The fastest way to kill an automation project is to make staff responsible for decisions they can't see or undo. They'll quietly check everything by hand, and your savings vanish.
Regulators say the same thing. India's AI Governance Guidelines ask that AI outputs can be reviewed, overridden or supplemented by human judgement before they cause harm (MeitY, 2025). The RBI's FREE-AI committee is blunter: accountability can't be delegated to the model (RBI, 2025).
1. Sort the work by risk, not by volume
Most verification work, like matching an invoice to a purchase order or checking a document is complete, is repetitive and low risk. Automate that fully. Keep people on the cases where a wrong answer costs money, breaks a law or hurts a customer.
| Lane | Example | Who decides |
|---|---|---|
| Routine | Fields match, amount under limit | System, with a weekly sample audit |
| Uncertain | Low confidence or a missing field | System suggests, person confirms |
| High risk | Large amount, new vendor, policy exception | Person decides; system only gathers facts |
2. Show the reason, not just the result
Every automated decision should say why: which fields matched, which rule fired. Staff trust what they can check in five seconds.
3. Make overriding easy and expected
The EU AI Act asks that people overseeing high-risk systems stay aware of automation bias, the habit of over-relying on a system's output, and can override it or stop it (EU, 2024). One click to reject, a short reason, no paperwork.
4. Measure trust directly
Track how often people override the system and why. A rising override rate means the automation is drifting. Zero overrides for months may mean nobody is really looking.
Is your workflow trustworthy?
Score each 0 (no), 1 (partly) or 2 (yes):
- Lanes: routine, uncertain and high-risk cases are separated.
- Reasons: every decision shows why it was made.
- Override: rejecting a decision takes one click.
- Audit: a sample of automatic decisions is reviewed each week.
- Owner: one named person is accountable for the workflow.
| Score | Next move |
|---|---|
| 8–10 | Widen the routine lane, one rule at a time. |
| 5–7 | Add reasons and one-click override before automating more. |
| 0–4 | Run in "suggest only" mode until staff agree with it. |
Your next step
Take one verification queue and sort last month's cases into the three lanes. The routine pile is your automation target; the high-risk pile stays with your people.
Sources
- MeitY (2025). India AI Governance Guidelines: Enabling Safe and Trusted AI Innovation. November 2025.
- RBI (2025). FREE-AI Committee Report. 13 August 2025.
- EU (2024). EU AI Act, Article 14: Human Oversight. Reproduced by artificialintelligenceact.eu.