Four-Part Bottom-Up AI Roadmap
Prove a prompt, add knowledge, automate it, then unify the experience
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 6
- Confidence
- 99%
The roadmap increases technical sophistication only when usage proves that the previous layer creates value. Begin with a specialized chat containing a clear prompt, even if users must manually gather and paste information. Add a vector database and retrieval when the capability requires a larger, consistent knowledge base. Once users repeatedly rely on the capability, connect triggers, data sources, and destinations so the workflow runs with less manual effort. Finally, combine several validated workflows into a coherent copilot that routes requests and exposes the right capability without requiring users to understand the underlying architecture. This bottom-up sequence makes each investment evidence-based. It also separates capability validation from integration work, reducing the risk that a costly, polished application merely automates a weak idea.
Origin
Extracted from Marketing Against The Grain. Ethan Dewal described the roadmap Asana used after testing specialized chats for outbound messaging, call preparation, call follow-up, and product questions.
Core principles
- 01Complexity must follow validated value
- 02Each architectural layer should solve a demonstrated limitation
- 03Manual prototypes create learning cheaply
- 04Automation belongs behind proven capabilities
- 05A copilot should unify useful workflows rather than conceal unproven ones
How to run it
- 1
Prove the prompted chat
Create a specialized chat with one purpose and a guided prompt. Ask users to manually provide the necessary context while evaluating whether the output is consistently useful.
Pro tip Keep the first version transparent so users understand what it can and cannot do.
Watch out Do not hide an unreliable capability behind a polished interface.
- 2
Add retrieval when needed
Connect the chat to a vector database when the use case depends on more information than users can reliably provide. Curate the underlying data so retrieval remains consistent and relevant.
Pro tip Use retrieval to solve a demonstrated data-volume problem, not merely because RAG is fashionable.
Watch out Poor or contradictory source data will undermine the output regardless of the model's quality.
- 3
Validate internal product-market fit
Measure whether the enhanced capability solves a recurring problem and earns repeated use. Refine the prompt, sources, and output before automating the surrounding workflow.
Pro tip Track repeat usage alongside qualitative examples of useful outputs.
Watch out One successful demonstration does not establish product-market fit.
- 4
Automate the workflow
Connect events, systems, and destinations around the validated capability. For example, an upcoming meeting can automatically trigger preparation and deliver the result where the representative works.
Pro tip Automate the minimum number of handoffs needed to remove meaningful friction.
Watch out Automation amplifies weak outputs as readily as strong ones.
- 5
Unify proven workflows
Bring multiple validated automations and chats into a coherent copilot experience. Let users state their objective while the system routes the request to the appropriate capability.
Pro tip Design around user intent rather than around the names of internal tools.
Watch out Do not bundle immature workflows simply to make the copilot appear comprehensive.
- 6
Iterate from usage
Observe how people invoke the copilot, where they abandon it, and which workflows create measurable value. Use that evidence to refine routing, interfaces, and the next capabilities.
Pro tip Treat the copilot as an evolving product rather than a completed deployment.
Watch out The appropriate phase-four experience is still uncertain, so avoid irreversible architectural bets.
In the wild
A team first validates a chat that creates useful call-preparation documents from manually pasted information. After representatives adopt it, the team connects the calendar and relevant data systems. Every qualifying meeting then triggers the same proven capability and delivers the preparation document automatically.
→ Manual data movement disappears without changing the validated core capability.
Several independently validated capabilities handle account research, call preparation, follow-up, and product questions. The organization combines them behind one interface where a representative can ask for help without knowing which chat or workflow performs the task.
→ Representatives gain one intent-driven interface for multiple proven sales workflows.
Common mistakes
Jumping straight to a copilot
A broad interface cannot rescue component workflows that users do not value. Validate each capability before combining them.
Automating before validation
Deep integrations make iteration expensive and can distribute poor output at scale. Confirm the manual workflow first.
Launching a million-dollar back-room project
Large hidden builds delay user feedback and commit resources before product-market fit exists. Experiment visibly and cheaply at the front of the organization.
Is it for you?
Best for
It is best for organizations progressing from early AI experiments toward integrated internal software.
Not ideal for
It is not ideal for a single fixed automation that requires neither substantial knowledge retrieval nor a shared user interface.
From the transcript
“the simplest architectures chat with a prompt uh a little more complex chat with a prompt and a vector database below that's rag”
“third you add automation to this uh which is an AI workflow and then ultimately maybe you take a couple workflows together bring them into…”
“try to avoid the million dooll back room project”
From the episode
He Automated His Sales Job With Ai… So His Boss Promoted Him