The 80/20 of business automation — start here.
You probably don't need an AI agent yet. You need a Zapier you actually maintain. The unglamorous order we use to sequence automation work.
Nobody needs an AI agent yet. Almost everyone needs a Zapier they actually maintain.
The temptation, especially in 2026, is to skip ahead to the agent demos. But across our automation engagements, the highest-leverage work has always been deeply unsexy: workflows that just run, every day, without breaking.
The three categories of automation
Before adding any AI, almost every team has gains in three areas:
1. The integration layer
Tools that don't talk to each other. Lead from a portal that copies into a CRM, that copies into a spreadsheet, that copies into a Slack channel. Each copy is a person, each person is a delay. Replacing this layer alone usually clears 60% of an ops team's coordination work.
2. The trigger layer
Things that should happen at a specific time but happen when someone remembers. Renewal reminders, follow-up sequences, contract expiry alerts, monthly reports. Computers are extremely good at this. People are extremely bad.
3. The dashboard layer
Reports that are produced manually because somebody downloads three CSVs and combines them in Excel. The CSVs always live in databases that could be queried. The Excel always lives somewhere a dashboard could live.
The order to do them in
For most teams, the right order is: integrations → triggers → dashboards. Why? Because integrations make the data trustworthy, triggers act on it, dashboards summarise it. Doing them out of order means dashboards lying to you about events that didn't happen — which is worse than no dashboards.
Doing dashboards before integrations is how you end up with a beautiful chart of garbage.
The watcher problem
The number-one reason teams give up on automation is silent failure. A workflow stops running, nobody notices, the business runs on bad data for two weeks, trust is gone forever.
Solution: every automation needs a watcher. A second-tier process that pings if a workflow hasn't run in expected windows, a Slack channel that gets every error, and a weekly digest of "things that ran successfully" that an operator skims. Without this, automation rot is inevitable.
Where AI fits
Once your integrations, triggers and dashboards are working, AI gets useful. The patterns that earn their keep so far in our work:
- Classification: routing inbound items to the right queue.
- Extraction: pulling structured data out of unstructured docs (contracts, emails).
- Summarisation: pre-writing the long-form bits humans were going to copy-paste anyway.
- Decision support: surfacing the next-best-action without taking it.
What earns its keep less often (so far): full agentic workflows where the AI takes meaningful actions autonomously. The cost of being wrong is high; the eval discipline required is significant. Walk before running.
The 80/20 sequence
- List your top ten coordination workflows. Rank by hours/month.
- Take the top three. Get them off email and into a tool.
- Build watchers. Get serious about alerting.
- Add the trigger layer (recurring sequences, expiries, reminders).
- Replace your top three manual reports with dashboards.
- Now look at AI.
Most teams that follow this sequence find they don't need AI yet. The teams that do find that AI is now ten times more effective than it would have been if they'd started there.