Jordan's roadmap is sequenced, the documentation phase is scoped, and now the question lands on the owner's desk in its rawest form: what does this cost, and against what? Twenty-five clinicians at roughly $59 a month is over $17,000 a year before training, integration, or a single enterprise feature, and the EHR vendor is whispering that its native AI is "basically free." Every one of those numbers is designed to be compared against zero, and zero is the wrong denominator. The right denominator is what the practice is already paying: five to ten unpaid documentation hours per clinician per week, a denial pattern bleeding revenue on claims the clinical work already justifies, and the quiet certainty that replacing even one departing clinician costs $25,000 to $60,000 in onboarding plus six to nine months of suboptimal productivity. This lesson teaches you to build the AI scribe ROI model for a therapy practice properly: per-clinician licensing (the Mentalyc, Upheal, Heidi, Twofold category) versus Eleos-style enterprise versus EHR-native AI, each priced against the unpaid hours, the denials, and the retention dollars, in a budget model an owner can defend to a board and a skeptic can audit.

The Wrong Question and the Right One

The wrong budgeting question is "how much does the tool cost?" The right question is "what is the practice currently paying for not having it, and which licensing structure captures the most of that loss for the least new risk?" Practices ask the wrong question because the tool's cost is visible and invoiced while the current cost is invisible and uninvoiced: no line item on any P&L says "unpaid clinician documentation, 7 hours/week x 25 clinicians," and no 1099 anywhere records the evening hours a paneled clinician donates to her own chart. But invisible is not small. A clinician earning $97 from Headway for a 90834 and then spending forty unpaid minutes on the note is working a meaningful slice of her week at functionally $0 an hour, and the practice is harvesting the consequences in thin notes, denial exposure, and turnover risk.

Carry one analogy through this lesson: the budget model is a leak audit. A building manager deciding whether to pay for plumbing does not start from "pipes cost money"; she starts by measuring what the leaks already cost: the water bill, the rotting floor, the tenant who moves out. Your practice has three measurable leaks, and your readiness assessment already metered them: the unpaid documentation hours (the water bill, paid daily and invisibly), the documentation-sensitive denial rate (the rotting floor, structural and compounding), and clinician turnover (the tenant who leaves, the single most expensive event in the building). The licensing decision is the plumber's quote, and a quote only means something when held against the metered leaks.

This framing also disciplines the vendor conversation, because vendors price against each other while your model prices against the leak. A vendor whose pitch is "we are $20 a month cheaper than the competitor" is answering the wrong question. The model you build in this lesson answers the right one, in numbers your billing manager can reproduce.

Metering the Leaks: The Three Numbers on the Cost Side

Start with unpaid documentation hours, the largest and most measurable leak. Your readiness assessment sampled time per note; now annualize it. Take the measured weekly unpaid documentation hours per clinician (the playbook's range is five to ten for a full caseload; use your own measured number), multiply across the workforce, and price the hours two ways. The conservative pricing is replacement value: what those hours would cost if you had to pay someone for them. The aggressive pricing is opportunity value: what those hours would earn if even a fraction converted to billable sessions, at your actual blended reimbursement. Hold both numbers, label them clearly, and let the conservative one carry the headline, because a skeptical board member will attack the aggressive assumption and you want the model to survive without it. The honest caveat belongs in the model too: an AI scribe does not eliminate documentation time, it compresses it, because the clinician still reads every word before signing, the signature being a legal attestation. Model the saving as a measured reduction (your pilot's time-per-note delta), never as the whole burden.

Second, the denial leak. From the readiness assessment you have denials by payer with stated reasons. Isolate the documentation-sensitive subset: Jordan's Anthem prior-auth denials for lack of medical necessity documentation, where the PHQ-9 trajectory already justifies the care, are the canonical example, and every one of those denials has a claim value, an appeal cost in staff hours, and a write-off rate. The model credits AI only for the documentation-sensitive slice, and only at the improvement rate your day-60 and day-180 roadmap readings actually show; a model that assumes denials go to zero is a vendor slide, not a budget.

Third, the retention leak, the largest single-event cost in the model. Replacing a clinician costs $25,000 to $60,000 in onboarding plus six to nine months of suboptimal productivity, which means a 25-clinician practice losing even two clinicians a year to burnout is carrying a six-figure leak before counting the disruption to clients and panels. The model handles retention conservatively: AI does not guarantee retention, so the model expresses it as avoided-replacement scenarios (if documentation relief retains one clinician who would otherwise have left, the program banks $25,000 to $60,000 plus the productivity ramp) and lets the satisfaction-pulse and turnover data from the roadmap's day-60 and day-180 reads tell you which scenario you are living in.

Option One: Per-Clinician Licensing (Mentalyc, Upheal, Heidi, Twofold)

The per-clinician model is the one your twelve self-adopters already discovered: a monthly subscription per clinician, in the territory of Carmen's $59-a-month Upheal plan, purchased in minutes, cancelled in minutes. Its budgeting profile: linear cost (25 clinicians costs roughly 25 times one clinician), minimal integration expense, fast deployment, and per-seat flexibility, you license the volunteers in phase one and add seats as the rollout expands, which maps perfectly onto the roadmap's 30/60/90 expansion gates. For a group practice in the documentation phase, this is usually the cheapest way to buy the first leak repair, and its month-to-month nature keeps the kill switch real: a tool that fails its pilot can actually be cancelled, which disciplines the vendor.

The hidden costs are administrative and contractual, and they scale with headcount. Twenty-five individual seats means twenty-five onboarding paths unless you negotiate a group plan; it means verifying that the practice-level BAA covers every seat (a clinician's personal subscription, like Carmen's, is her BAA problem and the practice's exposure, which is exactly the ungoverned pattern the readiness assessment flushed out); and it means the per-seat price that looked small at five clinicians compounds at scale: 25 seats at roughly $59 lands above $17,000 a year, every year, with no equity in anything. The budgeting discipline for this tier: price the group contract, not the retail seat; require the BAA, the zero-data-retention posture, and the subprocessor disclosure at the practice level; and book the admin time (license management, onboarding, the AI-use log) as a real cost line, because it is one.

Where this option wins on the model: small to mid groups, phased rollouts, practices whose EHR fit score makes deep integration unrealistic this year, and any practice that wants its first-year spend fully reversible. Where it strains: 40-plus clinician agencies where per-seat math crosses over enterprise pricing, CCBHC environments whose reporting use cases need deeper data integration than a standalone scribe offers, and practices that need supervision, analytics, or compliance tooling above the individual note.

Option Two: Enterprise Licensing (the Eleos Category)

Enterprise behavioral health AI, the category Eleos Health defines, prices differently because it sells a different thing: not a per-seat scribe but an organizational deployment, with EHR integration, workforce-level analytics, supervision and compliance surfaces, and an implementation project. The reference point for scale is public: Eleos's CalMHSA / Streamline deployment runs across the California public-sector behavioral health workforce, on the order of 27,000 clinicians at full saturation, which tells you the category is built for agencies, CMHCs, and CCBHCs, organizations whose problems are workforce-shaped, not seat-shaped.

The budgeting profile inverts the per-clinician model. Costs are front-loaded: implementation, EHR integration work, training time across the workforce, and a contract measured in years rather than months. The value is also differently shaped: an enterprise deployment can reach the use cases a standalone scribe cannot, the CCBHC reporting burden with its PPS daily encounters and quarterly narratives, supervision-level visibility into documentation quality, and organizational analytics that feed the day-180 program judgment. For an agency whose roadmap scored those reporting use cases as unexpectedly high ROI, the enterprise tier may be the only option that actually captures the leak, and the per-clinician comparison is apples to oranges.

The budgeting discipline for this tier is contract skepticism. Multi-year commitments mean the kill switch is weaker, so the pilot and milestone gates from the roadmap must be written into the contract as performance conditions, not left as internal hopes. Implementation timelines slip, so the model carries the slip: every month of delayed deployment is a month of fully metered leak continuing while the invoice runs. And the diligence questions from the next chapter (the BAA, the subprocessor map, 42 CFR Part 2 handling if any SUD program is in scope, HITRUST or SOC 2 Type II posture) are priced in here, because an enterprise contract that fails Part 2 diligence after signature is the most expensive document the practice will ever hold. For Jordan at 25 clinicians, the enterprise tier is worth a quote and a comparison row, but the honest model will usually show the crossover point sits above Jordan's headcount unless the CCBHC-style reporting burden applies.

A vendor prices the tool against its competitors. An owner prices the tool against the leak. The model that survives the board meeting is the one where every saving line is a number the practice itself measured.

Option Three: EHR-Native AI, and the "Basically Free" Trap

The third option arrives with the most seductive price: the AI already inside the EHR. SimplePractice ships Sidekick with Blueprint Health embedded for measurement-based care, TherapyNotes ships its own AI, and Therapy Brands has rolled native AI into the EHR layer for Medicaid-heavy clinics. The budgeting appeal is obvious: no new vendor, no integration project, no separate BAA negotiation, often no incremental invoice, or a modest add-on fee. For a practice whose EHR-fit score is high and whose documentation needs are mainstream, native AI can legitimately win the model.

But "basically free" deserves the leak audit's skepticism, because the native option carries three costs the invoice hides. The first is capability cost: a general-purpose EHR's AI may not draft the behavioral health note your payer mix demands, the time-in-session anchored, medical-necessity grounded 90837 note that survives an Optum audit, and a tool that saves ten minutes but produces a thinner note than a dedicated scribe is repairing the small leak while widening the structural one. Pilot it on your own notes against your own denial reasons before believing it. The second is the subprocessor cost, which Jordan has already paid once in this chapter: the EHR vendor's confident "yes" on AI-assisted scoring concealed a model provider in the subprocessor chain that does not sign a BAA at the tier Jordan pays for. Native does not mean compliant; it means bundled, and the bundle's diligence questions are identical to a standalone vendor's. The third is switching cost in reverse: native AI deepens EHR lock-in, so its true price includes the future negotiating leverage the practice surrenders. None of these costs makes native AI wrong; all of them belong in the model as rows, not footnotes.

The comparison discipline across all three options: hold the use case constant. Price each option's ability to deliver the roadmap's phase one and two (verified documentation drafting, MBC support with 96127 billing where covered, prior-auth drafting fed by real PHQ-9 trajectories), not its feature list. An option that cannot deliver the kill-quadrant discipline, that bundles, say, a client-facing chatbot you cannot disable, imports a governance cost the model must also count.

Building the Model: One Page, Three Columns, Honest Rows

Now assemble the budget model itself: one page, three columns (Per-Clinician, Enterprise, EHR-Native), and two row blocks. The cost block per column: license or contract cost annualized at your actual headcount and negotiated rates; implementation and integration; training hours priced at real loaded cost; administration (license management, the AI-use log, consent workflow); and the contract terms that change risk (term length, exit conditions, BAA and subprocessor posture, with any unresolved diligence item flagged in red rather than priced at zero).

The value block, identical across columns so the options compete on capture, not on storytelling: documentation hours recovered (your measured time-per-note delta from the pilot, multiplied across licensed seats, priced conservatively at replacement value, with the opportunity-value figure shown separately and labeled as the aggressive case); denial reduction (the documentation-sensitive denial slice, credited only at the improvement rate your roadmap's day-60 and day-180 readings demonstrate, valued at the real claims plus avoided appeal hours); and retention scenarios (zero, one, and two avoided replacements, each valued at $25,000 to $60,000 plus six to nine months of productivity ramp, with the scenario selection driven by your satisfaction-pulse and turnover data, not by hope). Beneath both blocks, one line per column: first-year net, steady-state annual net, and the break-even statement in operational terms an owner feels, such as how many avoided denials or how many retained clinicians pay for the year.

Two disciplines keep the model honest. First, every value row cites its source: the readiness assessment baseline, the pilot log, the day-60 denial reading. A row without a measured source gets the conservative default or a zero, because the model's credibility is worth more than its total. Second, the model is versioned like the roadmap: re-run at day 90 with real pilot deltas, re-run at day 180 with real denial and retention reads, and the version history kept, because the budget that funds year two will be judged against the promises this version made.

Presenting the Model: The Board, the Skeptic, and the Carrier

A budget model has three audiences, and the same page must survive all of them. The board or ownership group wants the headline economics and the risk posture: lead with the conservative case, show the leak metering (most owners have never seen their unpaid-hours number annualized, and that number alone reframes the meeting), and present the retention scenarios as scenarios, because a board that catches you treating hope as a forecast discounts everything else on the page. The clinical skeptic wants to know the money does not corrupt the clinic: show that every saving assumes the verification pass survives (the time-per-note delta is measured with clinicians reading every word before signing), that nothing in any priced option violates the hard limits (no AI scoring of the CSSRS, no risk levels, no Tarasoff or mandated-report determinations), and that the kill switch is funded, meaning the model still works if the practice cancels a failing tool at day 90.

The third audience is external: the malpractice carrier whose questionnaire started this chapter, and any payer or auditor who later asks how the practice governed its AI spend. For them, the model's value is its provenance: it descends from a dated readiness assessment, funds a roadmap with named milestones, credits only measured improvements, and carries the diligence flags (BAA, subprocessor, Part 2 if applicable) as explicit conditions. A practice that can produce that chain is a practice that made a governed investment; a practice that cannot is a practice that bought software.

One last word from the supervisor's chair: the cheapest option in the model is rarely the free one and never the unmeasured one. Jordan's twelve ungoverned subscriptions were "free" to the practice for a year, and they generated the largest exposure on the readiness assessment. The most expensive option is a tool, at any tier, whose savings nobody measures, because it converts the budget from an investment with a feedback loop into a recurring cost with a story. Build the model, meter the leaks, and let the numbers, not the vendors, set the price you are willing to pay.

The Applied Problem: The Three-Column AI Budget Model

Your deliverable is the Three-Column AI Budget Model: one page comparing Per-Clinician, Enterprise, and EHR-Native licensing for your own practice, against your own metered leaks. Build it in four passes. Pass one, meter the leaks from your readiness assessment and roadmap baselines: annualized unpaid documentation hours (measured weekly hours x clinicians x 52), priced at replacement value with the opportunity value shown separately; the documentation-sensitive denial slice by payer with claim values and appeal hours; and trailing turnover with each replacement priced at $25,000 to $60,000 plus six to nine months of suboptimal productivity.

Pass two, price the three columns with real numbers: a group quote for the per-clinician tier (not the retail seat price), an enterprise quote if your headcount or CCBHC/HCBS reporting burden justifies one, and the EHR-native option with its add-on fees and its diligence answers in writing, including the subprocessor and BAA-tier question. Book training, administration, and integration as explicit rows in every column. Pass three, draft with help if you want it, using an administrative prompt containing no PHI: "Here are my practice's metered costs (annualized unpaid documentation hours at two pricings, documentation-sensitive denials with claim values, turnover with replacement costs) and three licensing options with their cost rows. Build a one-page, three-column budget model with identical value rows per column, conservative and aggressive cases labeled, retention as zero/one/two avoided-replacement scenarios, first-year net, steady-state net, and a one-line operational break-even per column. Flag any cost or diligence row I left unpriced."

Pass four is the verification pass, run like a note review: every value row must cite a measured source or take the conservative default; the retention scenarios must be labeled as scenarios; no column may include a tool or feature that violates the kill quadrant or the hard limits; every diligence flag (BAA coverage at your tier, subprocessor disclosure, Part 2 handling if applicable) appears in red until resolved in writing. Done looks like this: one page a board member reads in five minutes, a conservative case that survives the skeptic, a break-even line in operational language, dated, versioned, and scheduled for re-running at day 90 and day 180 with the roadmap's real readings. File it with the readiness assessment and the roadmap; together the three documents are the chapter's answer to the question Jordan started with: not whether to adopt AI, but how to adopt it like an owner.

Key Takeaways

• Budget against the leak, not against zero. The practice is already paying for not having AI: five to ten unpaid documentation hours per clinician per week, a documentation-sensitive denial slice, and turnover at $25,000 to $60,000 per replacement plus six to nine months of suboptimal productivity. The licensing quote only means something against those metered numbers.
• Per-clinician licensing (the Mentalyc, Upheal, Heidi, Twofold tier, in the territory of $59 a month per seat) buys linear cost, fast deployment, and a real kill switch, but compounds at scale, demands a practice-level BAA covering every seat, and strains in 40-plus agencies and CCBHC reporting environments.
• Enterprise licensing (the Eleos category, proven at CalMHSA scale across roughly 27,000 public-sector clinicians) front-loads cost into implementation and multi-year terms but reaches workforce-shaped value: CCBHC PPS and quarterly reporting, supervision visibility, and analytics. Write the roadmap's milestone gates into the contract as performance conditions.
• EHR-native AI (SimplePractice Sidekick with Blueprint embedded, TherapyNotes AI, Therapy Brands) is never "basically free": price the capability gap against your payer mix and Optum 90837 audit exposure, run the same subprocessor and BAA-tier diligence Jordan learned the hard way, and count deepened EHR lock-in as a row, not a footnote.
• Keep the value block identical across columns and credit only measured improvements: the pilot's time-per-note delta priced conservatively at replacement value, denial reduction only at the rate your day-60 and day-180 readings show, and retention as zero/one/two avoided-replacement scenarios rather than a forecast.
• The model has three audiences: the board gets the conservative headline and the annualized unpaid-hours number, the clinical skeptic gets proof that every saving assumes the read-every-word verification pass and the hard limits survive, and the carrier or auditor gets the provenance chain from readiness assessment through roadmap to budget.
• The cheapest option is never the unmeasured one. An AI tool whose savings nobody measures is a recurring cost with a story; a versioned model re-run at day 90 and day 180 with real readings is an investment with a feedback loop, and that distinction is the whole discipline of this chapter.