The AI market is loud in a way that makes people think there are hundreds of decisions to make. There are three.
You can buy something finished. You can take something that mostly exists and bend it around your process. Or you can build. Everything else — which vendor, which model, which integration — is a detail underneath one of those three. And I've watched enough teams pick the wrong one of the three to know it costs more than picking the wrong vendor inside the right one.
Buy
Somebody else's product, with somebody else's opinions baked into it. You adapt to the tool.
This is right for work that's genuinely the same everywhere. Transcribing meetings. Drafting first-pass copy. Summarizing documents. Nobody has ever won business by having a distinctive approach to turning speech into text — pay the $30 a seat and stop thinking about it.
The failure mode is buying for work that isn't generic. You end up with a tool that does 70% of your process and can't be told about the other 30%, so your team runs the tool and the old way in parallel. That's worse than either one on its own, and it's the most common shape of failed rollout I see.
Configure
A system that already exists, set up around your rules. Your job types, your escalation policy, your tone, your data. Standard engine, your behavior.
Roughly 70% of SMB cases belong here, and it's the option people forget exists. They think the choice is between a $30-a-month tool and a $60,000 custom build, so they take the tool. In practice a configured system aimed at one specific job — inbound calls, lead routing, document intake, follow-up — lands somewhere between $500 and $2,500 a month all in, ours included.
The reason it works is that the hard, expensive parts of these systems are nearly identical across businesses. The parts that differ are exactly the parts that should be configuration: what counts as urgent, who gets called at 9pm, which questions must never be answered without a human.
Build
From scratch, or close to it, because the workflow itself is the thing you're good at.
Right when the process is the differentiator. The pricing model nobody else has, the underwriting logic that's the reason customers pick you, the proprietary data nobody else holds. Building that is defensible. Building your invoice reminders is not.
The honest cost isn't the build, it's that you now own it. The model changes, the edge cases, the person who has to care about it in year two. Budget for the second year before you commit to the first.
Three questions that decide it for you
Is this how we win?
If a competitor did this exact thing identically, would you lose anything? If not, you're in buy or configure territory. Very few processes inside a 40-person company are genuinely differentiating, and it pays to be unsentimental about which ones.
How often does the process change?
A process that changes monthly will fight a bought tool and outrun a custom build. Configuration handles change well — changing it means editing a rule instead of filing a feature request or booking developer time.
Who fixes it at 2am?
Buy: the vendor, on their timeline. Configure: your partner, or you, depending on the contract — ask before you sign. Build: you, forever. If the honest answer to this one is "nobody," you can't build, whatever the other two answers were.
The same problem, three ways
A conversation I keep having, in a few different shapes: a 30-person company with inbound leads, and all three vendor types have already quoted. Here's how each would actually play out.
Bought, they end up with a shared inbox tool with AI replies. Live this week, $50 a seat. Works if their leads all look alike and their reply is basically the same every time. Breaks the moment routing depends on something the tool has no way of knowing.
Configured, they end up with a qualification and routing agent set up around their criteria, their CRM, their handoff rules. Four to six weeks, low four figures monthly. Right when their definition of a good lead is specific and their sales team is small enough that misrouting hurts.
Built, they end up with custom scoring on their own historical close data, wired into something nobody sells. Months, and a real budget. Justified when lead scoring is genuinely their edge — which at 30 people, it usually isn't.
Most companies in that position should configure. Most companies in that position buy, because buying requires no decisions, and then conclude a year later that AI didn't work for them.
The expensive mistake in each direction
Bought when you should have configured. A tool nobody adopted, a team that now believes this doesn't work here, and — the part that actually costs you — twelve months before anyone will fund another attempt.
Built when you should have bought. $40,000 and four months producing something a $40-a-month product does better, because that product has ten thousand customers finding its edge cases and yours has one.
Configured when you should have built. Rarer, and much the cheapest of the three. You outgrow it, you know exactly what you need because you've run the thing for a year, and version two is better for having been informed by version one.
Which gives you a practical tiebreaker: when it's genuinely close, configure. It's the option that's cheapest to be wrong about.
Working out which of your processes belongs in which bucket is what the two-week audit is for, and its scoring maps almost directly onto this decision. Whichever of the three you land on, ask for the vendor's definition of done before you sign anything.