
A direct-certification match arrives and sets the meal benefit only — never an input to anything else about the child.
Decide
- 06
A verification sample falls due, and each lawful method is computed so the difference between them is visible.
- 07
A household does not reply, and the chase runs in the languages the district serves, with each attempt dated.
Out
- 08
A claim pack is assembled — counts by site and day, edit checks and exception resolutions — held for the official.
- 09
Execute write actions only inside the approval boundaries agreed during implementation.
Product statement
The agent assembles the claim; the school food authority official signs it and is responsible for reviewing and analysing the meal counts.
Example workflow
One claim month, service to signature
- 1Meals counted at servicePoint-of-service counts by site, by day and by meal type
- 2Eligibility assembledApplications, direct-certification matches and the documents behind each determination
- 3Claim pack assembledCounts by site and day, edit-check results, open exceptions and confidence
- 4Controls appliedEdit checks against attendance-adjusted enrolment, carry-over expiry, site rules and confidence threshold
No human action required
Stages 1 to 4 run unaided, and none of them determines or signs anything — the agent is assembling, and the official lane opens at the confidence gate.
Goes to the food authority official.
Adds a business-office read first.
Official review
The pack is held with its counts, its unresolved exceptions and the confidence.
Review · Correct · Send to the business office
Reviewed — the official signs
- 6Claim record updatedOnly where write access and approval policy allow it
- 7Outcome evaluatedEdit-check coverage, exception resolution, chase outcomes and findings raised in review
Corrections
Every correction the official makes is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Signing or submitting the Claim for Reimbursement.
- Approving or denying a household application.
- Terminating benefits, or closing out a household that did not reply.
- Choosing which verification sample the district will draw.
Automation boundary
Agent acts unaided
- ✓Assemble counts by site and by day into the monthly claim pack.
- ✓Run the edit check against attendance-adjusted eligible enrolment.
- ✓Compute each lawful sample and show whom each one would put at risk.
- ✓Chase non-responding households on a dated trail, and hold the file for a person.
Wrong after the claim is signed: the meals are eaten, and fiscal action reaches back to recover.
- Issuing the advance notice that precedes a reduction or termination.
- Confirming a sampled application, which a second individual must do.
- Judging an application questionable and verifying it for cause.
- Changes to edit-check, sampling or claim-control configuration.
Example output
One claim month, annotated
Everything the agent assembles is attached to the count or the document it came from.
Claim-pack output · single claim month
Illustrative example- Site
- Elementary site
- Logged finding
- Counts ran above the attendance-adjusted expectation and each flag was cleared with no reason
- Edit-check result
- Exception open
- Rule applied
- 7 CFR 210.8
- Confidence
- 91%
- Claim status
- Unsigned — exception unresolved
As received
Taken from the point-of-service counts and the eligibility file — the student record stays with our enrolment and records agent, and this same count is certified again for E-Rate under 47 CFR 54.505.
Record used
- Point-of-service count
- Attendance factor
- Eligible enrolment
Why this flag
The edit check is what makes the count claimable, and a flag cleared with no reason is the weaker record.
What the score decides
Below the configured threshold the pack picks up a business-office read before the official sees it.
Action
ReviewCorrectSend to the business officeValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
Every claim month
From the point of service
Assemble against the edit check
Draw on the counts by site and by day, the eligibility file and the attendance factor the district sets. The Food and Nutrition Service became the Food and Nutrition Administration on 1 June 2026; the desk that receives the claim and runs the review is still your state agency.
The record outlives its keeping
Routine count reconciliation and pack assembly arrive already done.
Point people at what a reviewer will reach for
Unresolved exceptions, expiring carry-over and an undated chase are marked. A verification request, an approval or denial and an advance notice of reduction are legally operative, and each is held for a person, as on our parent-communication agent.
Build an evidence trail
The count, the site record under it and the official who signed stay on the claim.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Point-of-service systems
- Titan
- LINQ
- Nutrikids
- Heartland
- MealsPlus
Eligibility and applications
- Online application portals
- State direct-certification files
District systems
- Attendance and enrolment feeds
- E-Rate eligibility extracts
Agent
Child nutrition claim and verification
- Reads the counts
- Runs the edit checks
- Holds for the official
Claiming and finance
- State claim portals
- District ERP and general ledger
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six layers between the model and the claim
Six meshes, finest at the end. Whatever the whole set lets by is drawn in the map below.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Test whether a count may stand
- L3
- Puts the pack in the hands of the official
- L4 – L5
- Keep the count and the record behind it
- L6
- Holds the claim at unsigned when signals degrade
How Nestack evaluates it
Evaluate the whole claim month — not only the pack at the end.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the claim the official signs
- E1Final-output evaluationDid each count in the pack trace to a site, a day and a check?
- E2Step-level evaluationDid the agent use the right sites, eligibility file and attendance factor?
- E3Tool evaluationDid it read the correct site record and write the correct claim month?
- E4Confidence calibrationDo low-confidence findings actually attract more official corrections?
- E5Slice evaluationHow far apart do exception rates sit across specific site types?
- E6Business outcomeHow many exceptions were resolved, and how many surfaced only in review?
Floor — the claim the district answers for
Failure modes
Where each failure originates in the agent
Seven failure modes, each sitting at the stage where it first appears.
Agent lifecycle
Direction of processing →
01 · Retrieval
1 mode
- HK-03
Match never flipped
A direct-certification match never reached the point of service.
Stage gathers
Counts, eligibility, enrolment and the attendance factor
02 · Reasoning
2 modes
- HK-04
Cleared with no reason
An exception is cleared with no reason recorded.
- HK-06
Carry-over run past expiry
Eligibility is claimed past the carry-over window.
Stage proposes
Counts by site and day, edit checks, confidence
03 · Tool / write
2 modes
- HK-02
Chase logged as no answer
A household that tried is filed as one that never replied.
- HK-05
Records gone before review
Records are lawfully gone; the state reconstructs them.
Stage writes
Only where write access and approval policy allow it
04 · Output
1 mode
- HK-01
Non-reimbursable meals counted
A mis-set register key adds adult or second meals to the claim.
Stage returns
The claim the official signs, and stands behind
05 · Change / Version
1 mode
- HK-07
Silent edit-rule drift
A configuration change loosens the edit check and open months do not re-run.
Stage tracks
Model, prompt, edit rules and sampling config
- Sev-1 · a determination or notice goes out
- Sev-2 · a wrong count reaches the signed claim
- Sev-3 · source degrades, pack goes to review
Affected slices
Most exceptions sit in one type of site
A district-level edit-check figure can read clean while one type of site holds most of the unresolved exceptions. Nestack reports the exception rate by site type, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| High-mobility elementary sites | 7.8% | 3.6× | Review | |
| Shared-kitchen satellite sites | 5.6% | 2.6× | Review | |
| CEP sites mid-cycle | 3.3% | 1.5× | Watch | |
| Established single-kitchen sites | 1.3% | 0.6× | Normal |
Bar: exception-rate lift vs. the established-single-kitchen baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
A cleared flag ends as a test
A cycle closes when the failed edit check is a regression case. That suite is what the next claim assembled is measured against.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Exception rate rises in one site type.
- 02
Diagnose
The site whose meal count was right every day and provable for three years is read back through the edit checks until the cause narrows to one.
- 03
Improve
The fix carries a number, and the serving sites that forced it come with it.
- 04
Verify
Nothing goes out until every touched site case has passed again.
- 05
Learn
It is kept permanently, and the edit rules are amended alongside it.
Learn → Detect
The return edge. The next month is assembled against a suite one case longer.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources, claim workflow, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Claim-month workflow discovery and refusal boundary.Week 1
- 02Point-of-service and eligibility sources.Week 1, Week 2
- 03Edit-check, carry-over and sampling rule mapping.Week 1, Week 2
- 04Count and eligibility ingestion.Week 2
- 05Edit-check logic and count binding.Week 2, Week 3
- 06Confidence scoring and finding routing.Week 3
- 07Food authority official review workflow.Week 3, Week 4
- 08Point-of-service and portal integration.Week 5
- 09Edit-check and verification cases.Week 4, Week 5
- 10Guardrails and claim controls.Week 4
- 11Claim-trail instrumentation.Week 5, Week 6
- 12Deployment, documentation and Agent Care handover.Week 6
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel
Final scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.
| Capability✓ in scope · — not at this tier | PilotOne district, one claim month | ProductionProduction claim and POS systems | AdvancedMultiple districts / site types |
|---|---|---|---|
| Introduced at Pilot | |||
| Claim assembly to your site rules | ✓ | ✓ | ✓ |
| Food authority official review | ✓ | ✓ | ✓ |
| Edit-check baseline | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Reporting by site | — | ✓ | ✓ |
| Review workflow in your systems | — | ✓ | ✓ |
| Approved claim write-back | — | ✓ | ✓ |
| Point-of-service integration | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-state claim rules | — | — | ✓ |
| Multi-stage district approvals | — | — | ✓ |
| High site count | — | — | ✓ |
| Multi-site claim controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on integrations, workflow complexity, site count, review controls and deployment requirements.
- Separate from build
- Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bring
We build with it
01Your sites, your registers and how counts arrive
→Count and eligibility ingestion, and site binding
Week 102Representative claim months from a past year
→Edit-check baseline, count normalisation and rule binding
Week 203Your attendance factor and site-level rules
→Edit-check, carry-over and sampling rule mapping
Week 104Access to relevant APIs, feeds or exports
→Point-of-service, eligibility and portal review, then integration setup
Week 205Claims you would not want reviewed
→Sample cases and the evaluation suite
Week 406What no signed claim may estimate
→Confidence scoring, finding routing, guardrails and claim controls
Week 307The named official who signs the claim
→Official review workflow, then pilot and production validation
Weeks 5–6
Nothing else is required
Deployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
The plan is drawn from the weeks the job truly needs, so the pair sitting in week five is real.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Claim workflow discovery, rule mapping and the refusal boundary
W2Source integration and the edit-check baseline
W3Count binding, carry-over checks and review controls
W4Evaluation suite, sample cases and failure-mode testing
W5Point-of-service and portal integration, a pilot month and targeted corrections
W6One claim month run under the food authority official, then Agent Care handover
- Reading the band
- A bar sits only on the weeks its work is named in. The fifth-week overlap is real, not padding.
- At the end of W6
- The final checks clear on live sites and monitoring moves to Agent Care.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Education AI agent
Build a claim agent around the official who signs.
Show us your sites, your registers and who signs the claim. How long does the record have to last — longer than 7 CFR 210.23 keeps it, because 7 CFR 210.18 can reach past it, and where the data has gone 7 CFR 210.19 has the state reconstruct your meal accounts.
- Typical delivery
- 4–6 weeks
- Deployment
- Your POS and claim stack
- Official signs
- Never cuts off
- After launch
- Agent Care available
Nestack Agents · Child nutrition claim and verification
AGT-ED-12 · Agent Care available after launch