Home › The business questions

Getting started

How do we get started?

Updated November 2026 · About 5 minutes to read

The short answer

Start with one question, a small team, about twelve weeks, and a finish line you write down before you begin.

At the end you decide: keep going, fix and try again, or stop. All three are good outcomes if you learned something cheaply.

Think of it as opening a test kitchen

Before a restaurant adds a dish to every branch, it tries it in one kitchen, with a few regulars, and asks them honestly what they think.

One dish, one kitchen, real diners, and a date when you decide.

Where the picture stops working: a recipe is finished once it works. The AI is only as good as the documents behind it, so someone has to keep them current after the test, too.

This guide starts once you know which question to answer. If you haven't picked one yet, start with choosing your first project.

The team: six roles, mostly part-time

You don't need an AI department. You need a few people who already know the work, and enough of their time.

A sponsor

A senior leader who wants the result and can clear obstacles.

Time: an hour a week, and the final decision.

A project owner

One person who runs it day to day and is answerable for the finish line.

Time: about half their week.

Document owners

The people who already look after the handbook or policies the AI will read.

Time: a few hours a week, more at the start.

Expert checkers

Two or three people who know the right answers and mark the AI's.

Time: a few hours a week while it's tested.

A builder

Your IT team or an outside firm, who sets it up and connects it to the documents.

Time: the only near-full-time role, during the build.

Someone for access

Usually IT or security, making sure people only see what they're allowed to.

Time: a few days, early on. More on SecureLLM.ai.

These time estimates are rough rules of thumb, not survey data. Yours will vary with the size of the job.

Twelve weeks, in three stages

An illustration of a typical first project, not a promise. Messier documents add weeks to the first stage.

Weeks 1 to 4
Get ready
Write down the "before" number and the finish line. Collect 50 real questions with known answers. Tidy the documents the project uses.
Weeks 5 to 8
Build and check
Set it up on those documents. The experts mark its answers to the 50 questions, and the documents get fixed where it went wrong.
Weeks 9 to 12
Try it with real people
A small group uses it for their daily work, with a "this was wrong" button. Then compare against the finish line and decide.

Our guide Is our information ready? has a one-afternoon test that tells you how long the first stage will take.

The budget: mostly people's time

In our cost guide, a made-up company of 800 people spends about $95,000 in the first year on an HR help desk: getting documents ready, building it, and checking answers. The AI bill adds under $1,000. The numbers are illustrations, not data.

Most of the document work and the building happen in these first twelve weeks, so expect most of the first year's cost up front. Count your own people's time as a cost too, even if no invoice arrives.

Budget for the people and the documents. The AI is the small line.

What success means: write it down first

Agree on these before anything is built. Afterwards, everyone remembers the goal a little differently.

What's the "before" number?For example, how long an HR question takes to answer today.
How often must it be right?Marked by your experts against the 50 test questions.
Will people use it?How many in the test group, how often, by week twelve.
What does it save?Time per answer or fewer questions to the experts, compared with the "before".
What result means stop?Decide now, so stopping is a decision, not a failure.

Three ways to run it

Your own team

Your IT team builds it, using an AI vendor's model and tools.

Best for: companies with developers who have time, and who want to keep the know-how.

Gives up: speed. A first project teaches your team as it goes, and that shows in the timeline.

A ready-made product

A tool already connected to where your documents live, such as one built into your office software.

Best for: a quick start, when the documents are already in that vendor's tools.

Gives up: control over how it finds and checks answers, and often a per-person price for everyone.

With an outside partner

A firm builds it alongside your people, then hands it over or keeps running it.

Best for: a first project, when nobody inside has done one before.

Gives up: a bigger bill up front, and your people need to learn enough to own it afterwards.

What goes wrong

No finish line. The test never ends. It just fades.
Experts with no time. Nobody checks the answers, so nobody trusts them.
A test group that isn't the real users. The IT team loves it; the people it's for never try it.
Nobody owns it after week twelve. The documents go stale, and so do the answers.
Growing the scope mid-way. "While we're at it, add finance." Finish the first question first.

Questions to ask your team or vendor

Who is the sponsor, and who owns the project day to day?
Whose time do we need, how much, and have their managers agreed?
What's our "before" number, and what result by what date means keep going, fix, or stop?
Who are the real users in the test group?
Who owns it, and the documents behind it, after the test ends?

Other names you may hear

Pilot: A small, time-limited first project with real users and a clear measure of success.
Proof of concept (POC): A smaller test that something can work at all, often on sample documents and without real users.
MVP: Minimum viable product: the simplest version that's useful enough for real people to try.
KPI: Key performance indicator: the one or two numbers you judge the project by.
Sponsor, stakeholder: The sponsor is the senior leader backing the project. Stakeholders are everyone it affects.
Rollout: Making it available to more people after the pilot, usually in stages.