A Checklist Before Starting A DX PoC
A paid PoC moves fastest when the client team brings enough context to make decisions quickly. The goal is not perfect documentation. The goal is to remove the blockers that usually slow the first week.
Most PoCs that miss their date do not miss because the technology was hard. They miss because nobody could answer a specific question in time: which field is authoritative, who can approve a draft, where is the test environment, who owns the API key. This checklist exists to surface those questions before the kickoff, not after.
Prepare The Workflow
Name the process owner
Describe the current workflow in five to ten steps
Identify the handoffs, bottlenecks, and rework
Define what success should look like after two to six weeks
The process owner is the single most important item on the list. This is not the executive sponsor and not the IT contact. It is the person who currently does the work or supervises it daily, who can answer "is this draft correct?" within hours, and who has authority to say "yes, this is good enough to ship." Without that role assigned, weekly demos turn into status meetings instead of decision meetings.
When describing the workflow, write it in the form trigger → action → output. For example: "Customer email arrives in shared inbox → support agent classifies, drafts reply, attaches relevant article → reply sent, ticket tagged in Zendesk." Five to ten steps is enough. If you cannot describe it in ten, the scope is still too big.
Prepare The Data
Bring safe sample files, screenshots, exports, email examples, or API documentation. If real data cannot be shared, create anonymized samples that still match the real structure.
In practice this usually means:
20-50 representative samples, not three. The long tail is where AI breaks; you need enough variety to see it.
Both happy-path and edge-case examples. The malformed PDFs, the angry customer emails, the records with missing fields. These are the cases that decide whether the PoC ships.
Field-level documentation for any structured source: column names, types, expected ranges, what null means, which fields are authoritative when two systems disagree.
API access or a stable export path. A one-time CSV dump is fine for week one, but at least know how the data would flow in production.
A written data-handling agreement. What can leave the corporate network, which fields must be masked, what the retention policy is for AI logs. APPI and any internal information security review should be cleared before the first model call.
Anonymization is not the same as deletion. Replace names, account numbers, and identifiers with realistic stand-ins; do not delete fields, because the AI workflow may depend on the shape of the data even when the contents are masked.
Prepare The Decision
Agree on acceptance criteria before the sprint starts. At the end of the PoC, the decision should be clear: stop, refine, integrate, or expand.
Acceptance criteria should be specific and measurable. Vague criteria like "the AI works well" turn the final review into a debate. Useful examples:
"The model classifies at least 85% of incoming tickets into the correct category, measured on a held-out set of 100 tickets."
"The drafted reply is accepted by the agent without edit in at least 40% of cases over a one-week trial."
"Extraction recall on the 20 sample invoices is at least 95% for the seven required fields, with a confidence score exposed to the reviewer."
Pair every metric with a fallback path: what happens to the cases the AI cannot handle. A PoC that ignores the bottom 10% is not a PoC, it is a demo.
Prepare The Environment
A few small infrastructure decisions remove most of week-one friction:
A sandbox environment with non-production credentials for any SaaS or internal system the workflow touches.
An identity for the AI workflow (service account, API token, or signed JWT) created with the minimum scope required.
A logging destination the client team can read — even a single shared Postgres table or Supabase project is enough for a PoC.
A demo URL or staging deploy decided in advance. Vercel, Render, Fly, or the client's own cloud — anywhere that the business owner can reach from a phone during the weekly demo.
That clarity is what turns a small PoC into a useful DX investment tool. The checklist is not bureaucracy; it is what makes "two to six weeks" an honest number instead of an optimistic one.