The CFO of a behavioral health enterprise opens the vendor quote and sees a clean number: per-clinician licenses, annual contract, modest discount at volume. She is looking at the tip of an iceberg and being asked to budget for the whole ship's passage around it. The license line is the visible tenth of a clinical AI investment; below the waterline sit integration, governance, training, and evaluation, the four lines that determine whether the visible tenth produces anything at all. And beside the iceberg floats the number nobody puts in the quote: the cost of doing nothing, denominated in clinician turnover, denial rate, parity exposure, and missed value-based upside. This lesson teaches you to build the Three-Year Capital Plan for a behavioral health enterprise: all five spend lines budgeted in full across three fiscal years, paired with a Cost-of-Inaction Model rigorous enough to put in front of a finance committee. By the end you will have built both, in one document, with your own numbers in every cell.

The Iceberg Budget: Why the License Line Misleads

Hold the controlling analogy through every section: a clinical AI investment is an iceberg, and the license quote is the tenth above the waterline. Enterprises that budget only the visible tenth run aground on the nine-tenths below: the EHR integration that was "included" until the scoping call, the governance committee whose hours appear on no invoice, the training curriculum that determines whether adoption reaches the saturation where enterprise tools earn their value, and the evaluation program without which the board's quarterly report is anecdote. The Eleos Health deployment across the CalMHSA/Streamline footprint, roughly 27,000 clinicians at full saturation across California's public-sector behavioral health workforce, is the standing demonstration that the asset class performs at enterprise scale; it is also the standing demonstration that saturation is a change-management program, not a license purchase, which is the polite way of saying the spend below the waterline is where the outcome is decided.

The multi-year structure matters as much as the line items. A one-year budget treats clinical AI as a software subscription; a three-year capital plan treats it as infrastructure, which is what it is. Year one is heavy below the waterline (integration, governance build-out, foundational training) and light on visible return. Year two rebalances: license spend scales with adoption, integration drops to maintenance, training shifts from foundational to phase-specific, and the first measured returns (note-closure lag, denial rate) appear against baseline. Year three is where the plan either compounds or gets quietly defunded: evaluation spend peaks as predictive capabilities arrive under stricter governance, and the return lines (turnover, parity posture, value-based eligibility) mature. A finance committee shown only year one sees cost; shown three years, it sees the shape of an infrastructure investment, which is the shape it actually has.

One framing sentence belongs at the top of the plan, borrowed from the board narrative of the previous lesson: declining this investment is also a decision with a price. The capital plan's second half, the Cost-of-Inaction Model, prices it, and a plan that arrives without that half is asking the committee to compare a known cost against an unknown one, a comparison the known cost always loses.

Line One: Licenses, Priced at the Tier That Actually Protects You

The license line is the simplest and still gets mispriced, because the cheapest tier is rarely the lawful one. The behavioral health vendor market in 2026 (Mentalyc past 30,000 clinicians, Upheal on high-volume telehealth caseloads, Eleos in the enterprise and public sector, Heidi arriving from primary care, Twofold among psychiatrists, TherapyNotes and Therapy Brands native in the EHR layer, Blueprint embedded in SimplePractice for measurement-based care) prices in tiers, and the tier determines the terms: whether a BAA is signed at your level, what the subprocessor list looks like, whether your clients' data trains models, what retention and deletion commitments exist. Jordan's near-miss in Sacramento, an EHR vendor whose subprocessor list included a model provider that does not sign a BAA at the purchased tier, is the budgeting lesson in one sentence: the compliant tier is the real unit price, and a capital plan built on the marketing tier is understated by design.

Budget the license line in three movements across the three years. Year one covers the foundation-phase tools: the documentation scribe and the measurement-based care platform, sized to your phased adoption curve rather than to full headcount on day one. Year two scales to saturation and adds the productivity perimeter (treatment-plan support, prior-authorization drafting, claims scrubbing). Year three adds the phase-three capabilities, predictive decision support and population-health analytics, gated on the exit criteria of your transformation map, which means year-three license dollars are conditionally committed: the plan shows them, the finance committee approves them, and the governance committee releases them only when the foundation has passed inspection. Conditional commitment is the financial expression of staging, and committees respect it because it caps exposure without re-opening the whole plan annually.

Line Two: Integration, the Largest Number Nobody Quotes

Integration is the first big mass below the waterline. A scribe that does not write into the EHR's structured fields produces PDFs nobody can compute on; an MBC platform whose PHQ-9 scores live in attachments rather than fields feeds nothing downstream; a predictive layer with no clean longitudinal data source is the dashboard displaying its own missing inputs. The integration line budgets the unglamorous plumbing: EHR interface work (whether SimplePractice, TherapyNotes, Valant, or a Therapy Brands stack), single sign-on and identity management, data mapping so outcome scores land as structured data, and, where SUD treatment is in scope, the segmentation work that keeps 42 CFR Part 2 data inside its consent and redisclosure boundaries under the 2024 final rule, a constraint that is architecture, not paperwork.

Budget integration front-loaded and honest: the year-one number is typically the plan's largest single line after licenses, and pretending otherwise produces the mid-year supplemental request that erodes committee trust. Years two and three carry maintenance plus the integration cost of each newly gated capability. Two disciplines keep this line defensible. First, scope integration in writing before contract signature, because "API available" and "integrated into your workflow" are different products at different prices. Second, put a data-portability clause in every contract and a line in the plan for exit costs: the vendor-acquisition scenario, in which your tool's new owner changes data practices mid-contract, is a budgeted contingency, not a surprise, and an enterprise that cannot afford to leave a vendor has not bought software, it has been bought.

Line Three: Governance, the Cheapest Insurance on the Plan

Governance is the line finance committees are most tempted to zero out because it buys no feature, and it is the line whose absence is most expensive. Budget it as real hours and real money: the AI governance committee's standing time (clinical leadership, compliance, IT or security, a frontline clinician, Part 2 fluency where SUD data is in scope), vendor diligence and annual BAA and subprocessor re-review, the incident process and kill-switch drills, consent-architecture maintenance as capabilities escalate, and external legal review when the regulatory perimeter moves. And in 2026 to 2028 it moves constantly: the Illinois WOPR Act and Nevada AB 406 drawing lines around AI in therapeutic decision-making, New York's AI companion law, the Colorado AI Act effective June 30, 2026, and the software-as-a-medical-device question that must be asked before any tool edging toward diagnosis or recommendation is procured.

The governance line also funds the bright line that protects every other line: the standing policy that AI never scores the CSSRS, never assigns a risk level, never makes the duty-to-protect or mandated-report determination, and that the clinician signs every note as a legal attestation after reading every word. That policy costs drafting hours, training time, and audit sampling, all budgetable, and it is the difference between a defensible program and a deposition exhibit. Price governance at a visible percentage of the total plan and defend it as insurance: the committee that asks why governance costs so much should be shown what one OCR inquiry, one board complaint, or one WOPR-class enforcement action costs, and reminded that the unsanctioned tools already inside the building (Jordan found twelve of twenty-five clinicians on a free scribe with no BAA) are generating the risk today, uninsured.

The license quote is the tenth of the iceberg above the waterline; integration, governance, training, and evaluation are the nine-tenths below it, and the cost of inaction is the ship you are steering either way.

Lines Four and Five: Training and Evaluation, Where Adoption and Proof Are Bought

Training is the line that converts licenses into adoption, and adoption is where enterprise tools earn their value. Budget it as a curriculum across the three years, not a go-live event: foundational training in year one (prompt discipline, pre-signature review, consent delivery in the clinician's own voice, the bright line), phase-specific training in year two (payer-defensible documentation, the verifiable details AI cannot supply: time-in-session minutes for a 90837, the specific PHQ-9 delta, the modality named in session), and phase-three training in year three (interpreting and overriding predictive flags without automation bias). Supervisors train one phase ahead of supervisees, because the supervisor's signature carries the supervisee's AI use, and the supervision addenda that make that oversight explicit are a training deliverable with hours attached. Underfunding this line produces the silent failure mode: licenses purchased, adoption stalled at forty percent, and a finance committee concluding the asset class does not perform when what did not perform was the budget.

Evaluation is the line that makes every other line provable. It funds the pre-launch baseline (documentation hours per clinician per week, note-closure lag, denial rate on the top three CPT codes, trailing turnover, MBC completion, clinician-burden survey), the quarterly re-measurement, note-quality audit sampling, and, in year three, the stricter protocols predictive tools require: false-positive and false-negative review on a schedule, and override-pattern monitoring, where a flag everyone ignores is a broken instrument and a flag nobody overrides is an automation-bias alarm. Evaluation is also the committee's own protection: the quarterly baseline-delta report it funds is what lets the finance committee distinguish a transformation from an expenditure, and a plan that skimps here is asking to be judged on anecdote in year three, which is how good programs get defunded and bad ones survive.

The Cost-of-Inaction Model: Pricing the Decline in Four Currencies

Now build the other half of the document, the half the vendor quote never includes. The Cost-of-Inaction Model prices the alternative scenario, the enterprise that stands still from 2026 to 2028, in four currencies, each computed from your own data. Currency one, clinician turnover: take your trailing-twelve-month turnover rate and your fully loaded replacement cost per clinician (recruiting, credentialing, ramp-up, lost caseload continuity), then model the documentation-burden share of departures conservatively; even a modest attributable fraction, multiplied across an enterprise workforce and three years, is usually the model's largest number. The labor market context does the rest of the argument: clinicians increasingly compare employers on documentation environment, and the APA Practitioner Pulse Survey's one-in-three monthly AI adoption means your recruits are comparing you to AI-supported alternatives already.

Currency two, denial rate: take the current denial rate on your top three CPT codes against annual claim volume, hold it flat (or trend it with payer tightening) across three years, and add the audit tail: high-frequency 90837 reviews against thin notes, with recoupment exposure of the $14,200-per-clinician class. Currency three, parity exposure: appeals not filed or lost for lack of documented medical necessity and outcome trajectory, in the environment where MHPAEA's 2025 to 2026 federal non-enforcement posture leaves state laws (California SB 855, New York's Timothy's Law) carrying the enforceable weight; price it as the revenue from denied behavioral health care that documentation-strong organizations recover and you do not. Currency four, missed value-based upside: the contracts, pilots, and county or MCO quality awards that flow only to outcome-reporting organizations, priced as the eligible revenue your enterprise cannot bid on without MBC infrastructure, with the Lyra Health and Spring Health pattern as the market proof that payers and purchasers already buy outcome stories. Sum the four currencies per year, line them against the capital plan's five spend lines, and the comparison the finance committee actually needs is on one page: the cost of the iceberg's nine-tenths against the cost of the ship that stayed in port.

Governing the Plan Over Three Years: Gates, Re-forecasts, and the Defunding Trap

A three-year plan survives contact with three years of reality only if its governance is built in. Tie every conditional commitment to the transformation map's exit criteria: year-two scale-up releases when phase-one criteria certify, year-three predictive spend releases when the foundation passes inspection, and the finance committee sees the certification, not just the request. Re-forecast semi-annually against the evaluation line's measured deltas, adjusting the model's assumptions in both directions, because a cost-of-inaction model that only ever grows reads as advocacy and loses the committee. And budget the kill-switch scenario explicitly: if governance suspends a tool, the plan shows the substitution cost and the contractual off-ramp, which is what the data-portability clause and exit line were for.

Watch for the defunding trap in year two. The below-the-waterline lines (governance, training, evaluation) produce no demos, and a budget cycle under pressure will reach for them first, which quietly converts a governed transformation into an ungoverned license subscription, the exact drift the whole plan exists to prevent. The defense is the structure you built: governance priced as insurance with the incident math attached, training tied to the adoption numbers in the quarterly report, evaluation tied to the committee's own ability to see returns. The plan's last page should say plainly what the previous lesson's board narrative said: the risks arrive whether or not the spend is approved; only the controls are optional. An enterprise that funds licenses and defunds controls has chosen the iceberg's visible tenth and the collision both.

The Applied Problem: Build the Three-Year Capital Plan with the Cost-of-Inaction Model

Your artifact is the Three-Year Capital Plan with Cost-of-Inaction Model: one document, two halves. The first half is a five-line budget (licenses, integration, governance, training, evaluation) across three fiscal years, with conditional commitments tied to exit criteria. The second half prices inaction in the four currencies (turnover, denial rate, parity exposure, missed value-based upside) from your own data, year by year, on a single comparison page. Build it in four steps.

Step one, pull your inputs: current vendor quotes at the compliant tier (BAA signed at your level, subprocessor list reviewed), an integration scoping estimate in writing from your EHR vendor, governance hours priced at loaded cost, training hours per clinician per phase, your evaluation baseline numbers, and the four cost-of-inaction inputs (turnover rate and replacement cost, denial rate and claim volume, parity appeal history, value-based opportunities you could not bid). Step two, draft with your AI assistant, no client data: "Build a three-year capital plan for a [size]-clinician behavioral health enterprise adopting clinical AI in a staged transformation. Five budget lines per year: (1) licenses at the compliant tier, phased to adoption, with year-three predictive spend conditionally committed on exit criteria; (2) integration, front-loaded in year one, including EHR structured-field work, SSO, 42 CFR Part 2 segmentation where applicable, and an exit-cost contingency; (3) governance: committee hours, annual BAA and subprocessor re-review, incident process and kill-switch drills, consent-architecture maintenance, external legal review; (4) training as a three-year curriculum with supervisors trained one phase ahead; (5) evaluation: baseline, quarterly re-measurement, note-quality audits, year-three predictive protocols with override monitoring. Then build a cost-of-inaction model in four currencies: clinician turnover, denial rate with audit tail, parity exposure under state-enforceable parity law, and missed value-based upside. Output a one-page comparison table, plan total versus inaction total, per year. Use placeholder cells I will fill; do not invent numbers."

Step three, run the verification pass: fill every placeholder from your own inputs and delete any number the model invented; confirm the license line uses the compliant tier, not the marketing tier; confirm year-three spend is marked conditional on certification; check that the turnover currency uses a conservative attributable fraction you can defend out loud; and walk the comparison page with your CFO before any committee sees it, because the model must survive the person who will be cross-examined on it. Step four, install the governance: calendar the semi-annual re-forecast, attach the exit-criteria certifications to each conditional release, and write the defunding-trap paragraph onto the last page. Done looks like this: a finance committee can see, on one page, what three years of governed adoption costs, what three years of standing still costs, and exactly which gate releases each conditional dollar.

Key Takeaways

• A clinical AI investment is an iceberg: the license quote is the visible tenth, and integration, governance, training, and evaluation are the nine-tenths below the waterline where the outcome is actually decided; a plan that budgets only licenses is understated by design.
• Budget licenses at the compliant tier, not the marketing tier: the tier determines the BAA, the subprocessor list, model-training terms, and retention commitments, and year-three predictive spend is conditionally committed, released by governance certification against exit criteria rather than by the calendar.
• Integration is typically the largest unquoted number: EHR structured-field work, SSO, data mapping so MBC scores are computable, 42 CFR Part 2 segmentation where SUD data is in scope, plus a written scope before signature and a budgeted exit-cost contingency for the vendor-acquisition scenario.
• Governance is the cheapest insurance on the plan: committee hours, annual BAA and subprocessor re-review, kill-switch drills, consent-architecture maintenance, and external legal review as IL WOPR, NV AB 406, NY's AI companion law, and the Colorado AI Act move the perimeter; it also funds the bright line that AI never scores risk instruments or makes protective determinations.
• Training buys adoption (a three-year curriculum, supervisors trained one phase ahead) and evaluation buys proof (baseline, quarterly deltas, note audits, override-pattern monitoring); underfunding either produces the silent failure where the committee concludes the asset class failed when the budget did.
• The Cost-of-Inaction Model prices standing still in four currencies from your own data: clinician turnover (usually the largest number), denial rate with the $14,200-class audit tail, parity exposure under state-enforceable laws like CA SB 855 and NY Timothy's Law, and missed value-based upside that flows only to outcome-reporting organizations.
• Govern the plan itself: conditional releases gated on certifications, semi-annual re-forecasts that move in both directions, a budgeted kill-switch scenario, and explicit defense against the year-two defunding trap, because an enterprise that funds licenses and defunds controls has chosen the visible tenth and the collision both.