Skip to main content
Published in Product Insights

5 Questions to Ask Before Investing in AI

5 Questions to Ask Before Investing in AI
Table of contents
  1. How do you test whether customers will pay for an AI feature?
  2. What will AI cost to run once everyone is using it?
  3. Is your data ready for AI?
  4. How will you know the AI investment paid off?
  5. Looking to the future  

Sharks with laser beams on their heads :dr_evil_face:

Just because it's possible (and kinda sounds cool) doesn't mean you should do it. That was true when the Internet of Things was mainstream (looking at you connected toaster oven), when blockchain was mainstream (ICO scam after ICO scam), and is true now with AI.

There are lots of ways in which AI can help solve current problems and make us more efficient: tools like Grammarly or Copilot for writing come to mind. But let's also think critically about when, where, and how we're implementing it in our products so we don't waste everyone's time and money.

Or if we're going to waste time and money. Let's make sure we have a lot of it and are having fun doing it (have you seen some of these AI movie trailers?!).

Trying to figure out if your project is sharks with laser beams, or actually useful? We have a list of five questions to ask before diving in.

1. Is there an existing AI product that suits your needs?#

One of the most challenging parts of implementing AI is if you're trying to develop something completely new. As it stands right now, there are a handful of things that AI tends to do really well:  generating text, summarizing content, etc.

If an existing AI product already does what you're looking for, the build is way less complicated. There's still the process of integrating the AI tool with your product. But, it's fairly straightforward.

That said, if you're trying for something totally new, things get a lot more complicated. There are lots of metaphorical "how do we attach the lasers to the sharks?" type questions. You can always try the duct tape route. But don't be surprised if your evil lair catches some stray laser beam shots along the way.  

2. What's your budget?#

If you're planning on investing in AI it's best you have a healthy budget (struggling very hard not to make another Austin Powers reference…) Even in the best-case-scenario where you can connect to an existing LLM (this stands for large language model and is essentially the AI's "brain") there are upfront costs like connecting it to your system and designing then building a user interface. Not to mention the ongoing costs like paying for tokens to access an existing LLM, and recurring maintenance.  

Depending on your project, it's possible you'll have to augment or train an existing LLM extensively, which takes a lot of time and resources. Or, in the most extreme cases, it's possible you'd need to build your own LLM. Which would likely cost in the millions, on the low end. For reference, OpenAI reportedly spent over $100 million developing their LLM. How many AI movie trailers could that have been?

Understanding your own goals with AI will inform the budget. If you're truly trying to create a new AI application, go all in with a multi-million dollar plan. If you're feeling pressure from a board or senior leaders to make progress with AI, start to explore what an LLM integration might look like and share the associated cost with leadership.

3. Do you have someone on staff, or retainer, with technical ability?#

AI is evolving at a very fast pace. That's exciting for potential new capabilities, but it also means that frequent updates and changes to AI products will likely be needed. If you don't have someone on staff, or retainer, to handle those changes your AI investment may never have the chance to succeed.

That type of shepherding isn't just limited to the evolving nature of AI, it's also crucial to monitor the quality and output of the AI's work internally. It's kind of like having a five-year-old make you breakfast. Everything may look right, but one bite in you find out peanut butter and glitter isn't a stellar pairing.

Similarly, AI might generate good looking code, but it could be buggy or downright wrong. So, it's crucial that someone checks the work. Otherwise, you might be in for a rough time.

4. Is there a non-AI solution that could be better suited for your use case?#

There are lots of things that AI can do, but that doesn't mean it's always the best or most cost-effective option. Recently, a real estate company asked for our help creating a tool that would use AI to alter existing property photos to show improved landscaping, lighting, etc. We declined.

It wasn't a bad idea. And AI is fairly good at generating images. But it's not good at determining what additions to an image would make it "better." It's always important to remember: AI can't actually think. It may appear to think (a la, walking Weekend at Bernie's style), but what it's actually doing is making highly educated guesses.  

In the end, it was actually more cost effective for the potential client to outsource the editing work instead of trying to build the tool. AI isn't a magic wand and it's important to remember just because it's a solution, doesn't mean it's the only solution.

5. Will AI make your offering better, or just different?#

AI is an investment that takes time and resources. It's a business decision. And just like any other investment, you should have a clear understanding of the benefits and risks it brings.

It's easy to get caught up in the hype and trying to keep up with the virtual Joneses, but if AI won't bring real, measurable value it's not worth the time. For now, we see a lot of products simply being made different by AI, not actually better or more efficient.

Map out expectations for your AI investment, and make sure it's one worth making.  

How do you test whether customers will pay for an AI feature?#

Ask customers for money before you build the AI feature. Show the offer with a real price, deliver the result by hand for a few customers, then charge for a pilot. Praise in a demo proves interest. A payment at your target price proves demand.

Question 5 asks whether AI will make your offering better or just different. Run these three tests in order, cheapest first, and stop at the first one that fails.

  1. Show the offer before the feature exists. Describe what the feature will do, show its real price, and count who asks to buy. This proves interest only.
  2. Deliver the result by hand. Give a few customers a simple screen, and have a person produce the output behind it. This gives early insight into whether the output is useful.
  3. Ask for money. Sell a paid pilot or a paid order at the price you plan to charge. This proves people will pay.

Nielsen Norman Group calls the second test the Wizard of Oz method. The user works with an interface "that appears to be autonomous but is (fully or partially) controlled by a human." The same article says the method "lowers investment risk into complex, costly technologies (such as generative AI)."

Release the build budget once buyers pay at your target price. Then automate the part they keep paying for.

Before you build, also compare the cost of the build with the cost of doing the work by hand. The real estate story in question 4 ends this way: "In the end, it was actually more cost effective for the potential client to outsource the editing work instead of trying to build the tool."

What will AI cost to run once everyone is using it?#

Multiply three numbers: requests a month, tokens per request, and the price per token. Input tokens and output tokens have separate prices, so price each and add the two. Do the math at full use. In the example below, a feature that costs $26 a month in a pilot costs $520 on Claude Sonnet 5.5 once everyone uses it.

Question 2 lists "paying for tokens" as an ongoing cost. Anthropic's Claude API pricing page was checked on November 14, 2026. It lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens. Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens on prompts up to 100,000 tokens. The Batch API processes requests asynchronously, with "a 50 percent discount on both input and output tokens."

This example assumes each request sends 4,000 input tokens, the question plus the records it reads, and gets 500 output tokens back.

ScenarioRequests a monthTokens a month (in / out)Monthly model bill
Pilot, Claude Sonnet 5.52,0008 million / 1 million$26
Full use, Claude Sonnet 5.540,000160 million / 20 million$520
Full use, Sonnet 5.5 through the Batch API, for work that can wait40,000160 million / 20 million$260
Full use, Claude Haiku 5.540,000160 million / 20 million$26

The pricing page describes Haiku 5.5 as "For high-volume, latency-sensitive tasks such as classification, extraction, and routing." That fits sorting and extracting work.

The model bill is also only one line of the budget. Question 2's other costs still apply: connecting the model to your system, building the screens and maintenance.

Some gains need no model call at all. In The Gnar Company's work for Counter Culture Coffee, one finding on the month-end close reads: "Many of the gains come from NetSuite capabilities Counter Culture already owns, with AI reserved for reconciliations and accrual worksheets."

Is your data ready for AI?#

Your data is ready when an AI tool can read it without a person explaining it. That means one home for each key list, field names everyone reads the same way, and access a system can use without manual exports. If only one person knows which field to trust, the AI will not know either.

The Gnar Company's work for Counter Culture Coffee states the reason: "An assistant is only as reliable as the data it reads, and it has no way to know which of two similar fields is the right one."

Run these five checks in your next leadership meeting.

  • Does each key list have one home? A yes means customers, products and prices are each edited in one system, and every other system reads from there.
  • Do teams agree on what each field means? A yes means no one has to ask which field to trust.
  • Can a system read the data directly? A yes means software can reach it without someone running a manual export.
  • Are you allowed to use it this way? A yes means your contracts and policies permit the use. It also means only people who can see the data today can see it through the AI.
  • Does someone own fixing bad records? A yes means a named person corrects errors at the source.

Each no is a data problem to fix before the AI reads that data.

For a structured version, Gnar's AI Readiness Assessment asks 10 questions. A fuller review checks six areas, laid out in what an AI readiness assessment should cover.

How will you know the AI investment paid off?#

Decide how you will judge the investment before you spend any money. Write down today's number, the one measure AI should move, the result that means stop, and the date you will check. Compare the change in that measure, in dollars, with what the AI costs to build and run. The cost section above covers the model bill, and the build cost is the other number to add.

Question 5 says to map out expectations. Write these four lines before money moves.

  1. The baseline. Measure today's number now.
  2. One measure. Pick the single number the investment should move.
  3. The stop line. Write down the result that means you stop or change course.
  4. The check date. Put the review on the calendar before the work starts.

Here is a hypothetical example with round numbers. Say 20 of every 100 supplier invoices need a person to fix a mismatch today. That share is the measure, and its goal is 10 of 100. The stop line is anything above 10 of 100 on the check date, so at or below 10 is a yes and anything above is a no. The check date is the last day of the quarter.

Now turn the change into time. In this example, each fix takes 15 minutes. Cutting fixes from 20 to 10 per 100 invoices saves 150 minutes per 100 invoices. Multiply that by how many hundreds of invoices you handle in a month. Put a dollar value on that monthly time. Then compare it with the monthly model bill and the monthly cost to maintain the tool. Spread the build cost over the months you expect to use the tool.

Measured data can change the plan itself. The Gnar Company built a prioritized AI roadmap for Counter Culture Coffee, a coffee roaster. It found 79 AI and automation opportunities. Then it scored and ranked each one. It also found $1.5M in potential savings across the top 10 initiatives.

The case study shows data changing one idea: "Coffee procurement initially looked like a market-timing opportunity. A full year of contract data showed the team was already managing that well, so the recommendation shifted to what would help most: a single, real-time view of the coffee position across contracts, importers, and warehouses." The case study also says: "The findings were presented to Counter Culture's board."

Ranking and ordering your own ideas is the next step, covered in an AI roadmap the board will fund. Gnar's AI strategy page describes the result: "You walk out with a strategy, a roadmap, and a fully scoped first build."

Looking to the future  #

In some form, AI will probably be around for years to come. But, it's still an evolving technology and one that isn't right for everyone right now. Be honest with yourself about why you want to add AI into your business and know what you want to accomplish with it. Is it to look cool, or actually do something useful? If you don't have a clear understanding, it may be worth exploring other options and setting a future date to revisit the AI conversation.

If you want a fresh set of eyes to help answer these questions, or you're ready for the sharks to meet the lasers, feel free to get in touch with us here. We're happy to help.

Frequently asked questions

What costs remain when using an existing LLM?

You still need to connect it to your system and design and build a user interface. Ongoing costs include tokens and recurring maintenance.

Why do I need technical help after launch?

AI products change quickly and may need frequent updates. Someone also needs to check output because generated code can look good while being buggy or wrong.

Can a non-AI solution be more cost effective?

Yes. For a real estate company that wanted to improve property photos, outsourcing the editing was more cost effective than building an AI tool.

How should I decide whether AI is worth adding?

Define the benefits, risks and expected value. If you cannot say what useful result you want, consider other options and revisit AI later.

What questions should I ask before investing in AI?

Start with the five in this post. Is there an existing AI product that fits, what is your budget, and who will maintain it? Is there a non-AI option, and will AI make your offering better or just different? Then add two more. Is your data ready, and how will you measure whether it paid off?

How do I test if customers will pay for an AI feature?

Ask for money before you build. Show the offer at a real price, then deliver the result by hand for a few customers. Then charge for a paid pilot or a paid order at your target price. Build once buyers pay at your target price.

How much does an AI feature cost to run each month?

Multiply requests a month by tokens per request by the price per token. Input and output tokens are priced separately, and you add the two. In this post's example, each request sends 4,000 tokens and gets 500 back. At that size, 40,000 requests a month cost $520 on Claude Sonnet 5.5 and $26 on Claude Haiku 5.5, a fit for sorting and extracting work. Those are Anthropic's prices on November 14, 2026.

Who can help build an AI roadmap to present to the board?

The Gnar Company built one for Counter Culture Coffee, a coffee roaster. It found 79 AI and automation opportunities. Then it scored and ranked each one. The findings were presented to Counter Culture's board. Gnar's AI strategy page says "You walk out with a strategy, a roadmap, and a fully scoped first build."

Stay In The Loop

Never miss an update. Get the latest blogs, webinars, and expert insights straight to your inbox from The Gnar.