Skip to content

AI · 6 min read

Where to Start with AI in Your Business: A Practical Framework

Most businesses don't need an AI moonshot. They need one well-chosen use case, a fast way to test it and a clear path to production. Here's how to find it.

Abstract illustration of data flowing through an intelligent system
On this page 11 sections
  1. Why most AI pilots stall
  2. Step 1: Map the work, not the technology
  3. Step 2: Score opportunities on four dimensions
  4. Step 3: Choose a pattern that fits
  5. Step 4: Prototype with real data, quickly
  6. Step 5: Engineer for production, not the demo
  7. Step 6: Treat adoption as part of the project
  8. What about governance?
  9. A realistic first 90 days
  10. The bottom line
  11. FAQs

Almost every leadership team we speak with is asking the same question: where should we actually start with AI? The tools are everywhere, the headlines are loud and teams are already experimenting on their own. Yet many organizations still struggle to point to a single AI system that runs reliably inside a real workflow.

The short answer is to start small, specific and measurable. The longer answer is a framework you can apply in a few weeks, not a few quarters. This article walks through it.

Why most AI pilots stall

Before choosing where to start, it helps to understand why so many early AI efforts never reach production. The reasons are rarely about the model itself. They are usually about the problem, the data or the organization around it.

  • The use case was chosen for novelty, not value. An impressive demo that saves nobody meaningful time is hard to justify once the excitement fades.
  • Success was never defined. Without an agreed measure — time saved, turnaround, accuracy, volume handled — nobody can say whether the pilot worked.
  • The data wasn’t ready. The information the system needed was scattered, outdated or locked in formats nobody had prepared.
  • Trust was never earned. Users saw a few wrong answers and quietly stopped using it, because nothing explained how quality was being controlled.
  • Nobody owned adoption. The technology worked, but no one changed the process around it.

A good starting framework is designed to avoid each of these failure points deliberately.

Step 1: Map the work, not the technology

Start with how work actually gets done. Pick two or three teams and list the recurring tasks that consume their time. Be concrete: “reading incoming emails and deciding who should handle them,” “pulling figures from supplier invoices into the accounting system,” “answering the same questions about our policies,” “drafting first versions of proposals.”

For each task, note roughly how often it happens, how long it takes, how many people do it and what happens when it goes wrong. You are looking for work that is frequent, repetitive and text- or document-heavy — the kind of work where current AI systems are genuinely strong.

Step 2: Score opportunities on four dimensions

Once you have a list, score each candidate from one to five on four dimensions:

DimensionQuestion to ask
ValueIf this worked well, how much time, cost or customer friction would it remove?
FeasibilityIs this the kind of task AI handles well today — reading, summarizing, classifying, extracting, drafting?
Data readinessIs the information the system needs available, accessible and reasonably current?
RiskWhat happens if the output is wrong? Can a person review it before it matters?

The best first use cases score well on value, feasibility and data readiness, and carry manageable risk. They are rarely the most exciting ideas on the list. That is fine: the goal of a first project is to deliver real value and to build the organizational muscle for the next one.

Step 3: Choose a pattern that fits

Most practical business AI falls into a handful of patterns. Recognizing which one your use case belongs to makes the design decisions much clearer.

Knowledge assistants

Staff ask questions in plain language and get answers drawn from your own policies, manuals, contracts or past work. The most reliable versions use retrieval-augmented generation (RAG): the system first finds the relevant passages in your documents, then uses a language model to write an answer grounded in them — ideally with links back to the sources so people can verify.

Document processing

The system reads invoices, applications, forms or emails, extracts the fields you care about and passes structured data into another system. People review exceptions rather than re-keying everything.

Triage and routing

Incoming requests are classified, summarized and sent to the right person or queue with the key details already pulled out. This is often one of the quickest ways to shorten response times.

Drafting and summarization

First drafts of proposals, reports, product descriptions or meeting notes, produced from templates and structured inputs, then reviewed and finalized by an expert.

Agents

Systems that plan and carry out multi-step tasks across several tools. Agents are powerful but also the most complex to make reliable, so they usually work best once an organization has experience with simpler patterns — and with human approval at the steps that matter.

Step 4: Prototype with real data, quickly

A prototype’s job is to answer one question: is this good enough to be worth building properly? Build the simplest version that uses real examples from your business, and put it in front of the people who do the work today.

Alongside the prototype, create a small evaluation set: a few dozen real inputs with the answers or outputs an expert would expect. This turns “it seems pretty good” into something you can measure and track as you improve the system.

Keep the prototype phase short. If the results aren’t promising within a few weeks, it’s usually better to move to the next candidate than to keep tuning.

Step 5: Engineer for production, not the demo

The gap between a working prototype and a system people rely on is where most of the engineering effort goes. Production AI needs:

  • Integration with the tools people already use, so the AI step fits into the workflow instead of becoming another tab.
  • Guardrails that define what the system should and shouldn’t do, and how it behaves when it isn’t confident.
  • Human review at the points where errors would be costly.
  • Access control and data handling appropriate to the information involved, including choosing model providers and settings that match your obligations.
  • Monitoring of usage, quality and cost, so problems are caught early and improvements can be measured.

Step 6: Treat adoption as part of the project

An AI system that people don’t use has no value. Involve the future users from the start, explain what the system does and doesn’t do, show them how quality is being checked and make it easy to flag problems. Measure adoption alongside quality, and adjust the process around the tool — not just the tool itself.

What about governance?

You don’t need a large governance program before your first use case, but you do need a few basics in place:

  • A short acceptable-use policy for AI tools, including what information must not be shared with them.
  • A named owner for each AI system in production.
  • A record of what each system does, which data it uses and where humans review its output.
  • A simple process for reporting and fixing issues.

These basics grow naturally as you add more use cases, and they make every subsequent project faster to approve.

A realistic first 90 days

For many organizations, a sensible first quarter looks something like this:

  1. Weeks 1–3: map workflows in two or three teams, score opportunities and choose one use case with a clear success measure.
  2. Weeks 4–6: build a prototype with real data and an evaluation set; test it with the people who do the work.
  3. Weeks 7–12: if the prototype earns it, engineer the production version with integration, guardrails, monitoring and training — then launch to a first group of users.

By the end, you should have one AI system delivering measurable value, a repeatable way to evaluate the next idea and a team that understands what good looks like.

The bottom line

The best place to start with AI is rarely the most ambitious idea on the whiteboard. It is the most valuable, feasible and measurable problem you can solve in a few weeks — built properly, adopted by the people it serves and used as the foundation for everything that follows.

If you’d like help identifying and prioritizing your first AI use cases, explore our AI Transformation approach or our AI & Automation capabilities — or start a conversation.

FAQ

Questions, answered.

What is the best first AI use case for a business?

The best first AI use case is usually a frequent, repetitive, text- or document-heavy task with a clear success measure. Examples include routing incoming emails, extracting invoice data or answering recurring policy questions. It should score well on value, feasibility and data readiness while carrying manageable risk, even if it is not the most exciting idea on the list.

How long does it take to implement AI in a business?

A focused first AI project can often move from idea to a live system in about a quarter. A practical plan spends the first few weeks mapping workflows and choosing one use case, the next few building and testing a prototype with real data, and the remainder engineering a production version with integration, guardrails, monitoring and user training.

Why do so many AI pilot projects fail to reach production?

Most AI pilots stall for organizational reasons rather than technical ones. Common causes include choosing a use case for novelty instead of value, never defining what success looks like, relying on scattered or outdated data, failing to earn user trust around answer quality, and having nobody responsible for changing the process so people actually adopt the new system.

How do you measure whether an AI project is successful?

Measure an AI project against a success metric agreed before work begins, such as time saved, turnaround time, accuracy or volume handled. A small evaluation set of real inputs with expert-approved outputs makes quality measurable as the system improves. After launch, track adoption, quality and cost together, because a system people do not use delivers no value.

What is retrieval-augmented generation (RAG) in business AI?

Retrieval-augmented generation is an approach where the system first finds relevant passages in your own documents, then uses a language model to write an answer grounded in them. It is commonly used for internal knowledge assistants built on policies, manuals or contracts. The most trustworthy versions link back to their sources so staff can verify each answer.

Does a company need perfect data before starting with AI?

No, but the specific information your chosen use case depends on must be available, accessible and reasonably current. Data readiness should be scored alongside value, feasibility and risk when selecting a first project. If the relevant data is scattered or locked in unprepared formats, it is often wiser to pick a different starting use case.

Should businesses start their AI journey with AI agents?

Usually not. Agents that plan and carry out multi-step tasks across several tools are powerful but the most complex pattern to make reliable. Most organizations get better results by starting with simpler patterns such as document processing, triage or drafting, then moving to agents once they have experience and can add human approval at the steps that matter.

What AI governance does a company need before its first project?

A large governance program is not required, but a few basics are. These include a short acceptable-use policy covering what information must not be shared with AI tools, a named owner for each production system, a record of what each system does and which data it uses, and a simple process for reporting and fixing issues.

Have something ambitious in mind?

Tell us where you are and where you want to be. We'll come back with honest thinking on how to get there.