Case study
ChatGPT and Gemini Won't Save Your Business — Start Over
TL;DR
- The function: one structured, recurring, rule-based function — finance, compliance, ops, sales intelligence, legal review, or customer operations.
- The lens: a 2-3 person mirror team designing from a blank sheet, not retrofitting existing workflows.
- The result: top 20% of organisations capture 74% of all AI economic value (PwC, 1,217 executives, 2026).
The framework
The Mirror Function. Pick one function. Run a small parallel version of it alongside the existing one for 6-8 weeks, designed from scratch around AI agents. Measure the gap between mirror and incumbent. The deliverable is not a tool — it is operational knowledge of how that function works when redesigned, not retrofitted.
The agent stack
- ChatGPT — operational engine for reading, processing, drafting, and synthesis across the workflow.
- Claude — operational engine for long-context reasoning and large-document analysis.
- Gemini — operational engine for multimodal tasks and Workspace-native work.
- Your internal data layer — documents, CRM, finance system, ticketing — feeding the agents.
- Human reviewer — judgment gate on agent outputs before they reach decision-makers.
Replicate it — step by step
- List every function in your business and rank them by share of structured, recurring, rule-based work.
- Pick the top one. Just one. Resist the urge to do three.
- Write the redesign question on one page: "If we built this function from scratch today with AI agents doing the structured work from the start, what would it look like?"
- Ring-fence 2-3 people, a small budget, and a 6-8 week timebox.
- Forbid the team from automating the current workflow — their job is to design a new one.
- Have them prototype the rebuilt function using ChatGPT, Claude, or Gemini as the operational engine.
- Define which decisions agents make autonomously and which require a human judgment gate.
- Run the mirror in parallel with the existing function on the same inputs for the full timebox.
- Measure output quality, cycle time, and resource requirement against the existing function.
- At week 8, present the gap to the board and pick the next function to mirror.
Before / after
- Worker AI access rose 50% in 2025 (Deloitte, 3,235 leaders, 2026).
- Only 34% of organisations are truly reimagining operations (Deloitte, 2026).
- Top 20% are 2x more likely to redesign workflows than add tools (PwC, 1,217 executives, 2026).
- Top 20% capture 74% of AI economic value; bottom 80% share 26% (PwC, 2026).
- Education was the #1 talent-strategy adjustment companies made for AI (Deloitte, 2026).
- [reader: measure your own baseline — cycle time, cost, throughput, decisions automated]
Replicate in one day
- 09:00 (60m) — Map every function in the business; rank by structured, recurring work.
- 10:00 (60m) — Pick the top function. Write the redesign question on one page.
- 11:00 (90m) — Name the 2-3 mirror team members, set the 6-8 week scope, ring-fence the budget.
- 12:30 (30m) — Lunch. No phones.
- 13:00 (90m) — Draft the success metrics: output quality, cycle time, resource cost.
- 14:30 (60m) — Define the agent stack and access permissions for the mirror team.
- 15:30 (90m) — Brief the team. Confirm: redesign, not retrofit. End of day one.