How to Scope an AI Project Before You Build

Why AI projects need a tighter scope than other software

Conventional software either does what the specification says or it has a bug. AI output is probabilistic: the same agent can route nine tickets correctly and send the tenth to the wrong team. A scope therefore has to define more than features. It has to set an acceptable error rate, say what happens when the model is unsure, and name who catches the mistakes. Tepia treats those three answers as the core of every AI scope.

The second difference is data. A conventional app can be built before its data exists, but an AI feature cannot be judged until it runs on your real tickets, contracts or call notes. When a request reaches Tepia as a broad wish for an AI agent, the first job is to narrow it to one task where the data already exists and the result can be checked.

Good scoping also stops weak ideas early. A week spent learning that your support history is split across three inboxes with no labels saves months of building on the wrong foundation.

Step by step: scoping an AI project or agent

1. Name the task and the owner. Write one sentence, for example: classify inbound service emails by trade and urgency and create a draft job on the dispatch board. A named manager owns the outcome and signs off on the scope.
2. Measure the baseline. Record volume, time per item, error rate and the downstream effect of errors over recent months. Without a baseline, nobody can say whether the AI helped.
3. Classify the work. Decide whether the task is retrieval, extraction, classification, drafting or taking action. Each has a different build pattern and a different way to test it.
4. Check the data. Confirm where past examples live, whether outcomes were recorded, whether personal or regulated data is involved, and who can grant API access.
5. Set the autonomy level. Choose suggest, draft for approval, or act within limits, and list the actions that always need a human, such as refunds or messages to customers.
6. Agree on success. Build a held out test set, set the threshold for launch, and define the fallback when the model is unsure.
7. Map integrations. List every system the AI reads from and writes to, and the permissions it needs in each.
8. Write the scope and pilot plan. Combine the above into one document with a time boxed proof of concept and a go or no go meeting at the end.

Tepia runs these steps in working sessions with the people who do the task today, because the real answers usually live in their heads rather than in a questionnaire.

Questions that expose a weak AI scope

Use these questions with any vendor or internal team. Tepia asks every one of them before agreeing to build.

A scope with good answers to all seven is ready for a proof of concept. Two or more red flags usually mean the first deliverable should be an AI readiness assessment rather than a build, and Tepia will recommend that directly.

Setting autonomy, risk and governance limits

For an AI agent, the scope must list tools as carefully as features. An agent that can read the CRM, create a draft job and send a text message holds three separate permissions, and each should be granted on purpose. Tepia writes a tool list showing read or write access, rate limits and the approval step for every action.

Regulated data changes the scope early. If patient information is involved, the scope names a model provider that will sign a BAA and a logging design that supports HIPAA; if customer records from California or the EU are involved, it covers CCPA and GDPR handling. Tepia uses the NIST AI Risk Management Framework as a checklist for mapping risks, owners and monitoring.

• Log every prompt, retrieved source, tool call and output so a reviewer can reconstruct any decision.
• Cap actions per hour and per record so a loop cannot send a hundred messages.
• Route low confidence cases to a person with the AI’s draft attached.
• Plan a switch that turns the AI off without taking the surrounding app down.

How Tepia approaches AI project scoping

Scoping is the first phase of Tepia’s six phase process, and the scope then shapes every phase after it.

1. Discovery: stakeholder interviews, baseline measurement, a data access check and a sample run on real records. Deliverables are an Investigation Summary, an Interview Summary and User Stories written with acceptance thresholds, such as a dispatcher approving AI drafted jobs.
2. Design: wireframes of the review queue, confidence indicators and override controls, plus the wording users see when the AI declines to answer.
3. Development and Testing: an Alpha proof of concept scored against the held out test set, then a Beta in production with a pilot team, with test plans for accuracy, permission boundaries and load.
4. Training: sessions on reviewing AI output, correcting it and reporting misses so corrections feed the test set.
5. Launch: a phased rollout by team or region, with dashboards tracking accuracy and override rates.
6. Support: monthly evaluation reruns, model upgrades tested before release, and new tasks scoped the same way.

As a rule of thumb, Discovery runs 1 to 3 months, development and QA 2 to 6 months, and launch about a month. Read more about our process or start small with an AI proof of concept.

What a finished AI scope document contains

When scoping is done, you should hold a document a different vendor could build from. Tepia’s AI scopes contain:

• The task statement and its owner.
• Baseline metrics with dates and where they came from.
• A data inventory with access approvals.
• The autonomy level and a tool list with permissions.
• The evaluation set location and the launch threshold.
• Every integration and the direction of data flow.
• Risks, mitigations and the human review plan.
• A pilot plan with dates and go or no go criteria.

If your scope cannot fill every line, that gap is the next task, not a reason to start building anyway. Compare approaches in custom AI vs off the shelf AI tools, read how to build an AI agent for your business, or ask Tepia to run the scoping sessions with your team.

Frequently asked questions