Ask any seasoned chef what they do on the first day in a new kitchen. They sharpen the knives they already own and walk the line. They do not stop to forge a new chef's knife from raw steel, no matter how interesting that project might be. The whole point of being good at the craft is knowing which tools are worth your time and which ones you should just pick up at the shop on the way in.
I keep thinking about that chef whenever I sit across from a leadership team that has decided to build their own ticketing system, their own CRM layer, their own analytics warehouse, and their own knowledge base, all in the same fiscal year, because they have access to a few large language models and a roomful of well-meaning engineers. They are forging knives instead of cooking dinner.
The risk in this moment is not that AI takes your job. The risk is that AI quietly redirects the attention of your best people toward problems that have already been solved, while the problems only your company can solve sit untouched on the side of the desk.
I hear the receipts every week. A COO who killed a working HRIS so the team could prototype an in-house version. Twelve months in, payroll is still being reconciled by hand on Friday afternoons. A VP of marketing who watched her team write a homegrown email platform on top of an open model, then quietly resubscribe to the tool she had cancelled because the deliverability never recovered. A founder I met last month who told me, with real pride, that his engineers had built their own help desk over a weekend. He stopped smiling when I asked who answered tickets at three in the morning.
None of these decisions look reckless on the day they are made. They look creative. They look thrifty. They look like the company is finally getting its money's worth out of all this AI talk. The cost shows up later, in dribs and drabs, on line items nobody is tracking.
Why so many smart teams fall into this
Let me be clear about what I actually believe. We are living through a generational shift in what software can do. Models can summarize, classify, draft, plan, and reason across domains that used to require a whole team. That is not hype. It is just true. Which is exactly why I get protective of how teams spend their AI budget.
When I push back on an internal AI build, it is not because I want the company to move slower. It is because I have watched too many teams aim a precision instrument at a problem a basic off-the-shelf product would have solved on a Tuesday. The model is doing fine. The use case is the mistake.
Here is what makes this trap so seductive. For the last twenty years, the build-versus-buy debate had a pretty stable answer. If a vendor existed and the price was reasonable, you bought. Building was for the rare case where your needs were truly unique. Modern AI tooling has changed the math just enough that almost every internal stakeholder can now produce a working prototype in an afternoon. The demo lands. The CFO sees a recurring subscription she can stop paying. The engineering team feels seen. Everybody in the room nods.
What that room is not pricing in is everything that comes attached to a real software product. A roadmap. A 24/7 on-call rotation. Two decades of compliance work. A long list of integrations that took years of partnership conversations to land. A support team that has already answered the question your CEO is about to ask. When you cancel the subscription, you are not just shedding a cost. You are taking on a second business, with no headcount and no plan, and pretending that is a savings.
Clayton Christensen wrote about this pattern for forty years. The thing that takes great companies down is almost never a scrappier competitor. It is the slow drift of leadership attention toward whatever feels modern, away from whatever the customer is actually paying for.
A better question than build or buy
The honest framing is not really build versus buy. It is essential versus optional. Where is your company actually winning, and where are you just keeping the lights on? AI deserves a starring role in the first column. In the second column, it almost always deserves to be a quiet, boring, off-the-shelf decision.
The other thing I notice is that the AI conversation tends to get loudest in companies that are quietly avoiding a harder one. A struggling product. A flat sales motion. A leadership team that has not had a real strategy debate in six quarters. Reorganizing the tech stack feels like progress, and it buys everyone another month or two of looking busy. I have caught myself doing this at Futuri, where we sell software to thousands of brands, and the temptation to chase the shiny build is always at the door.
A movie set looks like a real town from the front. Your customers eventually drive around the block.
"Before you spin up another internal AI project, ask whether you are using this tool to grow the business or to avoid a conversation about the business."
There is a clean test for this. If the workflow you are about to rebuild lives close to your unique judgment, your unique customer relationships, or your unique data, lean in. Experiment hard. Build, buy, or wire together whatever combination makes the people inside your company faster and sharper at the thing only they can do. If the workflow is generic, regulated, or held together by a hundred small integrations another company has already wrestled to the ground, hand it to them and move on.
The token bill nobody warned you about
Even if the engineering effort were free, internal AI has a cost structure most leadership teams have never had to manage before. A traditional SaaS contract is a fixed line on the budget. You sign for the year, you forecast cleanly, you move on. Generative systems do not work that way. Every prompt, every retrieval step, every chained reasoning call shows up as variable spend.
The demo always looks cheap. One question, one polished response, a few cents of compute. Production is a different animal. A single end-user request often fans out into six, ten, sometimes twenty invisible calls to retrieve context, reshape it, validate it, and re-rank it before anything appears on screen. Multiply that by daily active users and a couple of long-running background jobs and you are no longer running a tool. You are running a meter.
I am watching well-known enterprises blow through twelve months of approved AI spend in five. Some of them are seeing internal token bills that have started to approach the loaded cost of the headcount that those tools were supposed to make more productive. The CFO who proudly cut a subscription line is now staring at a usage-based line that is harder to predict and much harder to cap.
When you stand up your own AI system, you are not just signing up to write the code and keep it alive. You are personally underwriting the model providers' future pricing decisions. When you buy a platform whose entire business model is absorbing that volatility on your behalf, you have outsourced a risk you are not currently equipped to manage.
Four categories where building it yourself is almost always a mistake
If you want a shortcut for this decision, the next time someone in your leadership meeting suggests a new internal AI build, check whether the work sits in any of these four buckets. If it does, the right answer is usually to pay the vendor and keep your team's time for the work nobody else can do.
Project and operational coordination
Tracking owners, due dates, statuses, and dependencies across a real organization is a deceptively deep problem. The platforms that already do this well have been quietly working through every permission, every notification rule, and every awkward integration edge case for fifteen years. You are not going to leapfrog that with a Notion-and-LLM weekend hackathon, no matter how clean the demo feels.
Financial operations and reporting
The numbers your finance team produces have to stand up to auditors, regulators, and a board. Accounting platforms bring along multi-jurisdiction tax handling, revenue recognition standards that move every year, and audit trails written for the people who will eventually come asking. This is not a category where being clever is rewarded. It is a category where being correct is the whole job.
Compliance-sensitive data infrastructure
If you operate in healthcare, financial services, education, or anywhere personally identifiable information travels through your system, the data layer is a legal artifact, not just a technical one. The vendors who specialize here have certifications, indemnities, and a legal bench that has lived through your exact failure mode. Rebuilding that internally to save a subscription is not a cost-cutting move. It is a liability transfer in the wrong direction.
Sales intelligence and data enablement
Wrapping a model around a template to draft outbound email is a parlor trick. Building a system that ingests dozens of live data feeds, resolves identities across them, scores accounts against real intent, attributes pipeline movement back to specific actions, and writes proposal-grade content that reflects your point of view is a real company. If you try to staff that inside an org whose main business is something else, you have just opened a second startup with no funding round, no roadmap, and no product manager.
Where AI actually pays you back
The deployments I have watched return real money do not replace existing systems. They sit alongside the people who already know what they are doing and lift the ceiling of what those people can accomplish in a day.
A boutique environmental law firm I have followed has built nothing of their own. They use commercial tools to surface every relevant EPA enforcement action, every related ruling, every footnote in opposing counsel's history with a particular judge, and they hand the synthesized brief to their lead partner the morning of a deposition. A junior associate used to spend two weeks doing that work. The partner is still the one in the room making the call. The model is just sharpening the question she walks in with.
I see the same shape in revenue teams. The point is not to swap out the CRM or write a homegrown sequencer. The point is to sit right next to the seller's actual edge, which is a deep read of the customer, a feel for the moment, and the ability to show up to a meeting two questions ahead of the buyer. AI that strengthens those instincts wins. AI that tries to perform them loses.
That is the deployment style I keep betting on. Not software that pretends to be your company. Software that gives the actual humans inside your company a sharper version of the day they were already going to have.