AI automation for business in 2026: what to automate, n8n vs Zapier vs Make, and what it costs

What is AI automation?

AI automation is workflow automation where some steps use an AI model to read, extract, classify or draft, and the rest are fixed steps you can predict. An invoice arrives, a model reads it into fields, code checks the fields against your rules, and the bill is created in your accounting tool. The same steps run every time; only the reading needs AI.

That is what separates it from an AI agent, which decides its own next steps. Most repetitive back-office work, often called business process automation, is better served by AI workflow automation: it is cheaper to run, easier to test and easier to audit.

Every reliable AI workflow has the same five parts:

1. Trigger. A webhook, a new email, a file in a folder or a schedule.
2. Extract. The model turns messy input into structured data that matches a fixed schema.
3. Validate. Code, not the model, checks that data against your rules and decides whether it can proceed.
4. Act. The workflow writes to your systems, in a way that is safe to retry.
5. Log and review. Every run is recorded, and anything that failed a check waits in a queue for a person.

What should you automate first?

Pick work that is frequent, follows written rules, has a checkable right answer and can be undone. Good first candidates:

• Invoice and document processing: read the document, extract vendor, number, dates, lines and total, create the bill.
• Inbox triage: classify incoming email by request type and route it, with a drafted reply for a person to send.
• Lead enrichment and routing: research a new lead, score it against your criteria, assign it in the CRM.
• CRM and spreadsheet syncing: keep records consistent across tools so nobody types the same data twice.
• Scheduled reports: pull the week’s numbers and write a short summary.

Leave for later anything that moves money without approval, makes legal, medical or hiring decisions, or has no agreed right answer to test against.

Does AI automation pay off?

Adoption is broad and mostly shallow. In the Federal Reserve’s survey of small employer firms, 46% said they or their employees use AI, but just 7% of those users had fully integrated it, and accuracy was the most common challenge. The Census Bureau’s business survey found AI use hovering between 17% and 20% of US businesses in the first half of 2026.

Returns are real for a minority. In McKinsey’s 2026 survey, 89% of respondents said their organizations used AI regularly, but only 37% attributed any profit impact to it, and about 20% said AI running costs, tokens included, had limited their use. A 2025 MIT report, preliminary and with a disputed method, claimed 95% of generative AI pilots showed no measurable return; it also found tools bought from or built with outside specialists succeeded about 67% of the time, and internal builds about a third as often.

Time savings per person are modest. Generative AI users saved 5.4% of their work hours, about 2.2 hours a week, in a St. Louis Fed survey, and a large Danish study found average time savings of 3%. That is the case for automating whole workflows rather than adding another chat assistant: the time each person saves is small, while a workflow that runs without anyone removes the task.

Test extraction on your own documents, because there is no strong independent benchmark for it. In one 2026 study of electricity invoices, the best model scored an F1 of 97.6% when given a few worked examples, and more than 19 points lower without them. Examples in the prompt are cheap; use them.

How do you make an AI automation reliable?

The model is the least predictable part, so the design puts code around it.

Extract to a schema, then let code decide

Ask the model for JSON that matches a fixed schema, and validate it in code even when the API promises the format. Then route each result with checks the model can’t talk its way past:

type Invoice = {
  vendor: string;
  invoiceNumber: string;
  issueDate: string; // YYYY-MM-DD
  currency: 'USD' | 'EUR' | 'GBP';
  lines: { description: string; amount: number }[];
  total: number;
  confidence: number; // the model's own 0–1 estimate: one signal, not proof
};

/** Code, not the model, decides what happens next. */
function route(inv: Invoice, knownVendors: Set<string>, approvalLimit: number): 'auto' | 'review' {
  const sum = inv.lines.reduce((s, l) => s + l.amount, 0);
  const checks = [
    Math.abs(sum - inv.total) < 0.01, // the lines add up to the total
    knownVendors.has(inv.vendor), // a vendor you already pay
    inv.total <= approvalLimit, // under your approval limit
    inv.confidence >= 0.9, // and the model says it is sure
  ];
  return checks.every(Boolean) ? 'auto' : 'review';
}

The arithmetic and lookup checks do most of the work. A model’s own confidence score is easy to log but not reliable on its own, so never let it be the only gate.

Make every action safe to retry

APIs time out and workflows retry. Without a guard, a retry creates the same bill twice. Give each record a natural key and let the database refuse duplicates:

create table bills (
  id bigint generated always as identity primary key,
  vendor text not null,
  invoice_number text not null,
  total numeric(12, 2) not null,
  created_at timestamptz not null default now(),
  unique (vendor, invoice_number)
);

-- A retried run inserts nothing the second time.
insert into bills (vendor, invoice_number, total)
values ($1, $2, $3)
on conflict (vendor, invoice_number) do nothing;

The same idea applies to outside APIs: send an idempotency key where the API accepts one, and check for an existing record before creating one where it doesn’t.

Log every run, and alert on silence

Record for each run what came in, what the model returned, which checks passed, what was written and what it cost. Alert on failures, and also on silence: a workflow that stopped triggering looks exactly like a quiet week.

Cap the cost

Count tokens per run, keep a running monthly total, and stop the workflow at a ceiling you set. A loop that re-triggers itself is the usual way an automation runs up a bill overnight.

n8n vs Zapier vs Make vs Power Automate: what do they cost?

Prices from each vendor’s pricing page or documentation on October 10, 2026. The platforms meter different things, which matters more than the headline price.

n8n publishes its prices in euros. The Community Edition’s Sustainable Use License allows use “only for your own internal business purposes or for non-commercial or personal use,” so you can’t resell it as a hosted service.

Here is how the metering plays out for one workflow, as an illustration rather than a quote: 1,000 one-page invoices a month, each run doing six steps (trigger, AI extraction, vendor lookup, checks, create the bill, notify).

Zapier and Make charge more as a workflow gains steps; n8n charges per run whatever its length. That is why long workflows drift to n8n or code, and short ones stay on Zapier.

The model itself is rarely the expensive part. Prices per million tokens from each provider’s pricing page, and the cost of one invoice at about 3,000 tokens in and 500 out:

Start with a small model and move up only for the documents it gets wrong. The checks above tell you which ones those are.

Should you use a platform or write code?

Zapier or Make, if

• The workflow connects standard apps with a few steps, and volume is modest.
• Nobody on the team will maintain code, and the per-task or per-credit bill stays small.

n8n, if

• Workflows are long, so per-run pricing beats per-step pricing.
• You want to self-host for data control and have someone to patch it. n8n releases a new minor version most weeks.

A minimal self-hosted setup, with Postgres instead of the default SQLite and a pinned version:

docker volume create n8n_data
docker run -d --name n8n --restart unless-stopped \
  -p 127.0.0.1:5678:5678 \
  -e N8N_ENCRYPTION_KEY="$N8N_ENCRYPTION_KEY" \
  -e DB_TYPE=postgresdb \
  -e DB_POSTGRESDB_HOST="$DB_HOST" \
  -e DB_POSTGRESDB_DATABASE=n8n \
  -e DB_POSTGRESDB_USER=n8n \
  -e DB_POSTGRESDB_PASSWORD="$DB_PASSWORD" \
  -v n8n_data:/home/node/.n8n \
  n8nio/n8n:2.42.6

The encryption key protects the credentials n8n stores, so keep it in a password manager: without it, saved credentials can’t be decrypted. Bind to localhost and put a reverse proxy with TLS in front, and update deliberately rather than never.

Power Automate, if

• You live in Microsoft 365 and the workflow stays inside it.

Custom code, if

• Steps need judgment against your own rules, a review queue and per-run logs.
• Per-task fees would grow with volume, or you want the workflow in your own repository and accounts.

We build the last kind: one workflow, end to end, with the checks, review queue, logs and cost ceiling above, in about a week. See AI automations. If Zapier or Make already does the job, we’ll say so.

What can go wrong with AI automation?

Prompt injection

An AI step that reads email or documents reads whatever an attacker writes in them. OWASP ranks prompt injection first among LLM risks and notes that indirect injection comes from “external sources, such as websites or files”. It isn’t theoretical: in 2025 Microsoft fixed an AI command injection in Microsoft 365 Copilot that let an attacker disclose information over a network.

The defenses are structural:

• Give the extraction step no tools. It returns data that matches the schema, and nothing else.
• Let code choose actions from a fixed list. Text in a document never decides a recipient, an account or an amount.
• Keep a person on anything that moves money or data outside the business, as OWASP recommends for privileged operations.

Platform vulnerabilities

Self-hosting means you own the patching. In 2026 n8n fixed an unauthenticated flaw rated 10.0 in its form workflows, and CISA lists an n8n code-execution flaw from December 2025 as exploited in the wild. Keep instances off the public internet where you can, and update on a schedule. Hosted platforms have incidents too: in 2025 Zapier said customer data had been copied into code repositories accessed through a two-factor misconfiguration, and in 2026 it patched a bug chain that could have allowed account takeover, with no evidence it was exploited.

Trusting it too much

NIST’s generative AI guidance warns of “excessive deference to automated systems” and recommends recording where people override the system. Log every override from the review queue: it is your best list of what to fix next.

How do you roll out an AI automation?

1. Pick one workflow. Write down its trigger, its steps, the systems it touches and who approves what.
2. Collect real test cases. 50 to 100 past examples with the right outcome, including the awkward ones.
3. Build with the checks first. Schema, arithmetic, lookups, approval limits and duplicate guards before the happy path.
4. Run in shadow mode. Let it process real inputs and compare its output with what your team did, without writing anything.
5. Go live with a review queue. Start strict, so most items go to a person, and loosen the thresholds as the logs earn trust.
6. Watch the numbers. Share of runs handled automatically, review-queue size, errors, cost per run, and runs per day.

What next?

Write down the one task your team repeats most often, how many times a month it happens and how long each one takes. If it follows rules you could explain to a new hire, it can probably be automated. When you’re ready, send us your requirements or see how a one-week AI automation build works.