Four things we do, start to finish

Most of our work comes down to four moves. Here is how each one actually runs.

Designing an intent model

An intent model is the map that decides what a bot should answer. We build it from what customers actually say, not from guesses.

  1. 1Pull real transcripts
  2. 2Cluster what people actually ask
  3. 3Map each intent to an answer and a handoff
  4. 4Test against messages the model has not seen

Writing a production system prompt

The system prompt is the master instruction that sets the AI’s role and limits. We write it to be specific, then prove it holds up.

  1. 1Define the role and the hard limits
  2. 2Draft the prompt and add guardrails
  3. 3Build an evaluation set of real cases
  4. 4Tune until it passes every case

Building guardrails

Guardrails are the rules that keep an AI inside the lines. We make the unsafe answers impossible, not just unlikely.

  1. 1List what must never happen
  2. 2Write the rules and the refusals
  3. 3Add monitoring on live output
  4. 4Set the points where a person signs off

Running a discovery sprint

Before we build anything, we find out where the real value is. A discovery sprint takes two to four weeks.

  1. 1Audit the current experience
  2. 2Baseline the metrics that matter
  3. 3Find the highest-value fixes
  4. 4Deliver a roadmap you can act on

The enterprise work behind ICX

These are from Christi's prior enterprise roles, before founding ICX. Client names are withheld. They are the reason the practice exists.

A multilingual virtual agent for 20M+ users

The problem. A global tech company gave inconsistent support across eight markets and languages. Customers in different regions got different answers.

The approach. Before founding ICX, Christi built the multi-language intent architecture, the escalation paths, and the production prompts behind it.

What changed. One consistent experience across all eight markets, faster answers, and a system that held up at scale.

Faster responses25%
20M+Users served
8Global markets

Self-service that pulled volume off live agents

The problem. A mid-market e-commerce team was drowning in tickets, and the existing bot pushed more work onto agents than it took away.

The approach. Christi rebuilt the customer-facing AI end to end: new flows, guarded prompts, and a self-service strategy aimed at the questions people asked most.

What changed. More customers solved their own problem, happier customers, and far fewer tickets reaching a person.

CSAT improvement40%
Self-service adoption35%
Ticket reduction60%

Example artifacts

Since we can't show client deliverables, here is the kind of work we produce, written from scratch as teaching examples.

A customer-support system prompt

Illustrative example, not client work
ROLE
You are the support assistant for ACME, an online electronics store.
You help customers track orders, start returns, and answer product questions.

RULES
- Answer only from the ACME help center and order data you are given.
- If you do not have the information, say so and offer to connect a person.
- Never invent an order number, a price, a policy, or a delivery date.
- Never give legal, medical, or financial advice.

WHEN YOU ARE UNSURE
- Ask one clarifying question, then stop and wait.
- After two failed attempts, hand off to a human agent with a summary.

TONE
- Plain, warm, and brief. Short sentences. No corporate filler.
- Match the customer's language. Apologize once, then fix the problem.

An intent map

Illustrative example, not client work

The routing logic that decides what the bot should do with each request.

IntentExample messagesBot actionHandoff?
Track order"Where is my order?", "Has it shipped yet?"Look up status from order data, share tracking linkNo
Start a return"I want to return this", "It arrived broken"Check return window, generate labelIf outside window
Billing dispute"I was charged twice"Confirm the charge, log the caseYes, to a person
Out of scope"Can you write my essay?"Decline politely, restate what it can help withNo

A guardrail spec

Illustrative example, not client work

The rules that keep answers safe and on-brand, and what the bot does when one trips.

RuleWhat it blocksBehavior when triggered
No invented factsPrices, dates, or policies not in the source dataSay it isn't sure and offer a handoff
No off-topic adviceLegal, medical, or financial guidanceDecline and redirect to support topics
Stay in voiceSlang, jokes, or claims that break brand toneRewrite in the approved tone before sending
Escalate angerRepeated frustration or a threat to leaveHand off to a person with the full thread

A fallback flow

Illustrative example, not client work

What happens in the moment the AI doesn't understand, which is where most bots lose people.

1. Bot misunderstands the message.
2. Bot rephrases the question in plain words and asks the customer to confirm.
3. Customer still stuck?  Offer the top 3 things this bot can actually do.
4. Customer picks one  ->  continue in that flow.
5. Customer asks again, no match  ->  hand off to a person with the full thread,
   so the customer never has to repeat themselves.

Deflection estimator

A quick planning tool. Put in your own numbers to see how many conversations better self-service could take off your team's plate. It uses your figures and shows the math, nothing more.

2,000more conversations handled without a person, every month
Share of conversations contained65%

That is about $16,000 a month in agent time, using your cost per contact.

A planning estimate from your inputs, not a guarantee. Real results depend on your content, your traffic, and the design.

Want this kind of work on your team?

Get in Touch

Discovery calls are free. See what we do.