An AI agent is software that takes an action on your behalf rather than handing you a draft and waiting.

That distinction is the whole thing, and it gets buried under a great deal of language. A chatbot answers. An assistant drafts. An agent is given a goal, a set of tools, and permission to use them, and then it does something: updates a record, sends a reply, routes a case, books a meeting. Everything else in the category is a variation on how much rope it gets.
Dreamforce spent a week in September pitching this to rooms full of people. Here is the plain version, without the pitch.
What It Actually Does in a CRM
Strip away the demos and there are roughly four jobs agents are being sold for right now.
Deflection. A customer asks a question, and the agent answers it from your knowledge base without a person touching it. This is the most mature use and the easiest to evaluate, because you can measure it: what percentage of cases closed without a human, and did satisfaction hold.
Data work. The agent reads an email thread and updates the opportunity, logs the activity, or fills the fields a rep would have filled. Genuinely useful, unglamorous, and the thing most likely to pay for itself in a small org.
Orchestration. The agent decides what happens next in a process: routing, escalating, triggering the follow up. This is where most of the impressive demos live and where caution starts to matter, because the failure mode here is an action taken on a real customer rather than an answer someone can correct.
Configuration and administration. The newest pitch, and the one we would be most careful with. You describe a change in plain English (add a renewal date field and make it required, build a flow that emails the account owner when a case escalates) and the agent builds it in your org. Of the four, this is the most dangerous use by a wide margin, and the danger grows with how little Salesforce experience the person asking has.
An agent that answers a customer wrong produces one bad answer. An agent that changes your configuration changes the rules every later record, report, automation and integration runs on. It will do exactly what you asked, and nothing you didn’t think to ask. Make that field required and it has no reason to mention that the integration loading records overnight never fills it in, so those loads start failing while everyone is asleep. Ask for the flow and you’ll get one, built without regard for the two other flows and the validation rule already firing on the same object. Ask it to give the sales team access and it will, whether or not that access was the right size.
Experienced administrators and architects ask those questions before a change because they have been burned by not asking them. An agent has never been burned. With someone who knows your org directing it, reviewing every change, and trying it in a sandbox before production, it can save real hours. Standing in for that person, it is a fast way to build the kind of technical debt somebody eventually gets paid to unwind.
Where the Demo Hides the Work
The demo is running against clean data, a well maintained knowledge base, and a process somebody wrote down. Your environment is probably missing at least one of those.
That is the single most important thing to understand before you evaluate any of this. An agent is only as good as the material it acts on. Point one at a knowledge base that was last curated three years ago and it will confidently give customers answers that were true three years ago. Point one at the opportunity stage field we keep writing about, the one where three reps mean three different things, and it will make routing decisions on a definition that does not exist.
The organizations getting value from agents right now are, almost without exception, the ones who had their data and process in order first. The agent did not fix anything. It amplified something that was already working. The same is true of the org itself: an agent making changes to your Salesforce setup is only as safe as the person who understands what is already there.
This is not an argument against adopting it. It is an argument about sequence, and the sequence is not negotiable.
What It Costs
The figures below are Salesforce’s list prices plus ballpark numbers from the Agentforce work we’re doing with organizations right now. Your quote will depend on your edition, your contract and how the negotiation goes, so treat them as a starting point for that conversation, not a price sheet.
There are two ways to pay, and choosing between them matters more than the rate.
Pay as you go, with Flex Credits. You buy credits and the agent spends them as it works. A credit runs about half a cent, so 100,000 of them cost roughly $500. A standard agent action uses 20 credits, about a dime. A customer conversation the agent resolves on its own is priced at about $2. In our current work, a useful employee facing task (an agent reading a set of records and drafting something a person would otherwise write) runs around 40 credits a pass, or about 20 cents. Two working proofs of concept used about 3,000 credits between them. That’s roughly $15.
Per user, unmetered. Agentforce add-ons for Sales, Service and the other clouds start at about $125 per user per month, and the bundled Agentforce 1 editions start around $550. At $125, the license pays for itself once a person would otherwise burn about 25,000 credits a month. For that 40 credit task, it’s around 600 runs a month, or about 30 every working day. Below that line, credits are cheaper. Above it, the license wins.
For most small organizations the sensible start is credits. A few hundred dollars covers a lot of real testing, and you move a person onto the license only once their own usage proves it.
The implementation is the bigger number. In our current work, a well scoped agent got about 80 percent of the way there in roughly five hours of build time. The last 20 percent took real users, real records, and several rounds of feedback on what the output should look and sound like, and that is where most of the hours go. Before any of it, the data and the knowledge base the agent reads have to be in shape. If they aren’t, that cleanup comes first, and it usually costs more than everything above.
And if an agent will be changing your configuration, budget for an experienced administrator reviewing every change. Leaving that line out is how the cheapest option on this page becomes the most expensive one.
Where it stops paying off. If the task you’re handing over happens a few times a week, even half a cent a credit buys a capability you don’t need yet. The math starts working when the same task happens dozens of times a day across your team.
The Questions Worth Asking Before You Buy
Whichever way you pay, these are the questions that determine whether it works.
- What specific task are you handing over, and how often does it happen? If you cannot name the task and count the volume, you are buying a capability rather than solving a problem, and capabilities do not pay for themselves.
- What does the agent read to do its job, and who maintains it? Name the person. A knowledge base with no owner degrades, and a degrading knowledge base makes an agent worse over time rather than better.
- What is the worst thing it can do unsupervised? Answer a question wrong is recoverable. Email a customer the wrong contract terms is not. Set the permission boundary at what you can afford to have go wrong at three in the morning. If the agent can change your configuration, the honest answer is almost anything.
- How will you know if it is working? Decide the measure before you turn it on. Cases deflected, hours returned, response time. Without a baseline you will be relying on the vendor’s dashboard to tell you whether the vendor’s product is working.
What We Would Tell a Twenty Person Business
Wait on orchestration. Look hard at data work. Keep agents out of your setup menu.
The agentic capability that most likely pays off at small scale is the boring one: automatic activity logging, record updates from email, the administrative overhead that quietly consumes several hours a week per person. It is low risk, easy to measure, and does not require you to trust software with a customer relationship.
The customer facing agent is a bigger decision than it looks, and it should follow a knowledge base someone owns, not precede it.
The admin agent is the one to hold back hardest. Unless someone with real experience administering and architecting Salesforce is directing it and reviewing what it does, don’t let it change your org. Every change it makes becomes part of the foundation the other three uses run on.
And be honest about the threshold. Below a certain volume of repetitive work, the correct answer is that this is not for you yet, and any vendor unwilling to say that out loud is telling you something. The same filter applies here as to every technology wave, which we set out in AI strategy without the hype: start from the business problem, check the data foundation, then look at tools.
For the practical version of this as it develops, the Iron Insights newsletter is where we publish it first.
Ready to work out whether agentic AI does anything for an organization your size? Let’s talk about the problem before the product, starting with an honest read of what your data and processes will support through AI enablement, and what your Salesforce foundation would need first. Schedule a consultation today!
Responses