Generating ideas isn’t what’s preventing you from getting ROI from your AI tools. Someone in every department has a suggestion, the Slack thread has 140 messages in it, and the doc where you collected everything went out of date a month ago.
The challenge is deciding how to prioritize your AI ideas and which deserve investment.
"AI has introduced a whole new world of options when it comes to solving problems in organizations," says Clarinet Cofounder and Head of Product, Diane Sadowski-Joseph. "But it's also introduced a whole new set of problems when it comes to solving problems in organizations."
Part of the AI enablement work we do at Clarinet is helping our clients get out of the decision-making spiral of deciding how to resource their AI ideas.
We’ve determined these five paths as the most common resourcing scenarios:
- Build - Create it from scratch, in-house or with a partner.
- Buy - Pay a vendor who already solved it.
- Tailor - Buy something close, then add the piece that makes it fit.
- DIY - Create a prototype version yourself to see whether the idea holds.
- Skip - Do nothing new. Use what you already pay for, or wait for the feature to arrive.
The right path depends on the AI idea you’re evaluating: how central it is to the business, what breaks if it fails, whether your data is ready, and who maintains it after launch.
Five ways to resource an AI idea
Let's look at these paths in more detail using a fictional example:
Acme is a B2B SaaS company with 480 employees, and their CEO just declared 2026 their “AI-native year.” Their Chief of Staff, Alex, has been chosen to lead a new AI steering committee comprised of four people with incompatible agendas:
- Priya the CTO: "I have three platform engineers and a roadmap. Don't ask me to spare them for AI theater.”
- Marcus the Sales Leader: "Have you seen what our competitor just shipped? We'll be left behind by Q2."
- Jordan the Head of People: "I'm interested, but I'm also drowning.”
- Devin the Engineer: “I built a Slack bot over the weekend and want to ship it.”
Alex and the committee have 90 days to show measurable progress from their AI initiatives. In their first meeting, they generated 17 ideas. Through a prioritization process they end up focusing on performance reviews because the pain is high and recurring, it touches everyone at once, and making it easier is a visible first win the whole company will feel.
| Option | Best when | Watch out for |
|---|---|---|
| Build | The work is central to how your organization operates, the data is proprietary, or the workflow is specific to you | Maintenance, which outlasts and often outweighs the build |
| Buy | A proven vendor category exists, speed matters, and customization needs are low | Paying for a category you have not scoped, and redundancy with tools you own |
| Tailor | A tool you own covers most of it and the missing part is where the value sits | Automating a broken process instead of fixing it first |
| DIY | You are learning what is possible, pressure-testing a vendor, or the scope is small | Prototypes becoming production with no owner |
| Skip | A tool you already pay for gets you part of the way, or the problem improves on its own | Calling it a decision when it is avoidance |
Build
Acme's option for performance reviews: Devin the Engineer's prototype. Drafts reviews from Slack + 360 feedback survey data.
When it fits:
- The work is core to how your organization operates, the data is proprietary, no vendor has modeled your workflow, or you want the internal capability that comes from having built the thing yourself.
When it does not fit:
- When nobody has answered the maintenance question. "Maintenance is now forever," Diane says. "Once you build a custom suit, you can't return it. That sleeve goes wonky, that's on you."
AI builds need more upkeep than conventional software, because the models underneath change on someone else's schedule. Watch for hidden complexity too. Nick, Clarinet's Head of Build, points to the complex logic tree behind Slack's notification system, a full screen of branching decisions behind whether to send one alert about one message: "It's a good example of ways that the complexity can be obfuscated from a really well-built piece of software," he says.
Buy
Acme's option for performance reviews: Lattice AI Review Assist is priced per employee and already integrated.
When it fits:
- A proven vendor category beats what you would produce, speed to value matters, or you would rather not own the maintenance. AI note-takers are a good example. You could build one, and it would lose to a vendor who has spent three years on it.
When it does not fit:
- When you’re buying to avoid defining the problem. Vendors are good at demos, so ask for a free week of real use instead of a walkthrough, and trial two or three at once while switching costs are low.
Tailor
Acme's option for performance reviews: Lattice + a custom Claude calibration layer that ingests transcripts.
When it fits:
- The gap is narrow and specific, and adoption matters. Your team knows the interface, so you extend a habit instead of asking for a new one.
When it does not fit:
- When you cannot find a good tailor. Whoever does the work has to understand the business need and the engineering, and that combination is rare.
Diane describes a potential failure: "An engineer builds exactly what you asked for, but it turns out that that was an extremely inefficient thing to build." The process gets automated in its current shape and the inefficiencies compound.
DIY
Acme's option for performance reviews: A two-week Zapier + Claude prototype. See if the workflow holds.
When it fits:
- You are learning, or negotiating. A rough version you built yourself is the fastest way to tell whether a vendor's promise is ordinary or remarkable. Set a time budget and add at least 40% to it, which is the gap between how long people expect a build to take and how long it takes.
When it does not fit:
- As a long-term solution, unless the scope stays small. Diane's own example is instructive. She tried to build her own scheduling tool, what she got instead was an endless loop of calendar invites.
Skip
Acme's option for performance reviews: Rippling's new AI summary feature. Free. Already in their stack.
When it fits:
- "A lot of times there is a 30% solution that's just sitting right there," Diane says. There is a budget case too. So much AI tooling overlaps that when a team asks for something new, a fair response is: our stack already does that, show me why it is not enough.
When it does not fit:
- When the problem worsens on its own. Diane borrows a distinction from a former colleague: wine problems soften with time, while "a banana problem is a problem that if you leave it alone, it just gets worse and worse until it rots and attracts fruit flies." Skip is cheap for wine problems and expensive for banana problems. Either way, put a date on the revisit.
When you need to revisit an AI investment decision
"There is no perfect in the world of AI," Diane says. Part of the reason is that AI’s capabilities are changing quickly. So however you choose to resource your AI idea, don’t be surprised if it evolves while you’re deciding. Pricing changes. A vendor ships the thing you were building. A prototype turns out harder than the engineer thought.
None of that means the decision was wrong.
Diane’s advice is to frame it as, "This is the decision we're going to make today, knowing we may need to evolve it tomorrow." AI adoption is a practice rather than a project, and those who get ahead are the ones refining their ability to make the next decision well.
