A customer asks for a change to a quote. An AI assistant finds the record, drafts a reply, and updates a task. That could be helpful. It could also quietly promise a price the team never approved.

The question I would ask is simple: which parts may the system prepare, and which decisions belong to a person?

Why the approval line matters now

On October 6, OpenAI described research with Ironclad on AI agents that configure contracting workflows. The work includes intake forms, approval rules, and reusable legal terms. Their example makes the issue concrete: a purchasing process may require Finance approval above a threshold, Security review for certain requests, and Legal review of unusual terms. The model must keep those rules in view across the whole job. OpenAI says this is a research evaluation in Ironclad's environment, not measured customer time savings.

Source: OpenAI, October 6, 2026.

The same day, OpenAI announced an expanded Atlassian partnership. It describes AI using connected project information to surface blockers and suggest next steps. Deeper features for assigning work to agents, tracking decisions, and reviewing results are described as work the companies are exploring. They are not a finished capability that every business can use today.

Source: OpenAI, October 6, 2026.

Both updates concern enterprise software. My takeaway for a smaller company is a management question, not a product recommendation: when AI can reach more of the workflow, does everyone know where its authority stops?

Make an approval map for one recurring decision

Choose a decision your team makes often, such as responding to a request to change an estimate. List the information the team needs, the actions that can be prepared, the decision that changes the customer's commitment, and the person who can approve it. Keep the map short enough that someone on a busy shift can use it.

Example: a request to change an estimate

  • AI may prepare: Pull the approved estimate, summarize the customer's request, draft options using the current price list, and flag missing details.
  • A person decides: Whether to change scope, make a price exception, promise a deadline, or send a sensitive reply.
  • The system records: Who approved the change, what version was sent, when the customer was told, and the next follow-up date.
  • The team checks: That the customer received the approved answer and the record matches it.

This is a hypothetical example, not a story about an IntelliLine customer. The exact approvals should follow your own service promises, policies, and professional obligations.

Write down the stop conditions

Define the moments that should pause the process: missing customer information, a price outside the approved range, conflicting records, a complaint, or a request that calls for care and empathy. Name the person who receives each case and the backup when that person is unavailable. Show the person the original request and the AI's work so they can make an informed decision.

A vague instruction to “ask a human” is not enough. The handoff needs an owner, a way to reach that owner, and a clear status while the customer waits. If nobody is available, the system should say what the business can honestly promise about follow-up.

Check the decision after the message goes out

Approval is one checkpoint. The team also needs to confirm the right version was sent and the next action landed in the right place. Review a small sample of completed requests regularly. If a rule changes, update the source information and tell the people who use it. Give staff a simple way to report a wrong answer or a missed handoff.

The goal is a process another trained colleague can run consistently. AI can gather context, prepare routine work, and keep a record. People keep judgment, empathy, trust, approval, and accountability.

Our operations approach starts by mapping how the work actually gets done. The solutions overview shows where AI, automation, and people can fit together.

My take

If a new tool can take an action, draw the approval map before you give it the keys. That small piece of operating discipline can spare a customer a confusing promise and spare your team a cleanup job.

Sources