A concise briefing for Ontario CPA partners on why controlled AI operating systems are practical now, not next year — and which workflows to map first.

Chart 01 · SWE-bench Verified

Can AI complete real multi-step work?

The clearest public proxy for whether AI can complete complex work — not merely suggest it. The 75% line is the long-task reliability waterline.

Table view
Chart 01 · SWE-bench Verified, as charted
QuarterModelScore
Q2 ’24GPT-4 era40%
Q2 ’25Claude 3.549%
Q4 ’25Opus 4.581%
Q2 ’26Opus 4.7 · GPT-5.588%
Chart 02 · Output pricing

What changed in the economics?

  • $ / 1M output tokens

List price per million output tokens for leading models, as published by their makers. The highest price fell from $75 in early 2024 to $25–30 in 2026, about 2.5 to 3 times lower, while the score in chart 01 rose from 40% to 88%. The $10 and $15 bars are lower-priced models from the middle of 2024.

Table view
Chart 02 · Output list prices, as charted
Quarter$ / 1M output tokens
Q1 ’24$60, $75
Q2 ’24$10, $15
Q4 ’25$25
Q1 ’26$25
Q2 ’26$30

Eight new primitives — one operating layer.

Did not exist together a year ago

  • Persistent memory

    Context survives across sessions and tools.

  • Durable workflows

    Long jobs survive timeouts and restarts.

  • Sub-agent coordination

    Supervisors dispatch specialists in parallel.

  • Reliable computer use

    Drives accounting and tax software.

  • Messaging integration

    Email, SMS & Teams as native channels.

  • Governed tool access

    MCP-secured, audit-logged, partner-scoped.

  • Skills marketplace

    Reusable, vetted capability packages.

  • Structured outputs

    Schema-validated work product, not freeform.

What this means for Ontario CPA firms

01

Quality at scale

Standardized, consistent work product with fewer manual touchpoints across the firm.

02

Capacity unlocked

Expand advisory and higher-value work without proportional headcount growth.

03

Risk under control

PIPEDA-aware, audit-trailed, partner-reviewed workflows from day one.

04

Compounding advantage

Firms that map carefully in 2026 build an operating advantage for the next decade.

The inflection — by the numbers

Tax and accounting firms using AI, 2024 → 2025
9%41%
Wolters Kluwer, Future Ready Accountant, October 2025. A global survey of tax and accounting professionals.
Firms planning to increase AI investment over the next three years
77%
Wolters Kluwer, the same report.
Saved per employee, per month, as reported by firms using AI
18hrs
Karbon, State of AI in Accounting 2025, February 2025. Self-reported, from more than 500 accounting professionals.
Share of AI-driven returns captured by the 20% of companies that lead
74%
PwC, AI Performance Study, 2026. 1,217 companies surveyed.
Why this matters

Most of the return goes to a few. Map workflows before procurement — then decide what to automate, what to review, what stays human judgment.

The capability is real. The compliance posture is real. The firms that map this carefully in 2026 will compound a structural advantage — and the firms I trust most are still asking the right questions, not buying the loudest tool.

Ferdie Botden, CPAFounder · NorBot Systems · Stratford, Ontario
Sources

For the charts and the four figures above:

  • Anthropic release notes & Opus 4.7 announcement
  • OpenAI GPT-5.5 system card & API pricing
  • SWE-bench Verified leaderboard
  • Wolters Kluwer, Future Ready Accountant 2025
  • Karbon, State of AI in Accounting 2025
  • PwC, AI Performance Study 2026

Also consulted for the April brief:

  • Google Gemini release notes
  • CPA.com / AICPA productivity data
  • Thomson Reuters 2026 SOP report
  • Sage 2026 SMB accounting adoption
  • NorBot internal CPA market research