The most common enterprise AI mistake isn't choosing the wrong model or underinvesting in data. It's pointing AI at a broken process and expecting the technology to fix what's fundamentally a design problem.
What You Need to Know
- AI accelerates processes. It doesn't redesign them. If your claims process has seven unnecessary approval steps, AI will help you complete those seven steps faster. It won't remove them.
- The enterprises getting real value from AI redesign the process first, then deploy AI within the redesigned workflow. The order matters.
- "Automating" a broken process with AI creates new problems: faster errors, scaled inefficiencies, and an AI system that's hard to change because it's been built around the wrong workflow.
- Process redesign doesn't mean a 12-month transformation programme. It means asking "if we were starting from scratch with AI available, how would we design this?" and then closing the gap.
- The biggest operational gains come from eliminating steps, not accelerating them.
70%
of change programmes fail because they automate existing processes instead of redesigning them
Source: McKinsey & Company, The State of AI in 2022, December 2022
The Automation Trap
Here's a pattern we see repeatedly:
An enterprise identifies a time-consuming process, say contract review. Currently, a junior analyst reads each contract, extracts key terms, compares them against standard policies, flags deviations, and sends the analysis to a senior reviewer for sign-off. The whole cycle takes 4-6 hours per contract.
The AI brief: "Build an AI tool that extracts key terms from contracts, compares them against policies, and flags deviations."
The AI does this beautifully. Contract analysis time drops from 4-6 hours to 30 minutes. Everyone celebrates. Then someone asks: why are we still doing it this way?
In many cases, the manual steps existed because of historical constraints that AI removes. The junior analyst extracted terms because a human needed to read the document. AI can read it directly. The comparison against policies was manual because the policy database wasn't machine-accessible. With an AI foundation, it can be. The senior reviewer signed off because the junior analyst might miss something. With AI confidence scoring, the review can focus only on low-confidence results.
The redesigned process might be: AI processes the contract, flags only deviations with medium/low confidence, and the senior reviewer handles those 15% of cases. The other 85% flow through automatically with audit logging.
That's not a 4-6 hour process accelerated to 30 minutes. It's a 4-6 hour process redesigned to 5 minutes for 85% of cases and 20 minutes for the complex 15%.
The difference: Automation gave you a 12× improvement. Redesign + AI gave you a 70× improvement. Same technology. Different approach.
How to Think About AI and Process Design
Principle 1: Start with the Outcome, Not the Current Process
Don't ask "how can AI help with our current contract review process?" Ask "what does the ideal contract review outcome look like, and how would we achieve it if we were starting today?"
This isn't blue-sky thinking. It's practical. The current process was designed with human constraints. AI changes those constraints. The design should change too.
Principle 2: Eliminate Before You Accelerate
For every step in your current process, ask three questions:
- Does this step still need to exist? Some steps were created for constraints AI removes (like manual data extraction).
- Does this step need a human? Some steps are purely mechanical and can be fully automated.
- Does this step need to happen in this order? AI enables parallel processing that sequential human workflows can't.
The most valuable AI interventions aren't the ones that speed up existing steps. They're the ones that remove steps entirely.
Principle 3: Design for the Exception, Not the Rule
In most enterprise processes, 70-85% of cases are straightforward. They follow the standard path with standard inputs. The remaining 15-30% are the exceptions that require human judgement, contextual understanding, or creative problem-solving.
AI should handle the standard cases end-to-end (with appropriate governance and monitoring). Humans should focus on the exceptions: the interesting, complex, high-value work that actually uses their expertise.
This is better for the business (faster throughput, lower cost) and better for the team (they work on challenging problems instead of routine processing).
The best change management for AI is showing people that AI handles the boring 80% so they can focus on the interesting 20%. That's a change most people embrace.
Tim Hatherley-Greene
Chief Operating Officer
Principle 4: Build the Foundation for Continuous Redesign
Processes aren't static. Markets change, regulations evolve, customer expectations shift. Your AI systems should be designed for continuous improvement, not a one-time deployment.
This means:
- Feedback loops that capture where the AI struggles (these are your redesign signals)
- Modular architecture that lets you change one step without rebuilding the whole system
- Regular process reviews that ask "given what we've learned, how should this work now?"
The foundation approach makes this possible. When your AI infrastructure is shared and modular, redesigning a process doesn't mean rebuilding from scratch.
The Redesign Sprint
You don't need a consulting engagement to redesign a process. You need 2-3 days with the right people:
Day 1: Map the current process end-to-end. Mark every step that's mechanical, every step that's a handoff, every step that's a bottleneck, every step that exists "because we've always done it this way."
Day 2: Design the AI-native version. What would this process look like if AI could handle any mechanical step, access any knowledge, and work at any speed? Start idealistic, then constrain for reality.
Day 3: Gap analysis. What needs to change to move from current to redesigned? Prioritise by impact. Identify what the AI initiative needs to deliver.
This costs three days of your team's time. It's the difference between AI that gives you a 10× improvement and AI that gives you a 2× improvement.
- Doesn't process redesign slow down AI deployment?
- It adds 1-2 weeks to the planning phase and saves months of building the wrong thing. An AI system built around a redesigned process ships faster and delivers more value than one built around a process that's about to change anyway.
- What if the current process has compliance requirements that can't change?
- Compliance requirements define outcomes, not processes. "Every contract deviation must be reviewed by a qualified analyst" doesn't mean a human must read every contract. It means a human must review flagged deviations. AI handles the identification; the human handles the review. The compliance outcome is met with a fundamentally different (and faster) process.
- How do we get buy-in for process redesign from teams who've done things the same way for years?
- Start with the pain points they already recognise. Every team has steps they find tedious, bottlenecks they work around, and frustrations they've accepted. Frame the redesign as solving their problems, not changing their jobs. And involve them in the design. The people who do the work understand the work best.