Proactive suggestions are unusual among software features in that being right is not sufficient. A notification interrupts someone, and interruption has a cost that is paid whether or not the suggestion is useful. Get the ratio wrong and people stop reading — after which accuracy is irrelevant, because nobody is looking.
This guide is a rollout sequence for AI nudges that keeps that from happening.
1. Start with signals people already wish they had
The temptation is to launch the clever ones. Start instead with findings your team already tries to produce manually and cannot do reliably at scale — a purchase-order line that duplicates stock already inbound, an order sitting just below a better scheme slab, an invoice that was delivered but never raised, a customer who has gone quiet against their own pattern.
These share a property: the recipient immediately understands why it matters. That is what earns the right to interrupt them again tomorrow.
2. Require evidence on every signal
A suggestion without its working is a request to take something on faith, and faith depletes.
Compare: "This line may not be needed." Versus: "Stock pipeline already covers this — on hand 40, in transit 30, on order 25, committed 40, net 95, reorder point 65." The first invites a shrug. The second can be checked in ten seconds and either accepted or challenged on a specific assumption.
Insist on this before launch. It is much harder to add credibility later than to ship with it.
3. Attach an action, or do not send it
A signal that tells someone about a problem and leaves them to fix it elsewhere has moved work, not removed it. Every nudge should carry the correction — apply the suggested quantity, open the record, raise the invoice — so acting is one step.
This also disciplines the design. If you cannot define the action, the signal is probably an observation rather than a suggestion, and it belongs on a dashboard.
4. Route by role, not by broadcast
The fastest way to kill adoption is sending everything to everyone. A buyer wants stock signals; a credit controller wants receivables; a sales manager wants coverage and dormancy. Cross-delivery trains people that most of what arrives is not theirs, and they filter accordingly.
Scope by role and by record, and be willing to suppress a correct signal from someone who cannot act on it.
5. Launch narrow and expand on evidence
Pick two or three signals and one team. Run for a month. Then look at the numbers before adding anything.
Expanding on enthusiasm rather than evidence is how organisations end up with forty signal types and a workforce that has learned to dismiss the lot.
6. The metric that matters
Track the action rate: of the nudges shown, what fraction were acted on rather than dismissed or ignored?
- Above roughly 30% — the signal is earning its place.
- 10–30% — probably too broadly targeted. Tighten the threshold or the audience before concluding the signal is wrong.
- Below 10% — turn it off. A signal at this rate is training people to ignore the channel, which damages the signals that work.
Turning one off is a healthy act, not an admission of failure. The value of the channel is a function of its average usefulness, not the count of signal types.
7. Deal with the trust question directly
Two questions arrive early, and answering them plainly saves months.
"Can it see things it shouldn't?" If your assistant runs under each user's own permissions, say so and demonstrate it — show a user asking for something outside their scope and getting nothing.
"Can it change things on its own?" Be precise about the write surface. "It proposes; you approve; the change is audited" is a sentence people can evaluate. Vagueness here is read, correctly, as risk.
What this looks like in xMatix
Nine signal providers ship today across inventory, schemes, receivables, sales and service — purchase-order quantity, scheme gap, alternate item, overdue receivables, awaiting invoice, dormant customer, quote follow-up, service recommendation and stale draft. Each is computed deterministically and carries structured evidence and a one-click action.
Sense Assist delivers them in context, explains them in plain language on request, answers follow-ups against live data, and carries the action under the user's own permissions with the change written to the platform's own audit trail. Tenants can also author agents that publish their own nudges, so a pattern specific to your business can be encoded without waiting for a vendor.
Related: Deterministic signals, AI delivery · The nudge that says don't buy · Rolling out an AI assistant your CISO will sign off