Conversational AI
Research and design for AI agents, where trust, tone and error handling decide whether people come back.
What it includes
The scope of the service
- Conversational agent audits: what it understands, what it doesn't, and what happens when it fails.
- Guardrails, bias and error handling: how it responds to the unexpected, the ambiguous or the sensitive.
- Accessibility in conversational interfaces (web and WhatsApp).
- Responsible AI workshops based on NIST and OECD frameworks.
How I work
The process, step by step
I adapt the scope to your case, but the logic is always the same: the question first, the data after. The best work happens in collaboration: when the project calls for it, I bring in whoever knows each topic best and contrast perspectives, because a single point of view is a blind spot.
Framing
We define what the agent has to do well and which risks are unacceptable.
Audit
I test the agent in real, edge and adversarial scenarios. I document where and how it fails.
Diagnosis
Trust, tone, error handling, bias, accessibility and transparency.
Recommendations
Prioritized conversational design and guardrail changes, with the why behind each one.
What you take away
Concrete deliverables
Audit report
Scenarios tested and failures documented.
Prioritized recommendations
Conversational design, guardrails and error handling.
Accessibility diagnostic
Of the channel where the agent runs (web or WhatsApp).
Debrief session
With product and the technical team.
When it applies
Situations where it helps
"Our bot answers well in the demo and badly with real users."
"We don't know what the agent does when someone asks it something sensitive."
"We want to launch an accessible, trustworthy assistant from day one."
How we start
Where we begin
There is no single starting point. These are the most common formats; we define the scope based on your case. I work remotely, in Spanish and English, with teams in LatAm, the United States and Europe.
Focused diagnostic
An audit of one agent, one deliverable, a short timeline. To know where it fails or see how I work before committing to more.
Full project
Audit, conversational design and guardrails end to end, with a debrief for product and the technical team.
Ongoing support
For teams iterating on their agent that need recurring review. Agreed monthly dedication.
I also run responsible AI workshops for your team. Every proposal starts with a free 30-minute conversation, to see whether it makes sense to work together.
Before hiring, try the two self-assessments for free: Is your chatbot accessible? and Is your AI responsible? They give you a first read in two minutes.
Does your agent build trust or break it?
Tell me what your agent does and which channel it runs on, and we'll see where to start.