AI Engineer World's Fair 2024
Cooking with fire without burning down the kitchen
Loading slides…
Read the talk
Cooking with fire without burning down the kitchen
Twilio’s emerging technology team learned that incubating AI products requires more than fast prototypes: it requires real usage, explicit expectations, flexible roadmaps and shared learning.
From a talk by Dominik Kundel
What is our plan about AI?
The boss asks what the company is going to do about AI. Now someone has to turn that pressure into a product strategy—without exposing customers to experiments that undermine their trust. That was the situation Dominik Kundel faced at Twilio: how to work with a powerful, rapidly changing technology without burning down the kitchen.
At the time of the talk, Kundel led product and design for Twilio’s Emerging Tech and Innovation team. He describes a 16-person team spanning engineering, product, design and go-to-market. It operated as independently as possible alongside Twilio’s two established businesses: communications APIs for SMS, voice and email, and a customer data platform that helps companies understand whom they are engaging. Its remit was to explore the future of customer engagement through emerging technologies.
That did not make it Twilio’s sole AI team. Under the CustomerAI vision, AI was a capability to spread across products so customers could engage their own customers more effectively. The emerging technology team needed a narrower reason to exist than simply working on AI.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Improving the camera, replacing the camera
Innovation remained everyone’s responsibility. The distinction was between sustaining innovation and disruptive innovation, drawing on the framework associated with The Innovator’s Dilemma. Sustaining work improves an existing product for established customers: think of Sony making a better mirrorless camera. Disruptive offerings can begin with worse performance on those customers’ criteria while meeting the needs of a niche. Early Motorola phone cameras were hardly a reason to leave a dedicated camera at home; eventually, Kundel could take a vacation carrying only his phone. In the broader disruptive innovation framework, that overlooked foothold and subsequent movement into established markets matter—not poor quality by itself.
AI was developing along both paths. Photoshop’s Generative Fill augmented an existing workflow. Apple’s announcements embedded AI in familiar products. Kundel placed agents and custom GPTs closer to the disruptive side: potentially useful in new ways, but, in his assessment at this 2024 talk, not yet at the quality required for enterprise deployment.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A different baseline for enterprise and small businesses
Customer-facing failures made the quality problem concrete. Kundel cited a dealership chatbot producing an apparent offer for a Chevy Tahoe for $1, and Air Canada’s chatbot inventing a policy. The dealership vendor’s account describes prompted pricing text, not a completed vehicle sale. These examples illustrate the consequences of unreliable answers; they do not establish that either system was an autonomous agent or used a particular retrieval architecture. Kundel’s broader concern extended to RAG-based chatbots as well as agents.
Reliability was only one barrier. Regulatory uncertainty gave legal teams substantial influence over adoption. Meanwhile, enterprises often wanted AI to save money, but high-quality agents remained expensive in Kundel’s experience, even as lower-cost RAG chatbots improved. The desired business outcome and the available technology were pulling in different directions.
Startups and small and medium-sized businesses could have a different baseline. If support and sales were already stretched or inadequate, an agent could add a capability the business otherwise lacked. The comparison was not necessarily with a mature, well-staffed operation.
| Buyer context | Primary attraction | Main tension |
|---|---|---|
| Enterprise | Reduce existing costs | Quality requirements and agent expense |
| Startup or SMB | Add support or sales capacity | Improve an already inadequate experience |
That made smaller businesses a plausible initial niche for capabilities that established buyers were not yet ready to adopt.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Why explore something that does not pay yet?
An incumbent has sound reasons to prioritize sustaining work: customers already pay, problems are comparatively well defined, and the market is more predictable. Disruptive exploration offers fewer of those advantages and may generate no revenue initially. Yet Kundel argued that AI’s rapid quality improvements made ignoring it especially risky. A product that failed a quality threshold yesterday could cross it tomorrow, turning an unimportant competitor into a serious concern.
He pointed to ChatGPT and Perplexity’s pressure on Google search, with AI Overviews illustrating Google’s response. This supplied the team’s distinct mandate: investigate what might disrupt Twilio, prepare the company to respond, potentially create the disruption itself, and circulate what it learned.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Connecting conversations to customer memory
That mandate emerged through the work rather than preceding it. The group began the previous year as a three-person skunkworks team. Its initial assignment was to prototype a bridge between Twilio’s customer data and communications platforms: a customer engagement flywheel. The team would quickly build something, show it to customers, learn about their future direction as well as their present problems, and iterate.
Generative AI fit that bridge naturally. Communications contain unstructured language; customer profiles contain structured data. A JSON object is useful to software but is not a conversation, while a conversation contains information that is difficult to analyze or reuse until it has some structure. LLMs can translate in both directions, turning profile data into language and conversational information into profile data.
The team developed two complementary concepts, iterated with customers, and handed them to R&D to take to market:
- AI Personalization Engine: retrieval-augmented generation over a customer profile in Segment, making stored customer information available when generating an interaction.
- AI Perception Engine: extracting useful information from communications into the customer profile, so knowledge does not remain trapped in conversations or scattered human-agent notes.
Together, they formed customer memory. As a small illustrative data example, a conversation could yield a proposed profile addition while preserving an existing identifier:
json
{
"conversation": "Please contact me by email rather than phone.",
"profileBefore": {
"customerId": "customer-42"
},
"proposedProfile": {
"customerId": "customer-42",
"preferredContactChannel": "email"
}
}
The perception direction proposes structured information from the conversation. The personalization direction can later use that stored preference as context. The proposal above is not an executed profile update.
The handoff exposed a mismatch. The concepts still had quality and cost shortcomings, and their strongest use case served the emerging market for agents and generative AI applications. Positive feedback did not give them enough momentum to outrank sustaining work and profitability goals inside established R&D teams.
Customer conversations remained valuable: they suggested the team was addressing needs customers expected to have next, even if those needs were not yet large purchasing priorities. But a promising demonstration was insufficient. The incubation team needed to develop the ideas further before handing them over.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Let customers use the rough version
The next project, Twilio AI Assistants, built an agent builder on top of customer memory for omnichannel customer engagement chatbots. Rapid prototyping, working demonstrations and internal and external feedback remained central. What changed was who operated the prototype.
Instead of relying primarily on demonstrations, the team ran internal hackathons and gave customers visiting the office access to rough prototypes for a day. Participants could encounter the rough edges directly and tell the team both what was wrong and what they wanted the product to do. This was more exposing than a guided demo, but also more informative.
The team also sought internal uses, beginning with a relatively low-risk IT helpdesk application. An internal helpdesk did not perfectly represent the expected buyer. It nevertheless provided usage data that could expose quality problems and inform how the product should be structured. The purpose of dogfooding was to learn from actual use while keeping the initial stakes manageable.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Release early without promising maturity
Putting the product into users’ hands required an engineering organization capable of supporting it. It also brought the team into contact with Twilio’s established software development lifecycle, built to earn customer trust. Generative AI changed the sequence on which those expectations rested.
| Development approach | Before initial release | After initial release |
|---|---|---|
| Traditional model development | Gather data; train to a quality threshold | Optimize the released model |
| GenAI prototyping | Build a prototype or MVP using an existing model | Gather customer feedback; iterate |
In Kundel’s comparison, generative AI removes the requirement to assemble an initial training dataset before building a useful prototype. But that creates a dilemma: the system needs exposure to real users to improve while still falling short of the company’s normal quality bar.
The team looked to GitHub Next and Cloudflare Emerging Technologies and Incubation for precedents. Kundel associated those groups with Copilot and with Workers, D1 and Workers AI, respectively. The useful organizational pattern was to establish appropriate expectations about reliability, capabilities and availability for experimental work.
Twilio’s answer was a separate sub-brand, Twilio Alpha. That is the historical name used in the talk; the currently served AI Assistants announcement uses Twilio Forward. Alpha gave the team a way to ship early and often with explicit expectations, while making its work visible enough for interested customers to approach it rather than requiring the team to recruit every participant individually.
Operationally, that meant developer-preview waitlists, visible offerings and quick onboarding, with internal proofs of concept providing additional learning while customers came aboard. Kundel did not claim the team had discovered a release cadence fast enough for AI’s pace. It had established practices intended to help it move faster.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Own the problem, keep the solution flexible
Hiring emphasized curiosity and creativity over existing AI experience. Drawing on the view of developers expressed in Twilio co-founder Jeff Lawson’s Ask Your Developer, the team wanted engineers who could solve business problems as well as coding problems. Product should be able to give engineers a problem to investigate, rather than making them wait for a predetermined solution to implement.
Technical flexibility began with the assumption that a current model might soon become redundant. That did not mean pursuing every release or paper. It meant understanding the system’s known limitations and checking whether a new model could address them. Kundel reports that the team decided within a day of Claude 3.5 Sonnet becoming available whether it deserved immediate attention or should go into the backlog. That was a triage decision, not a claim of completing deployment or a production evaluation in a day.
Roadmap flexibility required similar discipline. As a new team without pre-existing customer commitments, the group had room to change direction, and it tried to defend that room. Kundel distinguished expectations that persist from expectations that move with the technology:
- Trust and safety: durable customer requirements worth committing to on the roadmap.
- Multimodality: a changing set of demands as new capabilities appear, requiring more freedom to adapt.
The shift in attention from images to voice illustrated how quickly a desired capability could change. Roadmap iteration therefore needed both flexibility and tolerance for failed experiments.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Dominik Kundel’s conference slides cover Twilio’s incubation process, customer experiments, roadmap flexibility and organizational lessons.
The developer-preview announcement explains Customer Memory, Segment integration, tools, knowledge sources and human handoff.
GitHub’s research group presents experimental approaches and tools for software development.
The June 2024 announcement describes the original Claude 3.5 Sonnet model and its contemporary capabilities.
Further reading
- Disruptive Innovation TheoryArticle
The Christensen Institute explains sustaining innovation, low-end disruption and new-market disruption.
The chatbot vendor describes the December 2023 prompting incident that produced a one-dollar car price.
Full transcript
Loading transcript…