Most "best AI automation tools" lists are affiliate pages. This is the stack we actually build on, why each piece is there, and — more usefully — where each one stops working.
The short version: the model matters less than people think, and the monitoring matters far more.
Orchestration: what runs the workflow
Zapier — the best onboarding experience and the largest integration library. Genuinely the right choice for simple connections at modest volume. Ceiling: per-task pricing at scale, and complex branching becomes unmaintainable.
Make — cheaper per operation, better visual branching. Ceiling: still a visual builder, so genuinely complex logic gets unwieldy.
n8n — self-hostable for a flat cost, native code steps, no per-task fee. Our default when volume is high or data must stay on the client's infrastructure. Ceiling: somebody has to own uptime, backups and upgrades. We compared it properly in n8n versus Zapier.
Custom code — Python or Node running on the client's infrastructure. Used when the logic genuinely needs it, not by default. Ceiling: nobody edits it on a Tuesday afternoon without a developer.
Most real builds are a mix. Keeping simple connections in a no-code tool and writing custom only where it breaks down is usually cheapest over three years.
Models: less important than the marketing suggests
We route each step to the cheapest model that does it well rather than sending everything to the largest one — which typically cuts running costs substantially with no quality loss.
- Frontier models (Claude, GPT) for reasoning, writing and anything needing real judgement.
- Smaller, cheaper models for classification, routing and simple extraction. Most steps in most workflows are this.
- Open-weight models where data cannot leave the client's infrastructure.
The important architectural decision is not which model — it is making the model choice swappable. Providers change pricing and capability frequently, and a build hard-wired to one is a build you will pay to change.
Document extraction
The highest-value capability for most businesses, and the one where AI clearly beats what came before. Older OCR needed a template per document layout; modern extraction reads formats it has never seen.
What matters in implementation is not the model — it is confidence scoring and a human queue behind it. Extraction without a threshold is how wrong numbers reach accounting systems.
Retrieval, for anything that answers from your documents
Chatbots and knowledge assistants need to search your content and answer only from it. That means a vector store — pgvector where a Postgres database already exists, dedicated services at larger scale.
The quality lever here is chunking and retrieval strategy, not the database. A well-tuned retrieval layer on a simple store beats a badly tuned one on an expensive service, every time.
The boring parts that decide whether it survives
This is the section that separates a system still running in a year from one quietly abandoned.
- A real queue — so a burst of work does not lose records.
- Retries with backoff — third-party APIs fail transiently; most failures resolve themselves on retry.
- Idempotency — so a retry does not create a duplicate invoice. Skipped surprisingly often.
- Structured logging — every action, input and decision, queryable when something goes wrong.
- Alerting to a person — failures and anomalies, through a channel they actually read.
- An evaluation set — real examples with known answers, re-runnable when a model updates.
None of this is exciting and all of it is why systems keep working. If a vendor's proposal covers models in detail and this in one line, that tells you what they have and have not run in production.
What we would not build with
- Anything that stores your credentials in the workflow definition. Secrets belong in a secrets manager.
- Platforms that own your automation. If it stops when you stop paying and you cannot export it, you are renting.
- Model providers who train on customer data. Enterprise endpoints, always.
- Anything with no audit trail. Fine for a prototype, not for a system touching money or customers.
Frequently asked questions
Which tool should I start with?
Zapier, honestly, unless you already know your volume is high. It is the fastest way to find out whether the process is worth automating at all, and outgrowing it is a good problem — we have written up what to do when you do.
Does the model choice matter much?
Less than the surrounding engineering. A mediocre model with good retrieval, confidence thresholds and exception handling beats the best model wired up carelessly. Model choice mostly affects running cost.
How often does the stack change?
The model layer moves constantly; the orchestration and reliability layers barely at all. That is why we keep model choice swappable and treat everything else as long-lived infrastructure.
Can we use our existing tools?
Usually, and we would rather you did. Building on what you already pay for is cheaper and less disruptive than a migration you did not ask for.
Want a stack recommendation for your specific situation? Free 30-minute call, and we will tell you when the answer is a tool you can configure yourself. See how we work or message us on WhatsApp.
