A customer-facing AI agent should not be judged only by how many conversations it handles without a person. In a building business, the more important question is whether it knows when to stop. A homeowner asking for showroom hours is not the same as a homeowner disputing a change order. A builder checking an order status is not the same as one demanding a promise about an installation date the system cannot verify.

OpenAI's July 2026 announcement for Presence, an enterprise voice-and-chat agent product, makes the operating model unusually clear. Each deployment starts with a specific job, receives only the knowledge and system access required for that job, and follows company policies for approved actions, required approvals, and human escalation. Before launch, teams test common requests, edge cases, higher-risk scenarios, tool use, policy compliance, and handoff behavior. Presence is an enterprise product, not evidence that a small contractor should buy a voice bot tomorrow. The useful signal is the control system serious deployments require.

Build the escalation ladder before the conversation

An escalation ladder is a written map from request type to allowed action. It tells the agent what it may answer, what evidence it must retrieve, which actions need approval, and when a person must take over. If the team writes these rules only after a bad interaction, the customer becomes the test environment.

  • Level one — answer: public, low-risk facts such as hours, service area, showroom location, or a published process, using an approved source.
  • Level two — retrieve: account or project facts such as appointment status, selection status, or order status after the caller or user is appropriately verified.
  • Level three — draft: prepare a reply, callback summary, warranty intake, or scheduling request for a staff member to review.
  • Level four — approve: pause before discounts, schedule commitments, scope changes, refunds, warranty decisions, or anything that changes a record or creates an obligation.
  • Level five — hand off: transfer disputes, safety concerns, legal threats, payment problems, emotional escalation, missing evidence, and any request outside the defined workflow.

Tie every answer to the right source

A support agent needs more than a folder of documents. It needs a source hierarchy. Public policies may govern general questions. The signed contract and approved change orders may govern scope. The current project schedule may govern internal planning but not authorize a promise to the client. The accounting system may show a balance without explaining a disputed charge.

For each request type, name the governing source, freshness requirement, verification step, and safe fallback. If two sources conflict, the agent should surface the conflict and escalate instead of choosing the answer that sounds most plausible. This is where a source-grounded workflow differs from a polished chatbot.

Make the handoff useful to the person

A handoff is not successful when the agent says, ‘Someone will contact you.’ The receiving employee needs a compact packet: verified identity, request category, customer wording, project or order reference, sources checked, actions already taken, unresolved question, urgency, sentiment, and the exact decision required. The customer should know who owns the next step and the expected response window, provided that window is real.

Anthropic's guidance on effective agents recommends starting with the simplest workable design and adding autonomy only when it measurably improves results. That is a sound rule here. A reliable intake-and-handoff workflow can be more valuable than an agent that tries to resolve every situation.

Evaluate the dangerous cases, not only the common ones

Before launch, build a test set from real categories of conversation. Include ordinary questions, ambiguous identities, conflicting documents, missing records, angry customers, urgent safety language, requests for unauthorized discounts, promises about completion dates, and attempts to make the agent reveal another customer's information. Score whether it used the right source, verified identity, stayed within permissions, preserved uncertainty, escalated at the correct level, and produced a useful handoff packet.

After launch, review escalations and a sample of completed conversations every week. Track unsupported claims, wrong-source answers, missed escalations, unnecessary escalations, approval bypasses, repeat contacts, customer corrections, and time to human resolution. OpenAI describes production sessions and escalations as inputs to a controlled improvement loop in which proposed changes are tested before rollout. Small firms can copy that pattern with a spreadsheet, a versioned instruction file, and a named owner.

The operating standard is the advantage

Google says there is no special AI-only schema for AI Overviews or AI Mode. Helpful, original, technically accessible content and accurate structured data remain the foundation. The part of this Field Note an AI summary cannot replace is the operating decision a company must make for itself: which requests are safe to answer, which systems govern the answer, which actions require approval, and who takes over when the boundary is reached.

Continue with the operating system

Sources Read

Next step, if this note maps to a problem on your desk: Private Training — a private working session for your team ($1,500+).

Related Field Notes