How to introduce AI into a business workflow safely

1. Start with one meaningful workflow

Justin Watt recommends beginning with one key workflow or one department’s work. Trying to organize every piece of company information before starting can delay the learning that a smaller project would provide.

Pick work with a recognizable beginning, an output and an owner. Explain which steps the pilot will change and where people will remain responsible. A shared workflow needs more thought than giving one person a chatbot account, because multiple people will depend on the same information and rules.

• Name the workflow and the person responsible for its result.
• Define the input, expected output and people who receive it.
• Keep the initial team and use case small enough to evaluate.

From the interviews: Justin Watt · 35:22

2. Map the handoffs and decision rules

Eddie Irvin describes asking where information comes from, who makes a decision and who must sign off. That process clarity helps a builder distinguish steps that can be automated from steps needing human involvement.

Walk through the work with the people doing it. Include the exceptions that an experienced employee handles without thinking. The useful output is a shared understanding of the process, not a diagram that hides the difficult decisions.

• Where does each input come from, and which version is authoritative?
• Who handles the next step and what do they need to know?
• What happens when information is missing or conflicting?
• Which decisions need approval before the work proceeds?

From the interviews: Eddie Irvin · 12:28

3. Test access as the people who will use it

Justin gives an example of an internal HR chatbot built with operations-level access and then released more broadly. People could retrieve information they should not have been able to see. The example shows why a tool working for its builder does not establish that it is ready for every employee.

Involve someone with the technical skills to review shared access. Define who can read and write each source, and test with the roles that will actually use the workflow. Apply the same questions when an assistant connects to several systems or takes actions on someone’s behalf.

• Check employee, manager and administrator access separately.
• Confirm that shared outputs respect the intended audience.
• Check write permissions as carefully as read permissions.
• Recheck access when another team joins the pilot.

From the interviews: Justin Watt · 21:53

4. Decide which information can reach the model

Steven Walchek describes an approach that detects and de-identifies sensitive information before it reaches a model, then restores the relevant information within the organization’s environment. This is his description of Liminal’s approach, not a claim that every AI tool provides those controls.

The practical question for your workflow is what information leaves your environment and what control applies before it does. Review the actual tool and data flow with the people responsible for information security. Give employees an approved way to use AI and explain which information belongs in that workflow.

• Identify the information needed for the task and remove unnecessary inputs.
• Check the tool’s actual data handling and connected systems.
• Agree which sources and tools are approved for this workflow.
• Explain the rules to the people who will use them.

From the interviews: Steven Walchek · 12:08; Steven Walchek · 24:55

5. Make human review a defined part of the work

Ronnie Coleman describes AI working with experts who provide the context and judgment that the model cannot simply download. Review works better when the person knows which output they are responsible for checking and what would make it unacceptable.

Define the point at which a person approves a draft, resolves an exception or authorizes an action. Include how to correct errors and how the team returns to the existing process if the pilot is not useful. These are practical design questions synthesized from the interviews, rather than a universal implementation formula.

• Name the reviewer and the decision they own.
• Specify the quality checks needed before an output is used.
• Explain how the team reports errors and handles exceptions.
• Keep a clear way to continue the work when the AI step is unavailable.

From the interviews: Ronnie Kwesi Coleman · 7:35; Eddie Irvin · 12:28

6. Measure the workflow before expanding it

Justin describes establishing shared norms in a smaller area before adding departments. Conflicting templates and definitions can make AI outputs less useful as the system grows. Agree the source of truth and check whether it remains clear when another team joins.

Compare the pilot with its starting point using preparation time, quality, rework and the business result that matters to the owner. Keep review and support costs in the comparison. Use the companion prioritization guide to decide whether the evidence supports expansion.

From the interviews: Justin Watt · 35:22; Ronnie Kwesi Coleman · 15:22

Worked example: preparing a customer briefing

Illustrative exercise, not a reported customer result: a team wants AI to prepare a briefing from approved account notes. Start with one team, one briefing format and the information its members are permitted to access. An account owner checks the draft before anyone sends it.

Test missing or contradictory notes and users with different access. Record the time saved after review, the corrections needed and whether the briefing helps the next conversation. Expand to another team only after checking that its information and access rules fit the workflow.

• Input: approved account notes with an agreed source of truth.
• AI role: prepare a draft with links back to the available information.
• Human role: verify the content and approve sharing.
• Evaluation: preparation effort, review effort, accuracy and usefulness.

Keep building your approach

Use the prioritization guide to agree the baseline, business outcome and decision about expansion.

Explore AI adoption, AI governance, AI agents, AI leadership.

Get the next conversation →