The Agent-Readiness Scorecard — 8 conditions before an agent is worth buying
An AI agent acts on what it reads, so most of what decides whether it works sits upstream of the agent. This is the checklist version of our data-foundation series: eight conditions, each linking to the piece that works it through. Score one job, not your whole company. The ones you fail are the work.
How to score this
Pick ONE job you would hand an agent first — not your business as a whole. Booking requests, say, or first-pass triage on inbound. Give yourself a point per condition only where you could show someone the evidence today. 7-8: the remaining questions are for the vendor. 4-6: it will work on the easy path and surprise you off it, so start somewhere narrow and reversible. 0-3: pick a smaller job, or fix these first. Scoring one job at a time is deliberate. Nobody has all of this everywhere, and waiting for that is how people never start.
Why “ready” never meant a spotless company-wide database · The five questions to ask the vendor
1. You can name the system that decides who a customer is
This asks for a winner rather than tidy records. When two systems disagree about who someone is, there should be one you would point at and say “that one wins”. Agents do not ask; they pick. If you have never named the winner, the agent names it for you, differently each time, and never mentions it.
2. The number it would act on has been checked against its twin this month
Most people can tell you their systems agree. Far fewer checked recently. This condition is about the checking, not the agreeing: when did a human last open both and compare? An agent inherits whatever drift accumulated since.
3. You would notice within a day if a feed stopped
A source that quietly stopped updating still answers questions, using facts from whenever it stopped. Nothing errors, so nothing alerts. Score this on how you would find out, not on whether it has happened.
4. Somebody would be annoyed if this job were done wrong
A specific person, by name, who would notice and care. Not a team, not a rota. This sounds soft and it is not: work with an owner gets checked, and work without one gets discovered at year-end.
5. You know which of your numbers answer different questions
Revenue on the invoice and revenue in the bank are both right and rarely equal. Wherever two of your numbers are measuring subtly different things, you need to know it before something automated picks one. Score a point if you could list those pairs.
6. You have said out loud which of its actions you cannot undo
Tagging a lead wrong costs a minute. Emailing a customer, quoting a price or cancelling tomorrow leaves the building. Plenty of good agent work is irreversible, so this asks for clarity rather than timidity: you named which is which before granting anything.
7. Something already tells you what happened, without being asked
Before an agent exists. If your current answer to “what did the system do last week” is somebody opening a tool and looking, that will not scale to something making its own decisions. Score a point if a report already arrives.
8. The second-most-technical person could explain the flow
Go and ask them. Not the person who built it. If their explanation has a gap in the middle, that gap is what you will own the day the builder leaves. Their answer is your real score for this one.
Failed a few? That list is the audit.
The conditions you missed are the data-foundation work, in priority order, for the one job you scored. Tell us which came back no and we will tell you which to fix first — and which you can leave alone for now, because some of them genuinely do not matter yet.