Matching Our Services to Customers

An organization with a growing catalog has a common problem: a client-facing person may not remember every service that could be relevant to a conversation. An AI assistant can help prepare a draft menu, but only if the catalog, the opportunity details, and the review boundary are clear.

This is a proposed workflow, not a claim about a production system or a live customer result.

Define the decision

The assistant is not deciding what a customer needs or making a promise on the organization’s behalf. Its task is narrower:

Given an approved service catalog and an opportunity summary, prepare a cited draft menu of possible services and the questions a human should confirm.

That scope makes the output useful for a conversation while keeping commercial judgment, eligibility, pricing, legal commitments, and customer consent with a responsible person.

Give it only approved inputs

The workflow should start with a versioned service catalog and a small opportunity brief:

• industry or business context;
• geography or site context, when relevant;
• scope of the work;
• prior services already delivered, if approved for use;
• stated constraints such as timing or budget; and
• the catalog version and source location.

Do not put confidential customer information into an unapproved tool. Do not let an assistant treat old marketing material, remembered conversations, or general web results as a catalog of record.

Use a source-bound prompt

You are a service-matching assistant.

Using only the approved catalog provided for this run, prepare a draft menu of
services that may be relevant to the opportunity summary.

For every suggested service:
- cite the catalog section and version;
- explain the connection using only facts in the opportunity summary;
- list what a client-facing reviewer must confirm.

If the catalog or opportunity details are incomplete, ask targeted questions.
If the user chooses to proceed with gaps, list the assumptions and label the
output "Draft — needs human confirmation."

Do not invent services, capabilities, pricing, availability, legal terms, or
customer facts. Do not convert a possible fit into a recommendation without a
human reviewer.

Make the draft reviewable

A useful output has four sections:

1. Possible fits — service name, catalog citation, and a one-sentence reason.
2. What to confirm — missing facts that could change the fit.
3. Assumptions and gaps — information the assistant did not receive.
4. Run record — catalog version, run date, prompt version, and reviewer.

The run record is not busywork. It lets a team understand why a service appeared in an older draft after the catalog or opportunity changed. NIST’s AI RMF similarly emphasizes documented roles, source and context limits, and ongoing monitoring for AI-supported workflows. (NIST AI RMF Core)

Test before scaling

Before sharing this pattern broadly, test it with fictional or approved sanitized cases:

• a complete brief with a current catalog;
• a brief missing geography or scope;
• a catalog that contains a retired service;
• a case where two services sound similar but have different eligibility conditions; and
• a request that tries to make the assistant invent an offering.

The test is successful when the assistant cites the supplied catalog, asks about material gaps, and refuses to turn weak evidence into a confident recommendation.

Keep the right human work human

The assistant can reduce the time spent searching a catalog. It cannot replace a conversation about what a customer actually needs, what the organization can responsibly offer, or whether the recommendation is appropriate.

That is a good division of labor: the system organizes possibilities; the person with context, accountability, and a relationship with the customer decides what to say next.