In various parts of my day, be it work or personal, I can find myself in conversations where folks are trying to make sense of the current AI landscape. The questions range from What is the best model? and Which agent or tool should we use? to How do we choose a provider without spending too much? Then there are comments like I’m doing task x, but the model I’m using is too expensive or Agent abc is freaking awesome and does everything I want when I prompt it this way.
These conversations aren’t limited to any one role or level of technical familiarity. Some people are trying to make a particular task easier, some are deciding what is appropriate for a team or business, and others are simply trying to understand what all this might mean for them. They each bring a different starting point, set of constraints, and idea of a useful outcome. Those differences don’t make any of the questions less valid; they make a single, simple answer less likely. Depending on whether I’m being asked to help or simply for my take, I usually come back, to varying degrees of success, with some version of “yes, and…” or “well, it depends…”.
That answer isn’t nearly as satisfying as the ads, articles, and industry professionals promising that if you just do this 1 thing or use this one tool in a way you probably haven’t tried yet, you will be that much more productive or able to get more out of the tool and your career. Sure. Why not?
Why Not
We’ve all likely seen or heard a few of those horror stories around agents wiping drives, companies being held liable discounts for chatbot misinformation, or instances of runaway token costs. The more stories we encounter, the more they erode our trust in these tools and platforms.
NOTE: There are other, more tragic stories around the various implications of the technology I’m not necessarily equipped to represent nor comment on. I’ll focus on the costing and “productivity” areas.
To me, the stories indicate a failing of first principles. There’s a recurring theme around really wanting to get out ahead and “win”. In order to get there first, corners are cut. Sometimes, that’s ok. Other times, the risk implications are catastrophic. There’s nothing wrong with wanting to “win” at what you’re doing, especially when it’s something you’re passionate about. Where it becomes an issue is when it’s blinding you from the decisions you need to make in order to achieve your goals.
AI gives us leverage, but leverage amplifies poor judgement as readily as good judgement.
Do you need to spend USD$1500 in 2 weeks on Claude or ChatGPT? Can you just share or publish that deck right after it was generated to “get it out there”? That post the AI just created, “looks good enough”, so it can be published right?
AI can reduce the work as well as the time required to produce something, but it does not remove the effort required to verify the result and assess its value of the work done. This part is on you. Whether you’re working independently, leading a team, or running a business, you remain responsible for defining the expected outcome and making sure everyone involved understands it. That’s no different with the introduction of AI.
Define what it is that you’re trying to accomplish first; that outcome or set of outcomes you’re looking for. Make sure you have clearly stated what is acceptable as well as unacceptable. I tend to think about this through my world and lens of software engineering. In Scrum and Agile teams, we often express it through Acceptance Criteria and a Definition of Done: what must be true for an outcome to be acceptable, and what standards the work must meet before we call it complete. The language will vary by field and industry, but the principle is the same. Everyone involved needs a shared understanding of what “good” looks like.
From there, use this clear contract to help identify what is the risk and impact. How can we limit the negative impact, typically referred to as “blast radius”. Identify clear boundaries and signals that allow you to introduce “stop” mechanisms. Where appropriate, try to keep the accountability to a person. Just remember that there will always be different motivations that will influence decisions. This is where the previous boundaries and mechanisms step in to help.
The important thing to remember here is that you can’t eliminate failure. You can make failures smaller, easier to detect, and easier to recover from. It’s part of the process.
Why
There are real gains that can be made with AI. You can absolutely save time, effort, and money by spending a bit on the latest tools. Using tools like Microsoft 365 Copilot or Google’s Gemini Notebooks for research, aggregation, and analysis of notes, documents, etc. can reduce the time and effort it can take to pull your ideas together. Bringing in the checks and balances enables experimentation.
The work isn’t just the cost of using AI responsibly. It is how you discover where AI is genuinely useful to you. Now that you’re not having to scour through multiple Google links or hundreds of PDF documents one-by-one, your effort can focus on a different question: Given my constraints, am I getting the expected results?
AI can shift some of our effort away from the repetitive or mundane. It can allow the possibility for us to apply ourselves towards being more the creative, the artist, the craftsperson. A craftsperson does not ask for the “best tool” without considering the job, material, required finish, working conditions, and their own familiarity with it. They learn what a tool does well by using it, observing the results, and understanding its limitations. A more expensive or powerful tool is not automatically the right one.
We all likely have a couple of screwdrivers in our home as well as a power drill. Yes, you can use manual screwdrivers for just about everything, but it’s much easier with a power drill. When building or assembling something, you will save your hands and your time with the power drill, but it can also do a lot of damage when used incorrectly. Bruises, broken bones, stripped screws, and IKEA furniture that doesn’t quite hold up. Knowing how to avoid injury matters, but knowing when to use the drill, which bit to select, what those numbers mean on that clicky ring around where you insert the bit, and when another tool is better is the actual craft.
NOTE: That numbered ring is called a torque ring to adjust the clutch, if you were curious but have never thought to look it up. Lower numbers equal lower torque for smaller, more delicate materials.
All of this to say, the best AI isn’t found by studying rankings or following someone else’s workflow. You find the right tools by putting them to work in your own context, within boundaries that let you learn without betting everything on the result. Experimentation, but with constraints.
Have a notebook (digital or physical). Note down your process. Capture what works and what doesn’t work. Iterate on what does. Revisit what doesn’t. With the pace AI evolves, things that worked today may not tomorrow. Things that sorta worked, might improve. BUT, don’t get stuck in experimentation. Make sure that the outcomes are the focus while the experiment is the side quest.
So What
You might be thinking… “not only do I need to be an expert in my actual craft, I need to be an expert in AI?” The short answer to this is “no”. The longer answer comes back to “being an expert in my craft”. AI is a tool. It may enable you in your craft, it may not. But, you need to figure this out yourself without breaking the bank.
First and foremost, ensure that you can define value. You must be able to measure what good looks like. You may have a clearly defined objective, but if you’re spending USD$100 to make USD$80… There’s a gap. A coding agent producing code faster is not necessarily valuable. If the result requires more review, correction, testing, or recovery, the apparent gain may be lost.
In software engineering, this has been a hot topic for a bit now. How are we justifying the spend? Because the output, i.e. code, can be a few steps removed from the business itself, how do you justify the spend? One area I keep coming back to, and one where I see the industry focusing attention, is the delivery feedback loop. We have a checkpoint where code, whether new or updated, is tested and reviewed relative to the intended outcomes, i.e. the Acceptance Criteria. This applies regardless of whether that code was produced by a human, an agent, or a combination of the two.
Whether the code comes from a seasoned software engineer, a fresh graduate, either working with a coding agent, or an autonomous agent, the output may not be correct the first time around. That mix makes it difficult to isolate the value of the tool from the skill of the person, the complexity of the task, and the surrounding review process. A single request will not tell you much. Over time, patterns across similar work become useful.
At that checkpoint, we look at two dimensions together: the quality of the work and the time it takes to deliver. Signals like churn (how many times a request is closed and reopened) and drag (how many times a reviewer or tester asks for changes) indicate how much correction was required. The time from the initial ask to a “done” state indicates how quickly an acceptable result arrived. Neither tells the whole story on its own. Faster delivery with more rework may not be a gain, while stronger quality that takes longer may or may not justify the spend. Viewed over time, along with the complexity of the work and the experience of the people involved, the combination better indicates whether the tool is improving the overall value of the work.
Your craft may have a different checkpoint, but the question is the same: how much value reached the other side, and what did it actually cost to get there?
But, before you get there, make sure you understand where you are now. How much does it cost you to do x today? Include not only the direct spend, but also the time and effort required to review, correct, and recover the work. Use that as your baseline.
During discovery, decide what you can afford to spend learning, how long the experiment will run, and what result would earn the tool an ongoing place in your workflow. That learning budget may temporarily consume more of the expected savings. Once the tool becomes part of the workflow, its ongoing cost needs to be justified by the value it demonstrates.
Most tools and platforms have tiered subscription models, with each tier containing a predefined bucket of use. Many also offer pay-as-you-go pricing, either on its own or after the included usage runs out. That transition can be easy to miss, and it is where costs can get away from you. Before you start, make sure to know if the service stops at its limit, slows down, or continues under something like metered billing. Set a hard spending cap where one is available, add alerts before you reach it, and review your usage before the bill becomes the signal. A sensible budget constrains what you spend. It doesn’t need to limit where you look.
OpenAI, Anthropic, Google, and Microsoft are just a few that dominate much of the conversation, but visibility is not the same as suitability. Depending on your needs and means, other hosted providers, specialist tools, and open-weight models may offer a better fit for a particular language, region, workload, privacy requirement, or budget. This is a rabbit hole. You likely do not need to explore all of it, but some of it may be worthwhile.
Platforms like Hugging Face provide broad model catalogs, hosted inference, and downloadable open-weight models where their licenses permit it. Services like Replicate let you run and fine-tune models through hosted APIs. If you want to build around those models, LangChain offers open-source frameworks for agent development, while Pi is a minimal, extensible coding-agent harness. These are different paths, each with its own trade-offs.
Some open-weight models can also be self-hosted. Taking that route can offer more control over data, configuration, and usage costs, but that control comes with responsibility and its own set of costs.
Choosing between someone else’s computers to host your work and a self-hosted path is less about finding the “universal best” and more about deciding which set of responsibilities you want to outsource and which you are prepared to own.
Wrapping It Up
None of the tools or platforms mentioned here sponsor this post, and I have no direct affiliation with them. I mention them because I have access to them and have tested or used them for personal and professional purposes. Their inclusion is not an endorsement, a ranking, nor advice to choose one over another. I regularly test and use others that were not mentioned. I also tinker with self-hosting.
Treat my examples and opinions as a starting point for your own evaluation, not a substitute for it. Test the tools that seem relevant to your needs. Compare them against constraints that actually matter to you, including cost, quality, privacy, reliability, and the effort it takes to use them to get the results you’re looking for. What works for me may not work for you. Learning where a tool helps, where it falls short, and whether it is worth the trade-offs is part of the work.
This space is vast, fast moving, and not going to stop any time soon. We may or may not be in a bubble. It may or may not pop soon.
What are the outcomes you’re looking for and the value those outcomes bring you?