Skip to main content
Discuss Book a call

Free decision tool · quick to answer

Build vs Buy vs AI Agent Decision Tool

You're deciding whether a piece of work should be built, bought or given to an AI agent. Answer about one task at a time, as it runs today, and change how much each question counts if your case calls for it.

Built by XONIK. Your answers stay in your browser.

Line drawing for a build vs buy vs AI agent decision tool: a decision path for one task, branching to build, buy and agent, with one branch ending in a dead end.

In brief

This build vs buy vs AI agent decision tool helps you choose whether a task should be built, bought, handed to an agent, or left alone for now. You answer questions about how the task varies, what a wrong result costs, who owns it and what it touches. It ranks the routes and shows what would change the result.

Step 1 of 6: Cost of a wrong result

0 of 12 answered

  1. Cost of a wrong result (current step)
  2. Shape of the work (not yet reached)
  3. Ownership and reach (not yet reached)
  4. Change and fallback (not yet reached)
  5. Scrutiny and rules (not yet reached)
  6. Running it and buying it (not yet reached)

Cost of a wrong result

1. What does it cost you when this task goes wrong?
How much this matters
2. If a result were wrong, how soon would someone see it?
How much this matters

Shape of the work

3. How much does this task differ from one case to the next?
How much this matters
4. Can someone who does this work today say what the right result is for a given case?
How much this matters

Ownership and reach

5. Who answers for the result, and can that person change its behavior, trace a single answer back and stop the work?
How much this matters
6. What must it reach in your systems and data?
How much this matters

Change and fallback

7. How often do the rules, products or inputs of this work change?
How much this matters
8. Could you stop it and carry on by hand the same day?
How much this matters

Scrutiny and rules

9. Who may ask how a particular answer was reached?
How much this matters
10. Which rules apply to this work?
How much this matters

Running it and buying it

11. Who will run it after launch?
How much this matters
12. Is there already a product that covers the bulk of this work on terms you'd accept?
How much this matters

Answer the 2 questions left on this step to go on.

On this page

What decision does a build vs buy vs AI agent tool frame, and what does it leave to you?

Take one task and ask how it should get done: written into your own systems, bought as a product, given to an agent, or left alone for now. The tool answers from what you tell it and shows its reasoning, so you can argue with the answer in a meeting.

It doesn't choose a vendor, a model or a budget, and it can't say whether a law or regulator treats your use as high-risk. Work on one task at a time, because a result for one task says nothing about the task beside it.

The resultWhat you take onWhat you trade away
Build into your own systemsRules and checks you can read, test and change, and an engineer who keeps them workingSpeed at the start, and the upkeep a vendor would have done
Buy a productA vendor's way of working, kept current without your effort, plus a named person who owns the configurationControl over how it behaves and when it changes
Deploy an agentSoftware that picks its own steps within a written remit, with a person reviewing what mattersPredictability, since behavior shifts when the model underneath changes
Do not automate this yetSettling ownership, the right result and the rules first, while people keep doing the workThe saving you were hoping for, for the time being

Who answers, and how long does the sitting take?

Answer for the task as it runs today, not as the sponsor wishes it ran. The questions touch operations, risk, technology and the team that does the work, so one person rarely knows every answer. Sit down with the task owner and a technology lead; it takes about as long as a working meeting. The questions on rules and on who runs it afterward are where people stop to find out.

When nobody has checked, choose the answer that says so. It's the accurate one, and the tool treats it as information. You can mark any question as mattering less or more. Do that only where you have a reason, and tell the room you did.

How should you read the ranking and the list of what would change it?

The result ranks every route. The gap between first and second matters more than the name at the top: a narrow lead means the task sits near a boundary. Routes level on points are all named.

Below the ranking are the answers that gave the leading route the bulk of its points. Challenge those first, since they carry the result. Next come the single answers that, changed alone, would hand the lead to another route. If one of those is a guess, find out before you decide. Read the row for not automating yet even when it doesn't lead.

How to take the result into the room

  1. 01

    Show the ranking before the reasons

    People react to the result first, and only then read the reasons.

  2. 02

    Test the answers that carried the lead

    The task owner confirms each one from what actually happens, not from how the work is described.

  3. 03

    Settle the answers that would flip it

    Give each uncertain answer to someone, with a date, and run the tool again when they report back.

  4. 04

    Write the conditions before the route

    The owner, the person who runs it, the way to stop it and the rules question get written down and signed by the sponsor.

When does a build vs buy decision tool fit, and when doesn't it?

It fits when

  • You've picked one task and need to decide how it gets done
  • Someone is proposing an agent or a purchase and the case is untested

Look elsewhere when

  • You're comparing named vendors, which the vendor evaluation matrix does
  • You have candidate tasks to rank; use the prioritization tool first
  • The question is whether the money works; use the business case calculator

The tool trades detail for speed. It shows which route your answers favor and what to check next. Only a trial on your own cases can show whether a particular product or build will work.

What do CTOs and product leaders ask before they use it?

Will it tell me which vendor or model to pick?

No. It doesn't choose a vendor, a model or a budget, and it can't say whether a law treats your use as high-risk, which is a question for counsel. It takes one task at a time and ranks build, buy, agent or wait from your answers.

When would it tell us not to automate a task yet?

When a gap would follow you into every route: nobody answers for the result, nobody can say what a right result looks like, a wrong result is costly and found late, or nobody has checked which rules apply. People keep doing the work while you settle it.

It ranks one route first. Is that my decision?

No, it's a first position to argue with. Points add up but can't enforce conditions, so a route can lead while no named person answers for the result or nobody is set to run it. Treat those conditions as gates before the build starts.

Can I change how much each question counts?

Yes. Mark any question as mattering less, normal or more, but only where you have a reason, and tell the room you did. The result also lists single answers that, changed alone, would hand the lead to another route, so check any of those that are guesses.

Want a second view on the result?

Send us the task and the route the tool favored. A senior person reads it and tells you what we would test before you commit.