“We need conversations that actually help.”

Conversation design

Conversation design is the work of writing the words, flows, and rules a chatbot or voice agent uses with a customer. We design those conversations so the bot answers the question the first time — in the customer's language, on the channel they already use. That means designing not just what the bot says but when it answers, asks, verifies, escalates, or hands off — our Behavior Decision Set.

What you get

  • Conversation flow documentation and intent maps (the routing logic that decides what the bot should answer)
  • Dialog scripts with persona and tone guidelines
  • Escalation paths (the bot-to-human handoff design)
  • Testing frameworks and conversation analytics setup
  • Multi-language deployment support
A designed customer conversationA short chat flow. The bot asks for an order number, gives tracking, then offers a live agent.My order is late.ASKWhat is your order number?It is 4821.ANSWERIt shipped today. Here is tracking.HAND OFFWant me to add a live agent?
Every turn is designed. The bot answers, asks, or hands a customer to a person.
“We need AI that sounds like us.”

Prompt engineering and LLM consulting

Prompt engineering is the work of writing the instructions that control a large language model (LLM) in production. Not every job needs an LLM, so we start with the Determinism Map — rule-based, LLM, or hybrid, decided per intent. Then we write the prompts and guardrails that keep answers accurate, in your brand voice, and within the limits you set, and test them against real examples before launch.

What you get

  • Production prompt libraries and system prompts (the master instructions that set the AI's role and limits)
  • Prompt testing and evaluation sets that score AI quality before launch
  • RAG pipelines (retrieval-augmented generation, the design that lets AI quote your own data instead of guessing)
  • LLM output audits and guardrail design (the rules that block off-brand or unsafe answers)
  • Few-shot example curation and tone calibration
From question to tested answerA simple flow. A customer question and your data feed a prompt with guardrails, then a tested answer.Customer questionYour dataPrompt + guardrailsyour rules, your voiceTested answer
We write the prompt, add guardrails, and test it before launch.
“We need a strategy that makes sense.”

CX strategy

CX strategy is the plan for what to automate, what to leave to people, and how to measure the difference. We build that plan, tie it to numbers your leaders already watch — containment, CSAT, handle time, cost per contact — and roll it out across teams. Part of that plan is the State Ledger: what the system should remember, update, expire, and clear, so the AI acts on what's true right now.

What you get

  • CX roadmaps tied to business goals
  • Workflow design and automation strategy
  • User research synthesis and persona development
  • Stakeholder alignment workshops
  • Adoption and success metrics frameworks
From plan to measured resultsA flow from choosing what to automate, to a roadmap, to rising results.What to automateWhat people handleCX roadmapResults we trackStartLiveTuned
We choose what to automate. Then we track the numbers your leaders watch.

How can we work together?

We offer two ways to work: an embedded team extension for ongoing work, and project-based consulting with a defined scope.

Embedded team extension

We join your product, engineering, or CX team as a dedicated AI specialist — in your standups, sprint planning, and cross-team work. This fits teams that need ongoing hands-on help.

  • Works inside your tools and rituals
  • Set hours per week or month
  • Best for multi-month work

Project-based consulting

We ship a defined scope with clear milestones. This fits teams that need a specific outcome: a chatbot design system, a prompt audit, a CX roadmap, or an agentic AI readiness review.

  • Set scope, timeline, and deliverables
  • Milestone reviews and approvals
  • Most projects run 4 to 12 weeks

How we run a project

We run every engagement in three phases: Discovery, Design and Build, and Measure and Optimize. Each phase ends with a review with you.

1

Discovery

We learn your users, your tech stack, and your goals, and we agree on what success looks like before any building starts.

2

Design and Build

We write the prompts and design the flows. You review at every milestone.

3

Measure and Optimize

We track performance with real data and keep tuning until the numbers show the value.

What is agentic AI readiness?

Agentic AI is AI that can plan a task, make decisions, and take actions on its own. Before you hand it real work, your data, your rules, and your existing chatbot setup have to be ready for that. We run the review that tells you whether they are — and what to fix first — treating the riskiest moves (stop, escalate, hand off) as deliberate decisions, not afterthoughts. See the full Conversation Behavior Framework.

AI governance and guardrails

We set the rules that decide what an AI agent is allowed to do, build the checks that catch it when it goes off-script, and define who reviews what — so every action can be traced and explained.

Data readiness review

An AI agent is only as good as the data it can reach. We check whether yours is clean enough, fast enough, and accessible enough for an agent to use it reliably — and flag the gaps.

Agent orchestration strategy

We design how several AI agents split up a complex job and hand work to each other — and where a person signs off before anything high-stakes goes through.

One framework behind every service

ICX runs every engagement through the Conversation Behavior Framework: four lenses that turn AI design judgment into something a team can review and test.

Determinism Map

Rule-based, LLM, or hybrid, decided per intent on latency, channel, memory, backend access, and compliance. See the lens →

Behavior Decision Set

The seven moves an assistant can make on any turn: answer, ask, verify, repair, stop, escalate, hand-off. See the lens →

State Ledger

What the system keeps, updates, expires, and clears, so AI acts on what is currently true. See the lens →

Turn Rhythm

Pacing and turn-taking that make a bot sound natural, because rhythm matters more than casual wording. See the lens →

Have a project in mind? Let's talk.

Get in Touch

Have questions about our process? See our FAQ.