All case studies
Case study

ChatGPT Just Changed — And Most Businesses Haven't Noticed

5 May 20267 min readListen to the episode

1. The function we redesigned

The board-level AI strategy process — specifically, the move from "how do we get our existing teams using AI tools?" to "if we were designing this function from scratch with AI agents at the foundation, what would it look like?" Redesign was needed because, per PwC's 2026 AI Performance Study, 74% of AI's economic value is being captured by just 20% of organisations, and the differentiator is structural — not budget, not tool choice. The retrofit framing produces portfolios of pilots that "match nobody's priorities" (PwC).

2. The fresh-expert lens

The fresh expert is senior leadership doing top-down, blank-sheet design — not the existing function lead, not a crowdsourced internal innovation programme. Incumbent thinking fails here for a specific reason: people inside a function optimise their own roles, layer tools onto current workflows, and protect headcount and process boundaries. PwC's explicit 2026 finding was that the most common AI mistake is "a ground-up rather than top-down approach." The fresh-expert lens unlocks three things the incumbent lens cannot: (a) permission to delete roles and decisions, not just augment them; (b) willingness to let AI make decisions autonomously rather than just assist (top performers are 2.8x more likely to do this — PwC); and (c) the discipline to point AI at growth and business-model reinvention, not just cost-cutting (top performers are 2–3x more likely to use AI to pursue new growth — PwC).

3. Agent stack

The redesign exercise itself is run with agents. The redesigned function is then operated by agents. Both sides of the stack:

Stack used to run the redesign exercise:

  • Workflow-inventory agent — extracts every recurring task in the chosen function from process docs, calendars, JIRA/ticketing, and SOPs.
  • Task-classification agent — sorts each task into structured/rules-based vs. judgement-based vs. relational.
  • Agent-capability mapping agent — matches each structured task to a current commercial agent capability (document intake, reconciliation, triage, drafting, monitoring).
  • Org-redesign drafting agent — proposes 2–3 redesigned org charts with role counts, role descriptions, and decision rights.
  • Quantification agent — estimates time, cost, and throughput delta vs. the current baseline.

Stack deployed inside the redesigned function (Google's named examples):

  • Threat-triage agents — Gemini-based agents triaging tens of thousands of unstructured threat reports a month inside Google's Security Operations Centre (Sundar Pichai, Google Cloud Next 2026).
  • Creative-generation agents — used by Google's marketing team to generate thousands of asset variations for the Gemini-in-Chrome launch (Google Cloud Next 2026).
  • Enterprise agent platforms named in the episode: Google Gemini Enterprise, OpenAI Workspace Agents, Anthropic Claude. Customers cited at Google Cloud Next 2026: The Home Depot, Unilever, Mars, Papa John's. Unilever is rolling agents across the entire organisation to serve 3.7 billion consumers (Google Cloud Next 2026).

4. The step-by-step process

  1. Pick one function. Choose the function in the business with the highest concentration of structured, recurring, rule-based work. Candidates named in the episode: finance, compliance, operations, sales support, customer operations, competitive monitoring.
  2. Force a top-down mandate. Senior leadership — not the function head — owns the redesign question. PwC's data: leaders crowdsourcing AI ideas from teams produced portfolios that did not lead to transformation.
  3. Inventory the current workflows in that function.
    • Inputs: process documents, SOPs, role descriptions, ticketing data, calendars.
    • Output: a flat list of every recurring task, with frequency and time-cost.
  4. Classify each task. For every task, label as: (a) structured / rules-based / data-processing, (b) judgement / interpretation, (c) relational / human-trust. Only (a) is the redesign target in v1.
  5. Map agents to structured tasks. For each (a) task, name the specific agent capability that handles it today (e.g. document intake, reconciliation, variance analysis, anomaly detection, first-pass commentary, exception flagging, communication drafting).
  6. Sketch the rebuilt function on a blank sheet. Answer five questions in writing:
    • How many humans?
    • With what skills?
    • Doing what work?
    • What decisions do agents make autonomously?
    • What does the function director spend their week on?
  7. Pick one workflow as the prototype. Smallest end-to-end loop with measurable input and output. Build it. Run it. Do not procure a platform first.
  8. Measure the prototype against a real baseline. Before/after on cycle time, cost per unit, error rate, throughput. Numbers, not adjectives.
  9. Codify the design pattern. Write down the document intake → routing → action → exception → reporting pattern. PwC found leading firms scale a single proven pattern across teams, regions, and functions — they do not let value sit in one pocket.
  10. Replicate the pattern in the next function. The episode's named example: an insurer that proves AI can cut invoice processing time in finance reuses the same intake + workflow model for contract review in legal and claims processing in operations.
  11. Reallocate resources towards what is working. PwC: leaders are 1.3x more likely to reallocate financial and human resources toward high-value AI projects as priorities shift. Treat the AI portfolio like a fund — place deliberate bets, move resource toward what is working, cut what is not.
  12. Re-run the design question every quarter. The blank-sheet question is not a one-off. As agent capability changes, the answer changes.

5. Before / after

Sourced numbers from the episode:

  • AI-fit companies deliver AI-driven financial performance 7.2x higher than other companies (PwC 2026 AI Performance Study, Joe Atkinson).
  • 74% of AI's economic value is captured by 20% of organisations (PwC).
  • Top performers are 2x more likely to redesign workflows around AI vs. layering AI onto existing processes (PwC).
  • Top performers are 2.6x more likely to report AI has improved their ability to reinvent their business model (PwC).
  • Top performers are 2.8x more likely to have increased the number of decisions made without human intervention (PwC).
  • Leading companies invest ~2.5x more of their revenue in AI than other companies (PwC).
  • ~80% of AI initiative value comes from redesigning work; ~20% from the technology itself (PwC).
  • Google Security Operations Centre, after deploying Gemini-based triage agents: >90% reduction in threat mitigation time (Sundar Pichai, Google Cloud Next 2026).
  • Google marketing, Gemini-in-Chrome launch: 70% faster turnaround, 20% increase in conversions (Google Cloud Next 2026).
  • Google committed $750 million to accelerate partner-led development of agentic AI systems (Google Cloud Next 2026).

For your own redesigned function: [reader: measure your own baseline] — cycle time, cost per unit, throughput, error rate, decision latency. Capture the numbers before you switch on the prototype, or you will not be able to defend the result later.

6. How to replicate in 1 day

  • 30 min — Pick the function and write the mandate. One paragraph from the CEO or board: "We are redesigning [function] from scratch around AI agents. Existing structure is not the constraint."
  • 90 min — Workflow inventory. Pull SOPs, role descriptions, and recurring tasks for the chosen function. List every recurring task, its frequency, and approximate time-cost.
  • 60 min — Classify tasks. Sort into structured / judgement / relational. Highlight the structured layer — that is the redesign target.
  • 90 min — Map agents. For each structured task, name the commercial agent capability that already exists for it (document intake, reconciliation, anomaly detection, drafting, triage, monitoring).
  • 60 min — Sketch the rebuilt function. Headcount, roles, skills, decision rights, agent-vs-human boundaries. Two pages maximum.
  • 2 hours — Build the smallest prototype. Pick one end-to-end workflow inside the function. Stand it up using one of the platforms named in the episode (Gemini Enterprise, OpenAI Workspace Agents, Claude). Do not procure — prototype.
  • 60 min — Capture the baseline and run a measurement plan. Define the before/after metrics, owner, and review date. Schedule the next-function replication on the calendar before you stop.

Total: ~8 hours. End of day deliverable: a written blank-sheet redesign of one function, a working prototype on one workflow, and a measurement plan with baseline numbers captured.