Preserving Your Brand & Customer Service In the Age of AI

A financial-technology company engaged Mercury to define how its brand should behave when an AI system answered, refused, admitted uncertainty, and handed a customer to a human.

The work connected a governed brand record to a behavior contract and built the sourced knowledge, answer policy, escalation paths, and evaluation coverage that made those behaviors enforceable and reviewable.

The company was preparing an AI system to participate in customer support. The harder issue was what would happen when software spoke under the company's name to people asking about financial services. Each answer would carry the brand into a moment of trust, judgment, and possible risk.

Mercury's objective was to make that moment governable. Success meant sourced answers, visible uncertainty, clear refusal rules, and escalation paths that sent the right conversations to people without dropping context. Review records and testing evidence also had to show when system conduct drifted from approved behavior.

The company's brand existed where brands usually live. A brand guide governed mission and vision, identity, voice and messaging, implementation rules, and the visual categories of logo, color, typography, and imagery. Those materials can keep a website and sales motion consistent. They cannot tell a support system what to do with a two-word customer question, a partially grounded answer, or a request that touches a subject a machine should not address.

Once AI enters a customer conversation, brand consistency stops being mainly visual. Customers experience the brand through behavior: whether an answer is sourced, whether the system admits what it does not know, whether it declines with the right tone, whether privacy boundaries hold, and whether a person can step in with enough context to help.

The risk lived at the edges, where a customer might ask about money, account status, or regulated territory. A wrong answer in a polished interface still feels like the company speaking. Short queries also exposed source gaps and answers that drifted from their evidence. Tone guidance alone could not resolve those failures.

The work began with the existing brand. Mercury extracted a governed brand record from the company's guide covering mission and vision, identity, voice and messaging, implementation and governance, and the visual categories. The record is review-ready and held from public use until the company approves it.

That comparison made the missing layer clear. Rules for voice had to be joined with rules for system conduct. Mercury shaped that layer as a brand behavior contract: what the system may say, what it must source, what it must refuse, when it escalates, what humans review, and what evidence proves those rules still hold. The visual identity governed how the company appeared. This contract governed how the company acted when software spoke for it.

The first behavior to make enforceable was sourcing. An unsourced answer is improvisation delivered in the company's voice, so the support model was organized around knowledge-base integration and sourced responses. As the team connected answers to sources, gaps became visible. Instead of allowing the system to cover those gaps with confident language, Mercury documented them and routed them into governed knowledge work.

Sourcing defined the answerable range. Mercury then built an answer-policy layer for the edge cases: topics the system should refuse, questions with thin source support, abstention behavior, and scripted flows where wording should stay controlled. Privacy, regulated-industry exposure, PII, CRM boundaries, and retention shaped this work as design constraints. They are described here only as constraints, not as certified controls or audited outcomes.

The next question was what happened after a refusal or abstention. A customer who reaches the edge of the system still needs help. Mercury built escalation paths and human off-ramps as first-class routes: the system needed to recognize when a conversation belonged with a person, stop trying to resolve it alone, and deliver the handoff with useful context intact.

The stopping point became a brand decision. Helpful answers mattered, and so did the moment a customer needed a person and the system either made that transfer cleanly or created more friction.

The contract then had to be testable. Mercury built evaluation coverage for safety, grounding, directives, flow behavior, tone, and model output. Results were preserved on review surfaces, and regression evidence made changes visible over time.

Testing exposed short-query retrieval weaknesses, faithfulness issues, and knowledge gaps. Each finding was documented, remediated through governed work, and fed back into knowledge, answer policy, and future evaluation. The evidence trail showed where the system needed correction.

The engagement produced a review-ready brand behavior model tying a governed brand record to sourced responses, answer policy, scripted flows, human off-ramps, evaluation coverage, review surfaces, regression evidence, and remediation.

The remediation trail is part of the outcome. The engagement shows the system being wrong in the ways customer-facing AI systems become wrong: short questions that expose retrieval assumptions, answers that drift from sources, and gaps where the knowledge base does not yet support a confident response. Because those failures were documented, traced, and repaired through the governed process, the company gained a more inspectable way to maintain the brand's behavior over time.

Brand review could now ask whether the system behaved correctly: whether answers stayed sourced, uncertainty showed, refusals held, sensitive situations moved to a person, and handoffs preserved context.

This case does not claim customer-facing rollout, production history, satisfaction, deflection, accuracy, response time, cost savings, compliance status, uptime, staffing impact, or ROI.

The engagement began with a practical concern: the company did not want its brand to become whatever a language model happened to say in a support conversation. It ended with the brand written into behavior, connected to the system that would express it, and checked by evidence people could review. The clearest expression of that behavior came when the system recognized the edge of its authority and handed the customer to a person.

For companies whose brands will increasingly be experienced through software behavior, Mercury builds governed AI systems following this pattern: brand record, behavior contract, sourced knowledge, answer policy, escalation, evaluations, and remediation.