Adding AI to a broken workflow makes the break move faster
The team introduces an AI assistant to create a first draft.
The draft arrives faster. It still needs to be copied into another system, checked against information the model cannot access, rewritten to match policy, approved in a message thread and manually recorded after it is sent.
One task became quicker. The complete workflow barely changed.
This is a common pattern in AI implementation. The company places a powerful capability inside an old process without reconsidering how the work should now operate.
The result is another tool, another handoff and another source of uncertainty.
Start with the job, not the AI
Define the outcome the workflow exists to produce.
Examples include:
- Resolve an eligible customer request accurately.
- Prepare a salesperson for a strategic account meeting.
- Review a contract and surface clauses requiring legal judgment.
- Turn product feedback into a prioritized decision input.
- Produce a financial variance explanation with traceable evidence.
Then map the current process from trigger to completed result.
For every step, capture:
- Who performs it
- What information they need
- Which system they use
- What decision is made
- How long it waits
- What commonly fails
- Which exceptions occur
- What evidence must be retained
Do not begin by asking which steps the model can automate. Begin by understanding why the steps exist.
Identify the real constraint
The obvious manual task may not be the bottleneck.
The workflow may actually be constrained by:
- Missing or unreliable data
- An approval that waits for one leader
- Context distributed across several systems
- Unclear decision rights
- A large number of exceptions
- Poorly defined quality standards
- A handoff between teams
- The absence of a recovery process
AI will not remove a constraint it cannot reach.
Decide what AI, software and people each own
AI is useful for
- Generating drafts
- Extracting and classifying information
- Summarizing large bodies of context
- Identifying patterns and anomalies
- Recommending options
- Translating unstructured information into a defined format
- Handling variable language where fixed rules are insufficient
Traditional software is useful for
- Deterministic rules
- Permissions
- Calculations
- Routing
- State management
- Records and audit trails
- Reliable execution of known actions
People are essential for
- Consequential judgment
- Ambiguous trade-offs
- Accountability
- Relationship-sensitive decisions
- Ethical and policy interpretation
- Novel exceptions
- Approval where error carries material impact
The workflow should combine these strengths instead of asking the model to imitate the entire employee.
Design the future workflow
1. Define the entry condition
What triggers the workflow, and which cases are eligible for AI assistance?
Not every request should enter the same path. A routine support issue and a contractual dispute may begin in the same inbox but require different handling.
2. Assemble the required context
Define which documents, records, policies and prior decisions the AI needs. Establish the source of truth and permissions before designing the prompt.
3. Define the AI task narrowly
Specify what the system should produce, for whom and under which constraints.
“Help with customer support” is too broad.
“Draft a response to eligible billing questions using the current account record and approved refund policy, then cite the information used” is operationally useful.
4. Place judgment deliberately
Decide which outputs can move automatically, which require human review and which should never be handled by the AI path.
Make the approval efficient. A reviewer should see the proposed output, relevant evidence, confidence or risk signal and available actions in one place.
5. Design exception handling
Define what happens when:
- Information is missing
- Sources conflict
- The system cannot complete the task
- The user rejects the output
- A high-risk condition appears
- An integration fails
- Cost or latency exceeds the operating limit
The exception path is part of the product.
6. Remove replaced work
If the team still completes every old step after the AI-assisted process, the company has added complexity rather than leverage.
State which steps, reports, handoffs or approvals will stop when the new workflow proves reliable.
7. Define the record
Decide what must be stored: input, sources, output, reviewer decision, final action and outcome. Traceability should match the consequence of the workflow.
Prototype with real cases and real users
Do not validate the workflow only with ideal examples.
Use:
- Common cases
- Difficult cases
- Incomplete inputs
- Conflicting information
- High-consequence exceptions
- Examples where the correct response is to stop or escalate
Observe intended users completing the work. Note where they leave the experience, build a workaround or seek information elsewhere.
The user is often revealing a missing system requirement, not resisting change.
Measure the complete flow
Track end-to-end performance:
- Time from trigger to completed outcome
- Human effort
- Successful completion
- Quality
- Rework
- Exceptions and escalations
- Cost per completed workflow
- Customer or employee impact
- Sustained adoption
A drafting step that becomes 80 percent faster has limited value if approval continues to wait two days.
Build governance into the workflow
Governance should appear where decisions happen.
Include:
- Data access controls
- Approved use cases
- Human-review requirements
- Logging and traceability
- Evaluation and monitoring
- Escalation thresholds
- Ownership for incidents and changes
NIST’s AI RMF describes risk management through Govern, Map, Measure and Manage. Applied well, these are operating responsibilities, not a document created after the system launches.
Common mistakes
Automating the visible task
The company improves generation while leaving the actual constraint untouched.
Designing from leadership assumptions
The people doing the work encounter dependencies and exceptions that never appeared in the project brief.
Requiring humans to verify everything from scratch
The new system creates a second version of the old job rather than useful review.
Preserving every old approval
Some approvals protect real risk. Others exist because the previous system lacked visibility or reliable rules.
Measuring model performance alone
The business experiences the workflow, not the benchmark.
Which operator fits?
AI Adoption Lead
Best when behavior, enablement and the operating model around the tool are the primary gaps.
Product/UX Lead
Best when the interaction, review experience or workflow design is creating friction.
Forward-Deployed Engineer
Best when the capability must be integrated into real systems and adapted alongside users.
AI Product Lead
Best when the company has not defined the user, outcome or product direction.
The mandate may require several capabilities, but one operator should own the complete workflow result.
What a useful mandate sounds like
“Add AI to the sales process” is not a complete mandate.
A stronger version is:
Redesign the strategic-account research workflow so sellers receive a traceable, decision-ready brief inside the CRM before each meeting, with human review for high-consequence claims and a feedback loop that improves quality over time.
The mandate should define the outcome, users, systems, authority, risks and measures.
What the company should retain
- A documented current and future workflow
- Clear roles for AI, software and people
- Defined human judgment and approvals
- Working integrations
- Exception and recovery paths
- Quality, cost and adoption measures
- Documentation and decision records
- A team capable of operating the new model
The best AI workflow does not make the process look futuristic. It makes the work easier to complete, easier to trust and easier to improve.