Skip to content

Tax Practice AI - Cost & Pricing Detail

Version: 4.0 Last Updated: 2025-12-25 Audience: Internal (SaaS business planning)


Document Purpose
AI_DELEGATION_STRATEGY.md Model routing, batch API, caching implementation
PROCESS_FLOWS.md Workflow states (PendingFinalQA, etc.)
ARCHITECTURE.md System architecture, infrastructure

Executive Summary

This document details the economics of Tax Practice AI as a SaaS product. We sell AI-powered tax analysis to accounting firms on a per-return fee basis.

The Business Model: - We charge accounting firms $20-30 per tax return - Our cost to deliver: $0.54-0.75 per return (AI + infrastructure) - Gross margin: 96-98%

At Scale (10 client firms, 10,000 returns/year): - Annual revenue: $200,000-300,000 - Annual cost: ~$7,600-9,700 - Annual profit: ~$190,000-292,000

Updated Dec 2025: Opus 4.5 pricing (3x cheaper) + batch/caching optimizations reduce AI costs significantly.


1. Understanding Tokens (Non-Technical Explanation)

What is a Token?

A token is how AI models measure text. Think of tokens like the "fuel" that powers AI—the more text processed, the more tokens consumed.

Simple rule of thumb: - 1 token ≈ 4 characters of English text - 1 token ≈ ¾ of a word - 100 tokens ≈ 75 words ≈ 1 short paragraph

Practical Examples

Document Approximate Size Tokens
A single sentence 15 words ~20 tokens
One paragraph 100 words ~130 tokens
One page of text 500 words ~670 tokens
A W-2 form (extracted data) ~200 words ~270 tokens
A 1099-B with 50 transactions ~800 words ~1,100 tokens
Full conversation (10 Q&A exchanges) ~2,000 words ~2,700 tokens

Tax Return Token Examples

Return Type Description Est. Tokens
Simple 1040 W-2 only, standard deduction, single state 40,000-60,000
Typical 1040 W-2, 1099s, itemized deductions, single state 90,000-120,000
Complex Individual K-1s, rental properties, multi-state 150,000-200,000
S-Corp (1120S) Business return with K-1 generation 200,000-300,000
Partnership (1065) Multiple partners, complex allocations 250,000-400,000

Why Tokens Matter for Pricing

We pay Anthropic (the AI provider) based on tokens consumed: - Input tokens (what we send to AI): Reading documents, providing context - Output tokens (what AI sends back): Answers, analysis, worksheets

Different AI models have different token costs—we optimize by using cheaper models for simple tasks and expensive models only for complex analysis.


2. Our Cost Structure

Fixed Monthly Costs (We Pay Regardless of Volume)

Item Monthly Annual Notes
AWS Aurora PostgreSQL $80 $960 Database for all client data
AWS S3 Storage $3 $36 Document storage
AWS Lambda $5 $60 Serverless compute
Airflow VM (EC2) $23 $276 Workflow orchestration
API Gateway $4 $48 API routing
CloudWatch $15 $180 Monitoring & logs
Secrets Manager $5 $60 Credential storage
Anthropic Base $50 $600 Minimum monthly AI spend
Total Fixed $185 $2,220

Variable Costs (Per Return)

Item Cost/Return Notes
AI tokens (blended) $0.33 With Opus 4.5 pricing, see Section 4
Persona (identity verification) $0.30 20% of returns are new clients × $1.50
Twilio (SMS) $0.12 Notifications and 2FA
Total Variable $0.75 Per return processed

With batch + caching optimizations, AI cost drops to ~$0.12/return. See Sections 5-6.

Cost Per Return at Different Scales

Scale Returns/Year Fixed/Return Variable/Return Total/Return
1 firm (500) 500 $4.44 $0.75 $5.19
1 firm (1,000) 1,000 $2.22 $0.75 $2.97
5 firms (5,000) 5,000 $0.44 $0.75 $1.19
10 firms (10,000) 10,000 $0.22 $0.75 $0.97

Key Insight: Fixed costs amortize quickly. At 10,000 returns, our all-in cost is <$1/return.

With Full Optimization (batch + caching):

Scale Returns/Year Fixed/Return Optimized Variable Total/Return
1 firm (1,000) 1,000 $2.22 $0.54 $2.76
5 firms (5,000) 5,000 $0.44 $0.54 $0.98
10 firms (10,000) 10,000 $0.22 $0.54 $0.76

Optimized variable = $0.12 AI + $0.30 Persona + $0.12 Twilio = $0.54


3. Full Lifecycle Token Usage

A tax return isn't just processed once—it's referenced throughout the year. Our token estimates account for the full lifecycle.

Tax Season (January - April)

Phase AI Activities Tokens/Return
Document Upload Classification, extraction (15 docs avg) 46,500
Analysis Prior year comparison, missing docs, anomalies 23,800
Preparer Q&A Staff questions during review (5 avg) 19,000
Worksheet Generation Export with citations 23,000
Subtotal 112,300

Post-Season (May - December)

Activity Frequency Tokens/Event Annual Tokens
Estimated tax reminders 4× per year 800 3,200
Client questions about filed return 3× per year 2,000 6,000
Amendment analysis (10% of returns) 0.1× 50,000 5,000
Subtotal ~14,200

Full Year Token Budget Per Return

Period Tokens % of Total
Tax Season (Jan-Apr) 112,300 89%
Post-Season (May-Dec) 14,200 11%
Annual Total 126,500 100%

Planning Note: Budget ~125,000 tokens per return per year, not just processing.


4. AI Model Pricing

Available Models (Anthropic Claude via AWS Bedrock)

Model Input Cost Output Cost Best For
Haiku 3 $0.00025/1K $0.00125/1K Extraction, classification
Sonnet 4 $0.003/1K $0.015/1K Analysis, Q&A
Opus 4.5 $0.005/1K $0.025/1K Complex tax code, judgment

Pricing as of Dec 2025 via AWS Bedrock. Opus 4.5 is 3x cheaper than previous Opus 4.1.

Cost Per Return by Strategy

Strategy Token Mix Cost/Return
All Sonnet (simple) 0% Haiku, 100% Sonnet $0.64
Delegation (recommended) 60% Haiku, 35% Sonnet, 5% Opus $0.33
Expert-heavy 40% Haiku, 40% Sonnet, 20% Opus $0.55

Savings: Delegation strategy cuts AI cost by 48% vs. all-Sonnet approach.

For model routing rules and touchpoint assignments, see AI_DELEGATION_STRATEGY.md.


5. Optimization Cost Impact

This section summarizes the cost savings from our optimization strategies. For implementation details, see AI_DELEGATION_STRATEGY.md.

Batch API Savings

Batch processing via Anthropic's batch API is 50% cheaper than interactive.

Scenario % Batch Annual Savings (1,000 returns)
Conservative 30% $72
Moderate 50% $120
Aggressive 70% $168

Realistic estimate: 50% batch = $0.12/return cost reduction.

For task classification and UX strategy, see AI_DELEGATION_STRATEGY.md.


Metadata Caching Savings

Extract once, reference forever. AI reads cached metadata instead of re-scanning original documents.

Scenario Without Caching With Caching Savings
First scan (W-2) 2,500 tokens 2,500 tokens
Q&A reference (×5) 12,500 tokens 1,500 tokens 88%
Worksheet generation 3,000 tokens 600 tokens 80%
Total per doc 18,000 4,600 74%

Per Return Impact:

Metric Without Cache With Cache
Tokens/return 126,500 ~50,000
AI cost/return (blended) $0.40 ~$0.16
Savings $0.24/return (60%)

For cache structure and implementation, see AI_DELEGATION_STRATEGY.md.

Combined Optimization Summary

Optimization Starting Cost After Savings
Model delegation (60/35/5) $0.64 $0.33 $0.31
+ Batch processing (50%) $0.33 $0.23 $0.10
+ Metadata caching (75%) $0.23 $0.12 $0.11
Fully optimized $0.12 $0.52

Result: ~81% cost reduction vs. naive all-Sonnet interactive approach.


6. Tiered Validation & Revenue Opportunities

Tiered Validation Strategy (FUT-002)

Quality assurance without breaking the bank:

Layer Model What It Does Cost/Return
Self-validation Haiku/Sonnet Each model QAs own work ~$0.002
Pre-review Sonnet Catches issues before reviewer ~$0.005
Final QA (batch) Opus Metadata-only overnight review ~$0.005
Total QA ~$0.012

Workflow: Approved → PendingFinalQA → (overnight) → PendingSignature or RevisionsNeeded

For workflow integration, see PROCESS_FLOWS.md and AI_DELEGATION_STRATEGY.md.

Expedited QA (Revenue Opportunity)

Mode Wait Time Our Cost Client Fee Margin
Standard (batch) 12-24 hrs $0.005 Included
Expedited ~2 min $0.03 $5-10 97%+

Use cases: Filing deadline, rush returns, client priority.

Revenue impact at 1,000 returns (10% expedited): - 100 expedited × $7.50 avg = $750 incremental revenue - Cost: 100 × $0.03 = $3 - Pure margin: $747

Image Quality Escalation Cost Impact

Phone photos and low-quality scans escalate to higher-cost models:

Scenario % of Docs Cost Impact
Native PDF/clean scan 70% Baseline
Phone photo 20% +$0.02/doc
Low-res/blurry 8% +$0.02/doc
Handwritten 2% → Human (no AI cost)

Estimated impact: +$0.03/return average (already included in estimates above).


7. SaaS Pricing Model

Our Revenue: Per-Return Fee

We charge accounting firms based on returns processed. Customer pays per return, we deliver AI analysis.

Pricing Tiers

Tier What's Included Our Price Our Cost Margin
Standard Haiku extraction + Sonnet analysis $20/return $0.82 96%
Professional Standard + Opus escalation (10%) $30/return $1.20 96%
Expert Opus-primary for all analysis $50/return $3.50 93%

Revenue Projections

Single Client Firm (1,000 returns/year)

Tier Revenue Our Cost Profit Margin
Standard @ $20 $20,000 $3,040 $16,960 85%
Professional @ $30 $30,000 $3,420 $26,580 89%
Expert @ $50 $50,000 $5,720 $44,280 89%

Cost includes $2,220 fixed + variable per return

Scaling to Multiple Client Firms

Clients Returns/Year Revenue @$25 Total Cost Annual Profit
1 1,000 $25,000 $3,040 $21,960
5 5,000 $125,000 $6,320 $118,680
10 10,000 $250,000 $10,420 $239,580
25 25,000 $625,000 $22,720 $602,280
50 50,000 $1,250,000 $43,220 $1,206,780

Annual Recurring Revenue (ARR) Targets

Milestone Client Firms Returns ARR @$25 Annual Profit
Year 1 3 3,000 $75,000 $70,000
Year 2 10 10,000 $250,000 $240,000
Year 3 25 25,000 $625,000 $600,000

8. Competitive Positioning

Market Comparison

Product Price Model Cost to Firm What They Get
SurePrep Per-return $12-100/return OCR + data extraction
UltraTax Annual license $7K-25K/year Tax prep software (no AI)
Tax Practice AI Per-return $20-30/return AI analysis + Q&A + worksheets

Value Proposition

If we REPLACE SurePrep: - Client saves $0-80/return (depending on SurePrep tier) - We provide OCR + AI analysis (more than SurePrep offers)

If we COMPLEMENT SurePrep: - Client adds $20-30/return for AI assistant - Justified by reduced prep time and error catching

Pricing Strategy Recommendation

Launch at $25/return (Professional tier) - Competitive with mid-range SurePrep - Strong 96% margin - Room to discount for volume deals - Room to upsell to Expert tier


9. Client Perspective (Their P&L Impact)

While we focus on our margins, understanding client economics helps with sales.

Client Cost Structure (1,000 returns/year)

Item Without Us With Us Change
UltraTax $15,000 $15,000
SurePrep ($20/return) $20,000 $0* -$20,000
SmartVault $1,800 $1,800
Tax Practice AI $0 $25,000 +$25,000
Total Tech $36,800 $41,800 +$5,000

If we replace SurePrep. If complementary, client pays both.

Client ROI Calculation

Time savings assumption: 30 minutes per return in reduced prep time - Preparer cost: $50/hour → $25 saved per return - 1,000 returns × $25 = $25,000 labor savings - Our cost: $25,000 - Net ROI: Break-even on labor alone, plus quality improvements


10. Existing Software Context

The following are costs our client firms already pay. These are NOT our costs, but understanding them informs our competitive positioning.

Service Purpose Client's Annual Cost Notes
UltraTax CS Tax prep software $7,000-25,000 Thomson Reuters, includes e-filing
SurePrep OCR/extraction $2,000-15,000 $12-100/return + setup
SmartVault Client portal $1,300-2,400 $45-65/user/mo

Integration approach: We integrate with these tools, not replace them (except potentially SurePrep).


11. Assumptions & Open Questions

Assumptions (To Be Validated)

ID Assumption Impact if Wrong Validation Method
A1 15 documents per return average ±30% on token costs Analyze pilot data
A2 5 Q&A interactions per return ±20% on token costs Track actual usage
A3 20% of returns are new clients ±10% on Persona costs Client data analysis
A4 Haiku handles 60% of tokens ±15% on AI costs Monitor model routing
A5 $25/return is acceptable price Revenue risk Market testing
A6 Batch processing saves 50% ±5% on total costs Measure actual savings

Open Questions for Business Partner Discussion

  1. Pricing strategy: Fixed per-return fee vs. tiered by return complexity?
  2. Volume discounts: Offer breaks at 2,500 / 5,000 / 10,000 returns?
  3. Annual contracts: Require annual commitment or allow month-to-month?
  4. SurePrep positioning: Market as replacement or complement?
  5. Expert tier demand: Will firms pay $50/return for Opus-level analysis?
  6. Year-round usage: Charge separately for post-season Q&A support?
  7. Overages: How to handle firms that exceed token estimates?

Follow-Up Analysis Needed

  • Pilot with 2-3 firms to validate token usage assumptions
  • Survey target market on price sensitivity
  • Benchmark SurePrep pricing for different firm sizes
  • Model churn scenarios and customer lifetime value
  • Calculate break-even point (how many clients to cover fixed costs)

Document History

Version Date Changes
4.0 2025-12-25 Document restructure: Moved implementation details to AI_DELEGATION_STRATEGY.md. Cost-only focus with references. Consolidated sections 5-6.
3.1 2025-12-25 Reconciliation: Opus 4.5 pricing, tiered validation, expedited QA, cross-ref AI_DELEGATION_STRATEGY.md
3.0 2024-12-25 Major revision: SaaS perspective, token explainer, lifecycle usage, assumptions
2.1 2024-12-25 Add model delegation strategy
2.0 2024-12-25 Add AI model strategy, token breakdown, vendor pricing
1.1 2024-12-25 Add cost per return analysis and sample P&L
1.0 2024-12-25 Initial version, extracted from ARCHITECTURE.md