In this article
Key Takeaways
- Yes. AI helps in three places: design time (suggested steps and flow templates), runtime (LLM steps that classify, draft, summarise), and observability (anomaly detection on failed runs).
- Gartner expects 40% of new integrations to use AI agents by end of 2026, up from a small fraction in 2024, with iPaaS as the main delivery layer.
- Albato exposes OpenAI as a step inside flows for classification, drafting, and summarisation. The pricing model is per successful action, which keeps AI-heavy flows predictable.
- The largest gains come from runtime use (AI inside a flow), not from AI building the flow itself. Start there.
The average company runs 342 SaaS apps in 2024 (Productiv, State of SaaS Usage 2024). McKinsey estimates 57% of US work hours are technically automatable (McKinsey, Agents, Robots, and Us 2025). AI is what closes the gap between deterministic if-this-then-that flows and the unstructured judgement calls that used to need a person. Gartner reports the iPaaS market grew 23.4% to $8.5 billion in 2024 (Gartner, iPaaS Market Share 2024) with AI agents projected to drive 40% of new integrations by end of 2026.
Three places AI changes workflow design
AI is not a single feature on top of automation. It changes work in three different layers.
| Layer | What AI does | Typical use |
|---|---|---|
| Design time | Suggests next steps, templates, flow patterns | New users picking the right trigger |
| Runtime | Classifies, drafts, summarises, decides | LLM step inside an existing flow |
| Observability | Detects anomalies, suggests fixes | Alert when a flow run looks off |
The table above gives the quick view. The sections below cover each point in more depth.
The largest gains in 2026 are at runtime, because that is where AI replaces a human judgement call inside an automated flow. Design-time help is real but secondary. Observability is the newest layer and the least mature.
AI at runtime: the real gains
Runtime AI means an LLM step lives inside a flow. The step receives data from a previous step (a lead, a ticket, a draft, a transcript) and outputs structured or unstructured text that the next step writes somewhere useful.
Four runtime patterns account for most of the value.
- Classification. Incoming ticket gets a category, priority, and routing tag before it lands on a queue.
- Drafting. Incoming email or deal note gets a suggested reply that the human can edit and send.
- Summarisation. Long transcript or thread gets a 3-bullet summary in Slack or CRM.
- Decision support. Lead with company size, industry, and stack gets a scoring note that helps sales prioritise.
For specific OpenAI patterns inside Albato flows, see what tools connect to OpenAI.
How Albato runs AI inside flows
Albato lists OpenAI at albato.com/apps/openai and integrates major LLM providers as steps. The step accepts a prompt built from prior step data, a model choice, and optional parameters. The output is plain text that any later step can consume.
The pricing model is per successful action. An AI step that returns a usable output counts once. Failed runs and filter steps do not. This matters because AI flows tend to fan out across many records (every lead, every ticket, every order), and per-task billing on other platforms can balloon at that scale.
Want to add an AI step to your real workflow? Start a free Albato account and try OpenAI inside any flow.
How to design a smarter workflow with AI
Five rules keep AI flows useful instead of expensive theatre.
1. Start with one judgement call. Pick the step in an existing flow where a human currently decides (categorise, prioritise, draft). Replace just that step with AI, not the whole flow.
2. Use the cheapest model that passes. GPT-4o-mini is enough for most classification and tagging. Upgrade only when output quality is measurably below the bar.
3. Validate output before it writes to systems of record. Add a structured-output check or a human-review queue for the first month. Then loosen the loop based on real data.
4. Log prompts and outputs. Without logs, debugging a flaky AI step takes longer than the manual work it replaced.
5. Re-check the savings monthly. If a flow runs 100 times per month, AI is overkill. If it runs 10,000 times, even modest accuracy is a win.
💡 Tip
AI inside flows is most valuable for tasks done many times per month that need judgement. One-off decisions and high-volume deterministic flows do not benefit.
What AI does not do well in workflows
Three categories where AI inside flows is the wrong tool.
A deterministic data transformation (CSV row to JSON, currency conversion, date format) belongs in a regular helper step, not in an LLM. The AI step will be slower, more expensive, and occasionally wrong on a job that has a guaranteed right answer.
A high-stakes irreversible action (refund, payment, account deletion) should not auto-fire from an AI decision without human review. Build a "draft and confirm" pattern instead.
A long-running deep analysis (build a research report, evaluate 50 vendors) belongs in a chat tool with iterative prompting, not a one-shot flow step. For these, see the guide on Claude artifacts.
Curious how AI inside flows fits your stack? Open a free Albato workspace and explore the OpenAI templates.
FAQ
Can AI design my workflow for me in Albato?
AI helps suggest steps and templates, but the human still decides the trigger and the goal. The larger value is AI inside the flow at runtime, where it replaces a human judgement call on every record (classify, draft, summarise).
What AI models does Albato support?
Albato connects to OpenAI (GPT-3.5, GPT-4, GPT-4o, GPT-4o-mini), with DALL-E for images and embeddings for semantic search. Other major model providers are available as separate connectors.
Is AI inside iPaaS reliable?
Yes for structured outputs (classification, tagging, simple drafting). Add a validation step or a review queue when output quality matters. Logs and retries inside Albato cover the operational side; output quality depends on the prompt and the model.
How much does an AI step cost in Albato?
Albato bills per successful action. OpenAI itself bills per token. A short classification on GPT-4o-mini costs fractions of a cent per call. A long draft on GPT-4 costs more but still under a few cents per call. Estimate monthly cost from the run volume and model choice.
Do I need to know prompt engineering?
Basic prompts work for most classification and short drafting jobs. Spend time on the prompt only when output quality is below the bar; otherwise simple prompts plus a small validation step usually pass.
Can AI rewrite my existing manual workflows?
No, you still describe the workflow. AI can suggest steps and accelerate the build, but the trigger, the data shape, and the destination are human decisions. Start with one judgement call, expand from there.
Want to keep exploring before committing? Sign up free and try Albato Automate.
Want more? Explore related guides from the Albato blog below.
