StackLift
All playbooks

Playbook

AI Project Discovery Checklist

Use this checklist before asking for an AI software estimate. It helps define the business goal, data sources, model behavior, users, integrations, risks, and launch constraints.

Best reader

Founders, CTOs, operators, and product leaders preparing an AI build

Outcome

A clearer brief that can become an architect-reviewed AI software estimate.

Use this sequence

1

Define the business decision or workflow the AI must improve.

2

List the data sources the AI can use and who owns them.

3

Identify users, roles, permissions, and human review points.

4

Document required integrations, constraints, and compliance needs.

5

Decide what must ship in version one and what can wait.

Start with the business workflow

An AI project should begin with the workflow it improves, not the model it might use.

Name the decision or task

Define success in business terms

Separate automation from assistance

Map the knowledge sources

The quality of an AI system depends on the data it can access and the rules around that access.

Documents and PDFs

CRM, ERP, or database records

Website and support content

Data freshness and permission boundaries

Plan the review loop

Production AI needs human review when confidence is low, risk is high, or business impact is meaningful.

Approval points

Fallback behavior

Audit trail

Escalation owner

Start in 60 seconds

Turn this checklist into a scoped plan.

Answer a few questions and get an AI-assisted, architect-reviewed scope, cost range, and timeline for your software project.

Frequently asked questions

What should be ready before estimating an AI app?

You should know the workflow, target users, data sources, integrations, and review requirements. A perfect PRD is not required, but clear constraints make the estimate more reliable.

Is model selection the first step?

No. Model selection comes after workflow, data, risk, latency, cost, and governance needs are clear.

The complete guide

AI Project Discovery Checklist: the complete guide

Updated November 18, 2026 ยท Published by , Founder, StackLift AI. Written with AI assistance; how we make content.

Short answer

An AI project discovery checklist is a short list of questions you answer before anyone estimates or builds. It covers the workflow the AI improves, how you will measure success, the data it can use, who may see what, where a person reviews its output, which systems it connects to, and what ships in version one. Clear answers make the estimate far more reliable.

What is AI project discovery, and why do it before estimating?

Discovery is the work of turning an idea like "we want AI in our support process" into a brief a team can price and build. It is not a long research project. It is a set of decisions about the workflow, the data, the users, the risks, and the first release.

AI projects need this more than ordinary software does. The behavior of the system depends on data you may not have inventoried, on judgment calls about what the AI may do alone, and on a quality bar you have to define yourself. If those questions stay open, they come back mid-build as change requests. A good brief is what makes a software estimate believable.

You do not need a perfect requirements document. You need clear constraints and honest gaps. Mark what you do not know as an open question with an owner, instead of writing "TBD" and hoping.

How do you run AI project discovery, step by step?

The five steps below are the backbone of the checklist. Each one produces something you can paste into a brief. Work through them in order, because later steps depend on earlier answers.

  1. Define the workflow or decision the AI must improve. Output: a one-paragraph problem statement and a success measure.
  2. List the data sources the AI can use and who owns each one. Output: a source inventory with owners, freshness and access rules.
  3. Identify users, roles, permissions and human review points. Output: a role table and an approval map.
  4. Document integrations, constraints and compliance needs. Output: a systems list and a constraints list.
  5. Decide what ships in version one and what waits. Output: a scoped first release and a later-phase list.

A working session of a few hours with the right people usually covers the first draft. If it feels too big to finish, that is a signal to narrow the workflow, not to skip steps. Our AI Scope Sprint is a 3-day version of this work if you want a team to run it with you.

Who should be in the room?

Discovery fails when only one function writes the brief. The business owner knows the pain but not the data. Engineering knows the systems but not the process. Legal and security know the limits but arrive late. Get each voice in early.

Roles to involve and what each contributes
RoleWhat they bringQuestions they answer
Business owner or sponsorThe goal and the budgetWhat outcome justifies this? Who approves scope?
Process expertHow the work is done todayWhat are the exceptions? What does a good result look like?
Data or systems ownerAccess and quality of sourcesWhere does the data live? How current is it? Who may see it?
Engineer or architectFeasibility and integrationWhat must connect? What is hard or risky?
Security, legal or risk reviewerRules and limitsWhat data may leave the building? What needs an audit trail?
End usersDay-to-day realityWould you actually use this? What would make you stop?

How do you define the workflow and what success looks like?

Start with the work, not the model. Name the decision or task: who does it now, how often, what inputs they use, and what output they produce. Then decide whether the AI will automate the task, assist the person doing it, or only draft for review. Those three choices lead to very different builds.

Success has to be measurable before the project begins. Capture a baseline now, such as how long the task takes or how often it is redone, so you can separate the AI's effect from other changes. Without a baseline you will argue about results later.

  • Name the single workflow for version one, in one sentence.
  • State who does it today and how many people touch it.
  • Choose automate, assist, or draft-only for each step.
  • Write two or three success measures a manager could check.
  • Record the current baseline for each measure.

Also ask whether AI is the right tool. If the steps never change, plain workflow automation is cheaper and easier to audit. If the work is mostly answering questions from your content, a chatbot may be enough.

What data does the AI need, and is it ready?

Data is the most common reason AI projects slip. A system that answers from your documents, records or tickets is only as good as that material: current, complete, trusted, and correctly permissioned. Check quality and access, not only existence.

Source inventory template
FieldWhat to writeWhy it matters
Source and typeDocuments, CRM or ERP records, database tables, website, ticketsSets the integration and ingestion approach
OwnerA named role responsible for itSomeone must approve access and fix errors
AuthorityIs this the official source or a copy?Conflicting sources produce conflicting answers
FreshnessHow often it changes and how updates arriveStale content causes wrong answers
Access rulesWho may see which recordsThe AI must respect the same limits people do
SensitivityPersonal, financial, health or confidential dataDecides what may go to a model provider
Format and cleanlinessScanned PDFs, duplicates, missing fieldsDrives preparation effort

If the system retrieves from documents, the RAG glossary entry explains the pattern and why permissions and freshness matter. Do not add new sources mid-build before the first version shows how it performs on the ones you started with.

Where does a person review the AI's work, and what are the risks?

Decide where humans sit in the loop before you design the screens. Review should happen when confidence is low, when the impact on money, customers or compliance is meaningful, or when the action cannot be undone. See the human-in-the-loop entry for the pattern.

  • Approval points: which outputs wait for a person before they take effect.
  • Fallback behavior: what the system does when it is unsure, such as asking a clarifying question or handing off.
  • Audit trail: what is logged so you can explain any decision later.
  • Escalation owner: the named person or team who gets the hard cases.
  • Failure examples: five to ten real cases that must never go wrong.

Write the quality bar as test cases from your real work, not as a feeling. A small set of representative tasks with known good answers becomes your launch gate and your regression check after every change.

What integrations, constraints and compliance needs should you document?

List every system the AI must read from or write to, and say which direction. Read-only connections are far simpler than systems where the AI creates or changes records. Note who owns each system's API and whether it has rate limits, sandbox access or a change freeze.

Constraints to capture before estimating
AreaQuestion to answer
SystemsWhich tools must connect, and does each have a usable API?
HostingMust it run in a specific cloud, region or private network?
Model providersMay data go to a third-party model provider, or must it stay private?
ComplianceWhich regulations apply, and what do they require of logging, retention and access?
Latency and volumeHow fast must answers arrive, and how many requests are expected?
Cost of AI usageWhat monthly budget is acceptable for model calls? See model pricing.
Ownership after launchWho monitors quality and updates content once it is live?

You do not need to resolve every compliance question yourself. Naming which rules apply lets the plan cover them from the start. For a deeper governance view, use the enterprise AI readiness checklist.

How do you decide what ships in version one?

Sort every capability into must-have, should-have and later. Version one should prove the core workflow with real users and real data, not cover every edge case. A smaller first release also gives you evidence about quality before you invest further.

  • Must have: the one workflow, its data sources, its review step, and the measures you will track.
  • Should have: extra sources, extra user roles, reporting views.
  • Later: new channels, new workflows, deeper automation, fine-tuning.

When the brief is ready, run it through the free estimate, which scopes a project feature by feature and returns cost, timeline and team in about five minutes. A senior architect reviews every StackLift estimate before it becomes a commitment. For cost drivers on AI builds, read the cost to build an AI app.

What mistakes should you avoid, and how do you know discovery is done?

Common mistakes and the fix
MistakeFix
Choosing the model firstSettle workflow, data, risk and cost needs; choose the model after
Scoping five workflows at oncePick one and sequence the rest
Assuming the data is readySample it and check access, freshness and permissions
No success baselineMeasure today's process before the pilot
Leaving review rules for laterDefine approvals and fallbacks before design
No owner after launchName who watches quality and updates content

You are done when someone outside the room can read the brief and explain the workflow, the data, the users, the review points and the first release without asking you questions. Open questions are fine if each has an owner and a date. If gaps remain that you cannot close alone, that is the moment to bring in help.

Questions people ask

What should be ready before estimating an AI app?

Know the workflow, target users, data sources, integrations and review requirements. A perfect requirements document is not needed, but clear constraints make the estimate more reliable, and honest gaps are better than guesses.

Is choosing a model the first step in an AI project?

No. Model choice comes after you know the workflow, the data, the risk level, the latency target and the budget for usage. The same workflow may suit different models, and the choice can change later if you keep the design flexible.

How long should AI discovery take?

A first draft of the brief often takes a working session or two with the right people. A structured version, like a 3-day Scope Sprint, adds architecture review and a cost and timeline range. Longer than that usually means the scope is too wide.

Do I need clean data before I start?

You need to know what state the data is in. Messy data is workable if you plan for preparation. Data you cannot access, or are not allowed to use, is a blocker, so check that first.

What is the difference between an AI assistant and an AI agent in a brief?

An assistant drafts or answers and a person acts. An agent takes steps in other systems on its own within limits. Say which one you want for each step, because agents need permissions, approvals and more testing.

Can I skip discovery and just build a prototype?

You can, and a prototype is useful for learning. But a prototype built without data, review and integration answers often has to be redone for production. A short discovery keeps the prototype pointed at the real problem.

How do I know if my idea is a good fit for AI?

Good fits involve varied input, judgment, and a clear outcome that people can check. If the steps never change, regular automation is simpler. If no one can say what a good result looks like, define that first.

Sources

Sources and further reading

The standards, rules, research and official documentation this page draws on.