Agentforce ROI: how to work out payback on Flex Credits, and when the number stops working

Short answer: Agentforce ROI is the labor time an agent removes, priced at what that time costs you, minus what Salesforce meters for the work. At the US median executive assistant wage of $76,590 a year, which is $36.82 an hour, a single $0.10 Agentforce action pays for itself if it saves a little under 10 seconds of that person's time. That ratio is so lopsided that cost is almost never what kills an Agentforce business case. What kills it is that saved minutes stay on the payroll unless somebody redeploys them.

Last updated August 2026. Rates read off salesforce.com/agentforce/pricing on August 21, 2026. Wage figure from the Bureau of Labor Statistics Occupational Employment and Wage Statistics, May 2025. We sell a competing product, so check the arithmetic rather than trusting it.

How do you calculate Agentforce ROI?

Three numbers, in this order. What the work costs you in labor today, what Salesforce will meter for the same work, and how much of the freed time you can actually convert into something else. The first two are arithmetic. The third is a management decision, and it is where most ROI cases quietly fall apart.

Start with the cost side, because it is the easy half. Salesforce sells Flex Credits at $500 per 100,000, so a credit is half a cent. One Agentforce action is 20 credits, which is $0.10, and an Agentforce Voice action is 30 credits, which is $0.15. Multiply your expected actions by $0.10 and you have the meter. The Agentforce pricing calculator runs the same sum against your headcount if you would rather move sliders than open a spreadsheet.

Then the labor side. Take the fully loaded hourly cost of whoever does the work now. The BLS median for executive assistants was $76,590 a year in May 2025, which across 2,080 hours is $36.82 an hour before benefits and overhead. Use your own numbers if you have them, and if you do not, defensible pay bands are worth building properly rather than guessing at, because every downstream figure inherits that error. Multiply the hours the agent removes by that rate and you have the value side.

What is a realistic payback period for Agentforce?

On the metered model, payback is usually measured in days rather than months, because the unit economics are extreme. One action costs $0.10 and one hour of assistant time costs $36.82. The action only has to save about 10 seconds to break even. Even if you assume the agent is wrong a third of the time and someone has to redo the work, you are still an order of magnitude ahead on the meter alone.

The per-user models change the shape of the question, because you are buying a fixed monthly figure rather than a variable one, and it has to be earned back whether the agent is used or not. Here is what each Salesforce buying model has to save, expressed as time rather than money.

Read the right-hand column, because it is the honest version of the ROI question. The $125 unmetered add-on needs to save each licensed user roughly 10 minutes a working day. That is a low bar for anyone drowning in case notes and a high bar for a user who logs in twice a week. An Agentforce 1 Edition at $550 needs to save about 45 minutes a day, every day, from every seat you buy. That is a real productivity claim, not a rounding error, and it should be tested on a pilot group before it is bought for a department.

Is Agentforce worth it?

It is worth it when the work already lives in Salesforce and the volume is high enough that per-action pricing stays cheap. Case summarization, record updates, appointment scheduling and first-line product questions all fit that description. Salesforce's own worked example puts 100 service users handling three cases a day at $1,800 a month on Flex Credits, which is trivial against the payroll of 100 service users.

It is not worth it when the agent is bought as a platform rather than for a job. The pattern to avoid is licensing a department, discovering that only a fraction of them use it, and paying the per-user rate on all of them anyway. The metered model exists precisely so you do not have to make that bet, and buyers who start on Flex Credits and switch later have far better information than buyers who start on the add-on.

What does an Agentforce ROI calculator leave out?

Four things, consistently. The first is the Salesforce license underneath, since Agentforce is priced on top of an existing edition and none of the credit arithmetic includes it. The second is Data Cloud and Data 360 consumption, which Salesforce excludes by name on every one of its own published examples. The third is implementation: connecting agents to real systems of record, writing the actions and testing them against edge cases is project work, and for mid-market buyers it usually exceeds the first year of metered usage.

The fourth is the one nobody models. Salesforce notes that credit consumption carries tiered multipliers by model type, so a premium reasoning model does not draw the same credits as a basic one. Any figure you produce, ours included, is a floor. Treat it as a way to compare buying models against each other rather than as a forecast of your invoice, and read the full rate card on our Agentforce pricing breakdown before you commit to a model.

When the ROI number stops working

Here is the part that gets skipped. Time saved is not money saved. If an agent removes 40 minutes a day from five people's work and nothing else changes, you have not saved $36.82 an hour five times over. You have five slightly less busy people and an identical payroll. The saving becomes real in exactly two ways: you redeploy those hours onto revenue work you were not doing, or you do not make a hire you would otherwise have made.

So write down which one you are claiming, before you buy. If it is redeployment, name the work the freed hours will go to and who will check that it happened. If it is avoided hiring, name the role. A business case that says "improved productivity" without either of those is not a business case, and it is the reason so many AI pilots show excellent unit economics and no measurable effect on the P&L twelve months later.

The second failure mode is quieter: the accuracy tax. An agent that gets 90 percent of records right and 10 percent wrong can cost more than doing nothing, because somebody now has to check all 100. Whatever hours-saved figure you use, discount it by the review time you will actually spend, and put a human approval step in front of anything that writes to a system of record or leaves your domain.

How to measure Agentforce ROI after you buy

Pick one process, baseline it for two weeks before the agent touches it, and measure the same thing after. Count the actions in Digital Wallet, which reports Flex Credit consumption in near real time, and put that against the process time you measured. Two weeks of real data beats any model, including this one.

Watch the meter monthly rather than quarterly. Salesforce states there is no overage penalty and bills excess in arrears, which is friendlier than Microsoft disabling custom agents once capacity is exhausted, but it also means nothing stops consumption from drifting upward. Unused Flex Credits do not roll over either, so a large prepaid balance is its own kind of waste.

Finally, re-run the comparison once you have a real usage number. The crossover between metered credits and the $125 unmetered add-on sits at about 62.5 actions per user per day at 20 working days a month. Most employee-facing deployments never get near that, which is why the metered model usually stays the cheaper answer long after people assume they have outgrown it.

Where a flat-rate agent changes the arithmetic

Every model above multiplies by headcount, so the ROI question is really a per-seat question. That is the right shape when the work is spread evenly across a large team. It is the wrong shape when a handful of recurring jobs, research, lead lists, outreach drafts, inbox triage, data entry and weekly reports, are what you actually want done.

WorkAgent is $149 a month flat for the whole account, with no credit meter and no per-seat multiplication, which in the table above is 4.05 hours of assistant time across the entire business rather than per person. We built this page, so weigh that accordingly, and note the honest boundary: Agentforce lives inside Salesforce and acts on Salesforce records. If that is your system of record and your agents need to write to it, Agentforce is doing something a general-purpose agent does not do. If the work is the repetitive business admin around it, the flat model is the cheaper shape and a far easier one to forecast.

Try it on a real task in the panel at the top of this page, then check the numbers yourself in the Agentforce pricing calculator. If Microsoft is also in your evaluation, the credit arithmetic runs the other way and is worth seeing side by side on Copilot Studio pricing and Agentforce vs Copilot Studio.