Key Takeaways
- Deloitte's October 2026 survey: 81% of companies say AI made decisions and execution faster, but only 25% use it to deeply transform the business.
- 60% have redesigned jobs around AI to some degree, yet only 15% report a significant shift in decision-making authority.
- Redesign starts with decisions, not tasks: list every decision in one workflow and give each an autonomy tier based on consequence and reversibility.
- Move approval rights, rewrite roles around exceptions, and track cost per case, not just the total AI bill.
- Report three numbers to the board: unit cost, cycle time and exception rate, before and after.
Deloitte's newest State of AI in the Enterprise report contains a finding that should make every COO pause. AI is working, in the narrow sense: 81% of companies say it has made decision-making and execution faster. Yet only 25% say they are using it to transform how the business actually runs.
The gap is not a technology problem. It is a design problem. Most organisations have dropped AI into processes built for people, kept every approval step, and then wondered why the board cannot see the return. This guide gives you a five-step framework for redesigning one workflow properly, and a way to show the result to your CFO.
Why does AI make work faster but not better?
Because most enterprises automate tasks inside an unchanged workflow. The AI drafts, sorts or summarises faster, but the same people still approve the same things in the same order. Speed improves at one step while the queue simply moves to the next approval, so cost and outcomes barely change.
Deloitte's survey of 3,500 director to C-suite leaders across 25 countries, published on 29 October 2026, puts numbers on the pattern. 60% of companies have redesigned jobs around AI capabilities at least to a moderate degree. Only 15% report a significant shift in decision-making authority. In Deloitte's words, employees may have better information and make faster decisions, but they are not empowered to act differently.
This matches what McKinsey's State of AI research has reported across several editions: of the organisational practices it tests, fundamental workflow redesign is the one most strongly linked to bottom-line impact from generative AI, and it remains one of the least common.
In practice the symptom is familiar. A claims team adopts an AI tool that drafts assessments in minutes instead of an hour. But every assessment still waits two days for a supervisor's sign-off, so customer turnaround is unchanged. The pilot report shows a productivity gain. The P&L does not.
What does redesigning work around AI actually involve?
It means changing who decides what, not just how fast tasks are done. Redesign covers four things in one workflow: which decisions AI can take on its own, which approvals can be removed or moved, what people now own instead, and how the workflow's cost and quality are measured.
Deloitte lists the levers directly: effective work redesign requires changes to roles, responsibilities, approval processes, decision rights and the division of work between people and AI. Notice that only one of those five is about technology.
Use this five-step framework on a single workflow first. It works for customer service, finance operations, procurement or HR, and it gives you an evidence base before you scale.
Step 1: Build a decision inventory, not a task list
Map the workflow as a chain of decisions. For a supplier invoice that might be: is this a valid supplier, does it match the purchase order, is the variance acceptable, should it be paid now, does it need escalation. Most process maps list tasks. Decisions are where the delays, costs and risks actually sit.
Step 2: Assign an autonomy tier to each decision
Score each decision on two questions: what is the consequence if it is wrong, and how easily can it be reversed? Then place it in one of four tiers.
| Tier | Who acts | Typical fit |
|---|---|---|
| 1. AI acts | AI decides and executes, logged | Low consequence, easily reversed (routing, tagging, standard replies) |
| 2. AI acts, human notified | AI executes, owner reviews a daily sample | Moderate consequence, reversible (small refunds within policy) |
| 3. AI recommends, human approves | Human signs off on AI's proposal | Material consequence (credit limits, contract terms) |
| 4. Human only | AI may supply information, no recommendation | High consequence or regulated judgement (dismissals, regulatory filings) |
This is exactly the principle Deloitte sets out: governance and safety controls should match the level of autonomy and the potential consequence involved.
Step 3: Move the approval rights
For every decision now in tier 1 or 2, remove the old approval step formally, in the policy and the system, not just in practice. For tier 3, move the approval to the lowest level that can judge it. This is the step most organisations skip, and it is why their speed gains disappear.
Step 4: Rewrite roles around exceptions and judgement
If AI handles the routine 80%, the team's job becomes the difficult 20%: exceptions, complaints, edge cases and relationship calls. Write that into job descriptions and KPIs. Name one accountable owner for each AI-run decision, because Deloitte found 23% of companies say accountability is unclear when AI produces a negative outcome.
Step 5: Measure the workflow's unit economics
Deloitte reports that 71% of companies track total AI inference and API spend, but only 44% track per-unit inference cost. Set a baseline before you redesign: cost per case, end-to-end cycle time, exception rate and error rate. Then compare after 60 to 90 days.
How does work redesign look in Hong Kong industries?
The pattern is the same across sectors: find the approval queue that sits after the AI step, then decide whether that approval still adds value. Three common Hong Kong examples are logistics document exceptions, property management repair requests and financial services client reviews.
Logistics: shipping document exceptions
A freight forwarder uses AI to check commercial invoices and packing lists against booking data. The faster check is only half the gain. The redesign question is which mismatches the AI may correct itself (tier 2, such as a unit format error) and which must go to a documentation lead (tier 3, such as a value discrepancy that affects declared customs value). Once the minor fixes stop queuing for a person, the documentation team spends its day on the cases that can actually hold up a shipment.
Property management: repair and complaint requests
A property manager receives requests by WhatsApp, email and phone across dozens of buildings. AI can classify, prioritise and dispatch standard jobs to approved contractors within a set budget (tier 2). Requests involving safety, insurance or owners' corporation disputes go straight to a named manager (tier 3 or 4). Front-desk staff stop re-keying tickets and start owning follow-up with residents.
Financial services: periodic client reviews
A wealth management firm uses AI to pre-fill periodic client review files from existing records. Under HKMA and SFC expectations, the risk judgement stays with humans (tier 3 or 4), but the redesign removes the old manual data-gathering stage and the second checker for unchanged low-risk profiles, with a sample audit instead. That shift from full checking to sampling is a decision-rights change, and it should be approved by compliance in writing.
What mistakes stop AI work redesign from paying off?
The most common mistakes are keeping every legacy approval, redesigning job titles without moving authority, leaving AI decisions without a named owner, measuring only speed, and trying to redesign every workflow at once. Each one keeps AI as a faster tool inside an unchanged process.
- Automating the approval chain intact. If every AI output still needs the same sign-offs, you have added a step, not removed one.
- Renaming roles without moving authority. A new "AI-assisted analyst" title means little if the analyst still cannot approve anything new.
- No named owner. Every AI-run decision needs one accountable person who reviews samples and can switch the automation off.
- Reporting speed as value. Faster drafting is a leading indicator. Cost per case and customer outcomes are what the CFO funds.
- Big-bang redesign. Redesigning ten workflows at once overloads change management. One workflow, measured properly, builds the case for the next.
Be honest about the human side too. Moving decision rights changes status inside a team. Supervisors who used to approve everything may see redesign as a loss of control. Involve them in setting the autonomy tiers, and give them ownership of the exception queue and the sample audits. Their judgement is what makes tier 2 safe.
How do you present AI work redesign to the board or CFO?
Present it as an operating change with a measured baseline, not a technology purchase. Show one redesigned workflow, the decisions moved between tiers, the named owners, and three before-and-after numbers: unit cost, cycle time and exception rate. Then ask for approval to repeat the method on the next workflow.
Boards have seen many AI pilot slides that report time saved. They are now asking the questions Deloitte's report raises: what does each workflow cost, what does it produce, and who is accountable when it goes wrong.
A one-page summary works well:
- The workflow and its volume: for example, 4,000 supplier invoices a month.
- Decision map: how many decisions moved to tier 1 or 2, and which stayed with humans and why.
- Accountability: the named owner for each AI-run decision and the audit sampling rate.
- Unit economics: cost per case, cycle time and exception rate, baseline versus 90 days.
- Risk controls: how errors are detected, reversed and reported, including PDPO considerations if personal data is involved.
- Next workflow: the candidate and the expected range of impact, stated as a range rather than a promise.
This format also protects you. If the numbers do not move, you know by day 90 and can stop or adjust, rather than defending a vague transformation programme a year later.
Speed is the starting point, not the result
The October data is clear: AI has made enterprise work faster almost everywhere, and changed it in only a few places. The difference is not the model or the vendor. It is whether leaders are willing to redraw decision rights, retire approvals that no longer add value and measure the workflow rather than the tool.
Start with one workflow. Map its decisions, assign autonomy tiers, move the approvals, name the owners and measure unit cost. That is a programme a CFO can fund and a board can oversee.
Empowering AI. Centered on you.
Redesign your first workflow with a partner who has done it
Start with a free AI Staff configuration assessment. We'll walk you through every step, from mapping the decisions in one workflow to deploying an AI staff member, setting autonomy tiers and tracking unit cost, with your data kept in Hong Kong.
Frequently Asked Questions
What did Deloitte's October 2026 State of AI survey find?
It surveyed 3,500 leaders in 25 countries. 81% reported faster decisions and execution, 25% use AI to deeply transform the business, 40% let AI act without human approval at least moderately, and 23% say accountability is unclear when AI goes wrong.
Which workflow should we redesign first?
Pick a high-volume workflow with a visible approval queue, measurable unit cost and limited regulatory exposure, such as invoice processing, internal IT requests or customer enquiries. Avoid starting with a regulated judgement process.
Does letting AI act without approval create PDPO risk?
It can if personal data is processed. Keep decisions involving personal data in tier 3 or 4 until you have documented the data flow, retention and access controls in line with the PCPD's AI guidance, and keep audit logs.
How long before redesign shows financial results?
Operational metrics such as cycle time and exception rate usually move within one or two quarters if approvals are actually removed. Deloitte notes financial returns arrive on a slower timeline than speed gains.
Do we need an in-house AI team to redesign work?
No. Redesign is mainly an operations and governance exercise: mapping decisions, changing policies and setting KPIs. Many enterprises pair an internal process owner with an external partner for deployment and monitoring.