Latest · Experimental · Published 11 Sep 2026 · GA Applications editorial

Automate one thing. Watch it for a week. Then decide.

Small businesses should treat CRM automation as an experiment, not an installation. Choose one automation that is reversible, observable and low-consequence — a follow-up reminder, an enquiry acknowledgement — run it for seven days with a written hypothesis and a named owner, and keep or kill it on what you observe. Scale only what survives the log.

The experiment format is a GA proposal, published to be run and judged. The accountability argument cites Harvard Business School, 8 April 2026.

Recognition

Why automation projects fail small businesses

Rarely the tools; almost always the sequence. A business automates ten things at once, nobody owns any of them, and when a wrong message reaches a customer nobody can say which rule sent it. The automation gets blamed, then abandoned, and the manual chaos returns with a subscription fee attached.

Harvard Business School’s Working Knowledge put the accountability rule bluntly on 8 April 2026: “AI made me do it” is no excuse. The same holds for every workflow rule and scheduled send — a named person must be able to answer for each automated act, or the automation is not ready to run.

The alternative is boring and effective: one automation, one owner, one week, one written log. This article publishes the format GA proposes. It is an experiment design — results belong to whoever runs it.

Selection criteria

Choosing the first automation

Five filters. A candidate that fails any one of them is a later automation, not a first one.

  1. Reversible

    If it misbehaves you can switch it off and undo the damage in minutes. Start with internal notifications, not customer-facing sends.

  2. Observable

    You can see every action it takes: a log, a copy of each message, a list you can check. Invisible automation is unaudited automation.

  3. Low consequence

    A mistake costs embarrassment, not money or safety. Invoicing and customer promises come later, if ever.

  4. Frequent enough to observe

    Something that fires several times a week. You cannot learn from an automation that runs twice a year.

  5. Named owner

    One person answers for it: checks the log, holds the kill switch, reports at week’s end.

Fill it in before touching the tool

The first-automation experiment card

One card per automation. If the card cannot be filled, the automation is not ready to exist.

FieldWhat you write before switching anything on
HypothesisOne falsifiable sentence: “If enquiries get an acknowledgement within five minutes, fewer prospects go cold.”
TriggerThe exact event: form submitted, deal stage unchanged for three days, quote sent.
ActionExactly what the system does, including the message text. No “and maybe” clauses.
OwnerThe named human who answers for every automated act.
RollbackHow it is switched off, and who can do it, in one line.
Success measureWhat you will count: response times, follow-ups completed, complaints.
Stop conditionThe observed result that ends the experiment immediately: one wrong recipient, one confused customer.

Copy this table into a document and treat it as the automation’s licence. No card, no automation.

What to watch, day by day

The seven-day observation log

Seven days is long enough for a frequent trigger and short enough that a bad automation cannot do much harm.

DayWhat to checkWhat a problem looks like
Day 1Every action the automation took; read each oneAnything sent that you would not have written yourself
Day 2Timing: did actions fire when expectedDuplicates, delays, silent skips
Day 3Recipient reactions: replies, confusion, complaintsA customer asking “what is this?”
Day 4Edge cases: missing fields, unusual names, after-hoursGibberish merge fields, 2am sends
Day 5Owner workload: did it shrink or growManual fixes outnumbering automated saves
Day 6The count against your success measureMovement in the wrong direction, or none at all
Day 7Decision: keep, adjust or kill — written on the card“Keep watching” with no decision date

Extend the experiment only with a new card and a new hypothesis. “Leave it running” is how unowned automation is born.

The accountability rule

When an automated message offends or confuses a customer, “the system did it” is not an answer — it is an admission. HBS’s warning about delegating blame to AI applies to every rule and workflow: a named person must be able to explain and own each automated act, or the automation does not run. The CRM itself is only the memory; how that memory is governed is the system.

Evidence honesty

Sources, method and what would change this

  • Verified fact

    Accountability for automated acts

    Harvard Business School’s analysis holds that blaming AI for harmful outputs is no excuse; accountability stays with the people who deployed the system.

    Harvard Business School Working Knowledge, 8 April 2026

  • Proposal

    Experiment format

    The card and the seven-day log are GA’s proposed minimum discipline for a first automation. They are a design, published to be run and judged — no outcome data is claimed.

    GA Applications method

  • Inference

    Failure pattern

    The claim that small-business automation fails through missing ownership rather than tooling is GA practitioner inference, not survey data.

    GA Applications editorial

  • Proposal

    What would change this conclusion

    If owners running the card report that the format blocks useful automation rather than catching bad automation, the format will be revised and the change dated here.

    GA Applications editorial

Bring one task you are sick of repeating

Describe the repeated task. GA replies with whether it is automatable, the experiment card filled in as a starting draft — and the honest answer if a better spreadsheet would do.

The first automation is often worth less than the discipline it installs.