There’s a gap between the automation people imagine (“the AI runs the business”) and the automation that quietly delivers value every day (“the AI moved this record, checked that condition, and drafted that reply, 400 times, without error”). The second kind is less exciting and far more profitable. Here’s how we think about where AI-driven workflow automation pays off — and where it doesn’t.
Automation ≠ a chatbot
A conversational assistant is one interface. The engine underneath — the part that does things — is a workflow: a defined sequence of triggers, conditions, and actions across your tools. AI makes those workflows smarter (it can read unstructured text, classify intent, draft language, decide a branch), but the reliability comes from the workflow structure, not the model’s improvisation. We build the two together: an orchestration layer that coordinates each step, with AI slotted in exactly where judgement is needed.
The sweet spot: high-volume, rule-shaped, judgement-light
The best first automations share a shape:
- Repetitive — it happens dozens or hundreds of times, so saved minutes compound.
- Mostly deterministic — the rules are knowable, with AI handling the fuzzy edges (reading a message, extracting fields, categorising).
- Low blast radius — a mistake is recoverable and reviewable, not catastrophic.
Concrete examples we see work:
- Lead handling — a form or chat comes in; the system enriches it, classifies it, writes it to the CRM, and notifies the right person. (Our own site assistant captures leads into a tracking sheet as part of the conversation.)
- Document processing — pull key fields from invoices, contracts, or forms and route them onward.
- Reporting — gather numbers from several systems on a schedule and assemble a draft summary a human signs off.
- Triage — read inbound support or email, tag by topic and urgency, draft a first response for review.
Where to keep a human in the loop
Automation should augment judgement, not quietly replace it. We deliberately keep a person in the loop when: the decision is irreversible (payments, contracts, anything legal), the cost of a confident error is high, or the process touches sensitive data and needs an audit trail. The right design isn’t “AI does it all” — it’s “AI does the 90% of grunt work and hands the 10% that needs a human to a human, with context.”
The failure modes to design against
We’ve learned to build against a few predictable ways automations go wrong:
- Silent failure. A workflow that breaks quietly is worse than one that errors loudly. Real monitoring and alerting isn’t optional.
- The happy-path trap. Demos handle the clean case. Production handles the malformed input, the duplicate, the timeout. That edge-case handling is most of the engineering.
- Fake confidence to the user. If a step didn’t actually complete, don’t tell the user it did. (This is a small thing that destroys trust fast — we treat “only report success when it truly succeeded” as a rule.)
How to start without overcommitting
Map one process end to end. Count how often it runs and how long each run takes by hand. Automate the deterministic spine first, add AI only where a human is reading or deciding something fuzzy, and instrument it so you can see it working. Prove the time saved on one workflow, then reinvest that into the next.
That “one workflow, measured, then widen” approach is how we take a team from a manual, error-prone process to a monitored automation that runs itself — with people focused on the decisions that actually need them.