Build or buy
Zapier, Make and Power Automate, or a custom AI agent?
How to tell whether a process needs an off-the-shelf automation platform, a custom AI agent, or a bit of both, and when it needs neither.
What the off-the-shelf platforms are good at
Zapier, Make and Microsoft Power Automate are trigger-and-action platforms: when something happens in one app, do something in another. They have large libraries of pre-built connections, visual editors that non-developers can learn, and increasingly their own AI steps for summarising or extracting text.
They are a good fit when the process is predictable, the data is structured, and the volume is moderate. Copying a web form into a CRM, posting a message when an invoice is paid, or filing attachments into the right folder are classic examples. Power Automate is attractive for businesses already on Microsoft 365 because it lives in the same environment; check which connectors your licence includes, because some are on higher plans.
Their weaknesses show up as processes get more complex: long chains of conditions become hard to maintain, usage-based pricing can climb with volume, and handling the messy exceptions often means bolting on more and more steps.
What a custom AI agent is good at
A custom agent is software that uses an AI model to decide what to do next, with access to a defined set of tools and data, rather than following a fixed path. It is well suited to work that involves reading unstructured information, choosing between several possible actions, and coping with cases nobody anticipated, such as triaging a shared inbox, preparing a quote from a set of plans and notes, or answering staff questions from policy documents with the source attached.
The trade-offs: it needs more design and testing, it has to be hosted and monitored, and it needs a clear boundary on what it may do without asking. Built well, with approval steps and logs, it handles the work that otherwise falls back on a person. Built badly, it is an unpredictable system with access to your data.
A quick way to decide
| The process involves… | Usually best served by | Why |
|---|---|---|
| Moving data between apps when something happens | Off-the-shelf platform | Predictable triggers and fields are exactly what these tools were made for. |
| Reading varied documents (invoices, referrals, forms) | AI step, often inside a platform | Fixed rules break on layout changes; an AI model can extract fields with a person checking. |
| Drafting replies or documents from context | AI step with human approval | Drafting is the strength of language models; sending without review is the risk. |
| Choosing between several actions based on judgement | Custom agent | When the right next step depends on reading the situation, a fixed flowchart gets long and fragile. |
| Anything touching money or legal commitments | Any approach, plus mandatory approval | The tool matters less than the approval step and the log. |
| A process that happens a few times a month | Often neither | A checklist or a template may be the better answer. |
Why most real projects are a mix
In practice, the plumbing (moving data, notifying people, filing things) is often cheapest on a platform you already pay for, the reading and drafting is done by an AI step, and the decision stays with a person. Building everything custom when a platform would do wastes money; forcing a judgement-heavy process into a rigid flowchart wastes time and produces something fragile.
The audit stage is where this gets decided per process, with the running costs of each option estimated from the providers’ current pricing. If an off-the-shelf tool can do the job and your team can maintain it, that is what we will recommend, even though it means a smaller engagement.
Questions to ask yourself before either
- How often does this process actually run, and how long does each run take?
- Can the people doing it describe every step, including the exceptions?
- Who would approve the output, and would they have time to?
- Which systems hold the data, and do they have working integrations?
- What is the worst thing that could happen if the automation got it wrong once?
If the last answer is serious and the first answer is “rarely”, the right automation may be no automation.