AI in education,
shipped on your LMS — not on a slide deck.

An `ai in education` development company like ours ships production AI workflows on your LMS + assessment surface — not slide decks, not pilots that die at procurement. The day-to-day work: scope which workflow returns instructor time without crossing the academic-integrity boundary (most often AI tutoring, AI essay grading with teacher review, AI quiz or lesson-plan generators, special-ed assist, or an LMS-integrated academic advisor), get the FERPA + COPPA + state-privacy attestations on file, build the integration against your Canvas / Blackboard / Moodle / Schoology / Google Classroom / PowerSchool stack via LTI 1.3, pick the right model per workflow (Sonnet 4.6 for tutoring quality and essay critique, Haiku 4.5 for high-volume Q&A, GPT-5.4-mini for structured quiz generation), bake in cohort-level bias auditing and WCAG 2.1 AA accessibility, ship behind a teacher sign-off flag, then operate the workflow long enough to prove cost-of-ownership before scaling. We do not sell a product — we ship one workflow at a time and report cost-per-decision monthly. First workflow live in 6–8 weeks; full pilot $10–25K fixed.

Yes — the operator-grade specifics. (1) FERPA: minimum-necessary data, education-records retention per institution policy, audit log on every inference (request ID, model, retrieved-context references, response, teacher who reviewed). (2) COPPA: for K-12 under-13 users, verifiable parental consent on file before the workflow activates, no behavioral advertising on the AI pipeline, minimum-necessary data collection, deletion-on-request paths. (3) State student-privacy laws: NY Ed Law 2-d (data inventory, breach-notification SLA, deletion paths, no secondary use), California SOPIPA (no targeted advertising, no profile-building for non-educational purposes), Illinois SOPPA (district-side contract with explicit data-use limitations). We sign your district's data-privacy contract before any pilot — and we will turn down the engagement if the contract conflicts with our infrastructure providers' BAA-equivalent terms. (4) IDEA: special-ed workflows ship only inside the IEP team's loop; AI does not set goals or log progress autonomously. What we do not claim: SOC 2 Type II or ISO 27001 certification as a product. If your procurement team needs SOC 2 as a hard gate, we'll say so up front.

`Ai tutoring` ships as 1:1 worked-example scaffolds inside the LMS assignment view — never as autonomous answer delivery. The pattern: a subject-specific tutor (Sonnet 4.6 for quality, Haiku 4.5 for district-scale volume) reads the assignment context via LTI 1.3 launch, applies the instructor-set difficulty band, runs Socratic scaffolds (next-step hint, worked example on a similar-but-not-identical problem, conceptual reframe), and writes a weekly mastery report to the teacher's dashboard. Hard guardrails baked in: the AI refuses to solve the assessed problem outright, refuses to write the essay for the student, refuses to leave the subject scope, and hard-handoffs on mental-health phrases or graduation-risk flags. Teacher visibility on every session is the design point — `ai tutor for students` only works if the teacher knows what their students asked the AI and how the AI responded. We've seen 40%+ tutor-attached attempt rates inside Canvas / Blackboard in the audits we've run.

Yes — with a teacher in the loop and a cohort-level bias audit on every grading cycle. The pattern: Sonnet 4.6 drafts a rubric-coded score with per-criterion rationale, the teacher reviews and overrides, the final mark is the teacher's. The bias audit runs at cohort close: scores grouped by demographic flag (where the institution has consented to that analysis) compared against the cohort distribution; any statistically significant skew triggers a rubric-recalibration cycle before the next term. AI text detectors are a different problem — we will not ship `ai detection` as the sole evidence in an academic-integrity hearing because the false-positive rates disproportionately flag non-native English writers and neurodivergent students. The honest answer on `ai for grading papers` is: AI saves the teacher 50–65% of grading time on rubric-anchored work; the teacher saves the AI from shipping a bias-skewed score; and the audit log is the joint contract between the two.

Yes — all six, via LTI 1.3 (and Google Classroom's native Classroom API where LTI doesn't reach). Quick recap of the per-LMS failure modes (covered in §6 above). Canvas: keysets rotate — fetch JWKS dynamically, never cache past TTL; AGS lineitems must be pre-created via Deep Linking 2.0 or admin-side. Blackboard Ultra: role claims sometimes return in legacy format; normalize before reading; score scale defaults to 0-100. Moodle: `context_type` is course vs CourseSection — handle both; grade scale is configurable per gradebook item. Schoology: rate-limit at 50 req/min on free tiers; AGS support is limited compared to Canvas. Google Classroom: no LTI NRPS or AGS — use the Classroom API directly with `coursework.studentSubmissions.patch`. PowerSchool: iframe sandbox blocks WebSockets in some configs — ship a long-poll fallback. `Canvas ai` and `khanmigo`-style integrations are not magic — they're 90% LTI 1.3 plumbing (OIDC + JWT + AGS) and 10% model choice. Picking the right model is the easy part; the LTI integration layer is the engagement.

Two cost buckets, transparent. Model layer: an `ai tutor` on Sonnet 4.6 with prompt caching warm runs roughly $0.008–$0.018 per session (a session = a multi-turn conversation around one assignment); a typical middle-school math student opens 30–60 tutor sessions per term, so model cost lands at $0.24–$1.08 per student per term. An `ai essay grading` workflow runs $0.04–$0.10 per essay on Sonnet 4.6; a first-year writing student submits 6–10 essays per term, so model cost lands at $0.24–$1.00 per student per term. Infrastructure layer (LTI launch, FERPA audit log, AGS write-back, bias-audit job): typically $0.01–$0.04 per session. On a 1,000-student pilot for either workflow, model + infra sits in the $200–$1,200/month range. The honest cost ceiling is the engineering: getting the LTI 1.3 integration right, the FERPA audit log right, the cohort bias audit running on schedule, the teacher-in-loop UI to a place teachers actually trust. That work is the engagement, not the per-decision spend.

Both vendors have explicit education positioning, and the $38.32 CPC peak on `claude for education` and $78.56 CPC peak on `openai education` tell you institutional buyers are real and active. OpenAI Education / ChatGPT Edu: institutional admin console, SSO with district IDs, COPPA-attestation paths for K-12 deployments, model fine-tuning on district curriculum where licensed, and (importantly) an enterprise tier that's actually data-policy-aligned for student data — consumer ChatGPT is not student-data-safe by default. Claude for Education (Anthropic): institutional access via the Anthropic API on enterprise terms, strong long-context handling for curriculum-aligned use cases, Constitutional-AI safety frame that aligns with academic-integrity zones, and (per Anthropic's published education-partner work with Northeastern, LSE, Champlain, and others) a clear posture on student-data minimization. Our position as a model-agnostic `ai in education` company: we pick the model per workflow. Sonnet 4.6 wins where pedagogical reasoning or rubric critique matters; GPT-5.4-mini wins on high-volume structured quiz generation; Haiku 4.5 is the district-scale tutoring routing default; GPT-5.4 is the long-context default for IEP-synthesis or capstone-feedback work.

Four places we'll say no — covered in §8 above and worth repeating. (1) Autonomous grading of credential-bearing work — final grades stay with the teacher; bias audits run on every cohort. (2) K-12 under-13 workflows without COPPA + state-privacy attestation on file — NY Ed Law 2-d, California SOPIPA, Illinois SOPPA, and equivalents in your state govern the data-use contract, and we won't ship until it's signed. (3) Academic-integrity zones — summative state exams, bar / licensure / board exams, capstone defenses; AI is BLOCKED, proctored, lockdown-browser-enforced. We also refuse to ship `ai detection` as the sole evidence in an integrity hearing because detector false-positive rates disproportionately flag non-native English writers and neurodivergent students. (4) Special-ed workflows that bypass IDEA or Section 508 / WCAG 2.1 AA — the AI sits inside the IEP team's loop, never sets goals, never logs progress autonomously, and every output is accessible. Beyond those four: any workflow where the academic-integrity policy hasn't been mapped against the Bloom × AI-band rubric is a workflow we'll refuse until that mapping is done.

No — augment, not replace. The AI we build for K-12 districts, higher-ed institutions, and adult-ed / workforce-training programs is teacher-in-loop or instructional-designer-in-loop on every consequential decision. Tutor sessions write a weekly mastery report to the teacher; essay grades draft with per-criterion rationale a teacher overrides; quiz items generate but a teacher approves before publish; lesson plans draft but a teacher edits before delivery; special-ed scaffolds suggest but the IEP team decides. Time saved (30–45 min per lesson plan, 50–65% of essay grading time, 40–55% of routine advising ticket time) goes back into student-facing work — the only ROI metric that matters in education. The realistic claim a `ai in education` development company should make: teachers keep doing the pedagogy, the assessment, the relational work; the AI handles the drafts, the lookups, the scaffolds, and the routine routing. If a vendor is selling `replace your teachers with AI` energy, that's the deck to pass on — `ai in classrooms` works because it returns teacher hours, not because it removes teachers.

Three tiers, pricing-locked across the cluster. (1) AI-in-education audit: $3K fixed, 1–2 weeks. We shadow academic-affairs + IT + instructional-design + (where relevant) special-ed coordinators, map your assessment surface against the Bloom × AI-band rubric, rank candidate workflows by instructor hours returned × time-to-ship × policy risk, and deliver a 90-day roadmap with per-workflow cost bands and an honest "these won't pay back yet" list. FERPA + COPPA + state-privacy + IDEA + accessibility posture reviewed before code. (2) Pilot to production: $10–25K fixed, 4–8 weeks. One workflow shipped end-to-end via LTI 1.3 on your LMS, teacher-in-loop, bias-audited, with FERPA audit-log + accessibility regression suite tested and a walk-away point at week 6 — if the metric won't move, we stop before production hardening and you don't pay phase 2. (3) Continuous AI-in-education team: from $5K/month, no annual contract. Embedded PM + AI engineer + instructional-design analyst shipping the next workflow on your roadmap, with per-cohort bias audits and cost-of-ownership reports each month. Most `ai in education` engagements we run start with the audit, ship the first workflow on the pilot, and move to monthly for workflows two through five. Cost-per-decision reported monthly on every shipped workflow — no per-decision number, no engagement.