Buy it, build it, or wait for it

Search first

A team spent three weeks building a receipt-filing agent. Their accounting package had done receipt capture with OCR for two years, on the mobile app, under a menu nobody had opened. This is common enough to be a rule: spend an hour reading the feature list of the tool that already holds the data before you design anything at all.

Look specifically inside the system of record: the CRM, the accounting package, the helpdesk, the spreadsheet suite. Vendors ship AI features into those products faster than they update their marketing pages, and a native feature comes with the auth, the audit log and the support contract already attached.

Three costs, and the one that gets forgotten

Every automation has a build cost, a run cost and a maintenance cost. The first is a week you can see. The second is a monthly bill you can price. The third is what kills projects, and it is paid in attention rather than money.

Over a year, expect a model you called by name to be deprecated or repriced, an API you depend on to change a field, a login to expire, a colleague who understood it to leave, and the upstream file format to shift. None of these are unusual events; together they mean the thing needs a competent person several times a year. Budget the person. There is no trustworthy industry figure for what proportion of build cost this is, and anyone quoting one precisely is guessing, but the direction is not in doubt, and a project with no named maintainer has a life expectancy measured in months.

When buying wins

You are mostly paying for connectors and for somebody else’s on-call rota. A hosted automation tool maintains hundreds of integrations, each with its own auth dance, pagination style and rate limits, and absorbs the breakage when a vendor changes an endpoint. If your flow touches nine services, that maintenance is the product and it is worth real money.

You are also buying compliance paperwork: the security questionnaire, the data-processing agreement, the certifications your customer’s procurement team will ask for.

When building wins

Three situations. First, when the data cannot leave: a hospital’s records, a lab’s formulations, anything under a contract that names the countries it may sit in. Second, when the vendor’s shape fights the work: the moment you need a loop with a condition inside it, a code review, or a rollback to last Tuesday’s version, a visual builder becomes the expensive option. Third, volume.

Per-task pricing is the quiet one. If a tool bills per step and one job is five steps, then two thousand jobs a month is ten thousand billed tasks. Compare that against a small virtual server running n8n or a plain cron job on a machine you already have. Prices move constantly, so check the current ones rather than trusting any number you read in a course, but the shape of the curve does not move: per-task pricing is cheap at pilot volume and turns hard against you somewhere in the low tens of thousands, and pilots grow.

When waiting wins

If the platform that already holds your data has the feature in beta, an integration you build now has a short life. Ask about the roadmap, treat the answer as marketing, and wait only when there is a dated commitment or a beta you can actually log into. “Coming soon” is not a plan you can schedule around.

The “it’s just an API call” fallacy

The call is about five per cent of it. The rest is authentication and token refresh, rate limits and backoff, pagination, retry on the transient errors and not on the permanent ones, mapping their error codes to yours, handling the schema changing under you, storing the secret somewhere that is not the repository, logging enough to debug it at 3am, and a named person who will do that. Every one of those is an afternoon, and together they are the integration.

The free paths

Self-hosting is real and it is not free: it moves the cost from a bill to your evenings, which is a trade, not a saving. With that said: n8n self-hosts and is the usual answer for visual flows, though its licence is “fair-code” rather than OSI-approved open source, so read it if you plan to resell. Node-RED is Apache-2.0 and excellent for event wiring. Windmill runs scripts with a UI on top. Apache Airflow or Prefect for scheduled data pipelines. And beneath all of them, cron plus thirty lines of Python still runs an enormous amount of the world’s automation. Zapier and Make have free tiers with task caps that are fine for proving a flow works and too small to run one.

The decision in one paragraph

Write down what happens if you do nothing for six months. Sometimes the answer is “we keep paying two hours a week”, which is a hundred hours a year and worth solving. Sometimes it is “nothing, this ran twice last quarter”. The second answer is not a defeat. It is the cheapest result this course can give you.