Buy, build, or wait: choosing the right kind of AI for your business

Once you know which problem to solve first and you’ve mapped how the work actually flows, there’s one question left: what kind of AI belongs there? In our experience there are only three honest answers — buy it, build it, or wait — and being tech-agnostic means we recommend all three, regularly.

Buy: when the software you already run does it

The systems small businesses already pay for — AppFolio, Clio, QuickBooks, Bullhorn, Applied Epic, NetSuite, and the rest — are shipping AI features every quarter. If the workflow is generic (categorising transactions, summarising a call, drafting a standard reply) and lives entirely inside one system, the right move is to turn the feature on and test it for two weeks before spending a dollar on custom work.

The signs point to buy when: it’s one system; the task is the same across every business like yours; you don’t need much customisation; and the vendor already holds the data. The cost is often nothing, or a small add-on. Two cautions: read where the data goes and whether it’s used for training, and confirm you can switch it off. If a native feature covers 90% of the job, take the 90%.

Build: when the workflow is yours

Custom makes sense when the process spans systems — email to your property-management software to a spreadsheet to an owner report. When your edge cases are the job: the self-employed borrower, the multi-location claim, the carrier with a lapsed insurance certificate. When the off-the-shelf version ignores the 20% of cases that cause 80% of the pain. Or when volume is high enough that a percentage point of accuracy is real money.

What a build looks like with us: three to six weeks, typically $8,000–$35,000 depending on scope, integrated into the tools you already use, with a human review step exactly where it belongs, handoff documentation, and a 30-day check-in after launch. What “build” almost never means is training your own model. It means wiring proven models into your workflow with the right guardrails, so the output arrives where your team already works — the case file, the ticket, the PO — rather than in yet another tab.

Wait: when the honest answer is “not yet”

This is the recommendation most vendors won’t give you, and it’s the one we make surprisingly often. Three situations call for it:

• The process is broken. Automating a broken process just makes it break faster. If two people on the team disagree about how a thing is supposed to happen, fix the handoff first. That costs nothing and often recovers more time than the automation would have.
• The data isn’t there. The information lives in people’s heads, or in inconsistent PDFs with no history. You may need 60–90 days of simply capturing it properly before there’s anything to automate.
• The volume doesn’t justify it. If it’s two hours a month, a checklist beats a build. Every time.

Waiting isn’t failure — it’s sequencing. In a meaningful share of our audits, the first recommendation on the roadmap is a process change that costs nothing, and the AI build comes second.

Five questions that usually settle it

1. Does it happen more than twenty times a week?
2. Could a person check the output in under a minute?
3. Does the input already exist in a system, not in someone’s head?
4. Does one vendor’s tool cover the whole workflow — or does it stop at the edge of their system?
5. What does being wrong once cost — an apology, or a regulator?

Mostly yes to 1–3 and no to 4: build. Yes to 4: buy, and test it first. No to 1 or 3: wait, and fix the input. Question 5 doesn’t change the answer so much as it decides where the human review goes — and for anything with regulatory weight, that review step is non-negotiable.

A note on vendor bias

Everyone who recommends AI to you has an incentive. Software vendors want you inside their walls. Agencies want to build. We’re not exempt — we do builds. That’s exactly why we charge for the operations audit separately from any build, and put buy and wait recommendations in writing alongside the build ones. A fair question to ask anyone advising you on AI: what’s the last thing you told a client not to build? If they can’t answer, keep asking around.