Agentforce readiness: why the agent talks nonsense, and what to fix first
"We turned AI on and it talks nonsense" is now a sentence we hear every month. The agent is rarely the problem. Here is what is, how to find it in five days, and what a first agent that actually helps looks like.
Bohdan PyrihCEO, AdvirixIt is the data, not the prompt
We learned this on our own prompt tests at a wealth-management client. The assistant's answers about clients were weak, and the first instinct was to rewrite the prompts. The weak scores traced back instead to the org: a Life Event object configured and empty, opportunities created only after the client had already said yes, relationship data nobody maintained. The assistant was answering accurately from records that did not say much. Better prompts would have made it confidently wrong faster.
That pattern repeats everywhere. Agentforce and the Einstein features ground their answers in your records and your knowledge articles. If the fields are empty, the articles are stale and the duplicates are unresolved, the agent inherits all of it and adds fluency.
Five things to check before switching anything on
- Empty fields on the objects the agent will read. Pull the fill rate for the twenty fields a human would look at to answer the same question. Under half, and the agent has nothing to say.
- Stale knowledge. Knowledge articles last updated before the last product change. The agent will quote them.
- Duplicates. Two records for one customer means the agent picks one, and it is not always the right one.
- Permissions. An agent sees what its running user sees. If that user has more access than a support rep, so does the agent, and so does whoever is chatting with it.
- Who owns the answers. If no person owns the knowledge base, nobody will fix the first wrong answer, and there will be a first wrong answer.
This is the AI readiness section of our Health Check: which objects have gaps, which knowledge is stale, what to fix before an agent is allowed to answer, and what not to switch on yet. It takes five days and read-only access.
What a first agent should look like
Narrow, measured and reversible. One channel: web chat, a customer portal, Slack or email. Three topics: knowledge questions, record lookups, and a clean handoff to a person with the conversation attached. Tone and greetings set, up to ten knowledge articles prepared and checked, guardrails tested before go-live against the questions your team actually gets. Analytics on from day one. Two weeks of work, a defined group of users, and a date thirty days later to count.
What "guardrails" mean in practice
The agent answers only from the knowledge and the records you point it at. It says "I do not know" instead of guessing. It hands off to a person on the topics you define: pricing exceptions, complaints, anything with a number in it, anything legal. The limits are tested before launch with a script, and the script is kept so the next change can be tested the same way.
Thirty days later: three numbers
How many conversations the agent closed on its own. How many it handed off. How many times it was wrong, and about what. Those three numbers decide the next step: more topics, more channels, or stop. An agent that closes a third of the repeat questions and hands the rest to a person with context is a success. One that answers everything is a liability waiting for a screenshot.
When not to switch it on
When the data checks above fail, and when nobody owns the answers. We have said no to AI work on that basis. At GoodLeap, a lender with several orgs, Service Cloud Voice and Agentforce were discussed and scoped, and not contracted: the integration and release work had to come first. The case study says so, because the sentence is more useful than a logo.
The money in your own base
The other AI conversation is quieter and usually worth more. Hundreds of clients, slow growth, and an expansion pipeline that does not exist in the CRM at all. The base already says who should be offered what next: what similar clients have that this one does not, which signals have fired, which relationship has gone quiet. At the wealth firm above, counting what the base already held found $40,100 a year of recoverable revenue and $53,000 a year of manual work. No model guessed that. The data said it, once someone asked.
A readiness scorecard you can run yourself
| Check | How to measure | Ready when |
|---|---|---|
| Field fill rate | The twenty fields a human reads to answer the question; % filled on records from the last year | above 80% |
| Knowledge freshness | Share of articles updated since the last product or policy change | above 90% |
| Duplicates | Accounts and contacts matching by email, name and domain | under 2% |
| Running-user access | Objects and fields the agent's user can read, compared with a support rep | equal or narrower |
| Answer ownership | A named person who reviews wrong answers weekly | named |
Three of five failing means the agent will be fluent and wrong. The Health Check runs these in five days and prices what it takes to pass each one.
The three steps, each priced on its own, are on the Agentforce and AI page. The readiness check is part of every Health Check.