How do you calculate Agentforce ROI?

Baseline your current fully loaded cost per resolved interaction and monthly volume, model platform costs from published pricing ($2 per resolved conversation on the Help Agent model, Flex Credits at roughly $0.10 per action, edition licensing), apply a conservative deflection rate to the routine share of your queue, and compare net monthly benefit against a fixed implementation cost. A worked example: 10,000 cases a month at $11 each with 40% deflection yields roughly $30,000 net monthly benefit after platform costs, on illustrative inputs you should replace with your own.

How much does Agentforce cost in 2026?

The published pieces: $2 per resolved conversation for the Help Agent lineage with no charge on escalation; Flex Credits at $500 per 100,000 credits, roughly $0.10 per action and more for voice; the Agentforce add-on at $125 per user per month; and the Core, Advanced, and Max editions at $195, $395, and $550 per user per month with 500,000 to 2.75 million bundled credits. Salesforce has not published which prebuilt agents ship with which editions, so confirm on a real quote.

What case deflection rate is realistic with Agentforce?

It depends on case mix and knowledge quality. Published results from our builds range from 48% of routine enquiries self-served to 92% autonomous resolution on a well-scoped warranty queue, and most organizations land between 30 and 50% within the first 90 days. Model conservatively against only the routine share of your queue, and define before the pilot the deflection threshold below which the case does not clear.

How long does Agentforce take to pay back?

On conservative mid-market inputs (10,000 monthly cases, $11 fully loaded cost, 40% steady-state deflection), the net benefit of roughly $30,000 a month pays back a fixed-price implementation within the first quarter after steady state. The honest caveat: payback is hostage to the deflection rate, which is why the pilot should be instrumented with a control group and a written kill threshold rather than launched on a vendor statistic.

Agentforce ROI: The Playbook for a Business Case a CFO Will Sign

The short version

Most Agentforce business cases fail at the baseline, not the math: if you cannot state today's cost per resolved interaction, no projection built on top of it survives a CFO. Gartner has predicted over 40% of agentic AI projects will be cancelled by the end of 2027, mostly on cost and value grounds.
The 2026 cost side is knowable: $2 per resolved conversation on the Help Agent model (no charge on escalation), Flex Credits at $500 per 100,000 (roughly $0.10 per action, more for voice), and the Core, Advanced, and Max editions at $195, $395, and $550 per user per month with bundled credit allowances.
The value side must be measured, not quoted: deflection benchmarks from our production builds run 48% to 92% depending on case mix, and most organizations land 30 to 50% within the first 90 days. Model YOUR number conservatively before you commit.
A pilot designed to prove ROI looks different from a pilot designed to launch: control group, instrumented session tracing from day one, and a sensitivity table instead of a single rosy projection.

Why Agentforce Business Cases Fail

The pattern is consistent: a team gets excited by a demo, quotes a vendor deflection statistic in a slide, buys licenses, and six months later finance asks what changed and nobody can answer with a number. Gartner has predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, largely on cost and risk-control grounds, and the cancellations will not be the projects that failed technically. They will be the ones that never defined what success cost or saved.

The fix is boring and reliable: a baseline you measured, a cost model built on published pricing, a value model built on conservative deflection, a payback calculation a spreadsheet can audit, and a pilot instrumented to prove or kill the case in 60 days. This playbook walks all five, with the real 2026 numbers. If you want the arithmetic done for you, our free Agentforce ROI calculator runs this exact model.

Step 1: Baseline What the Work Costs Today

Everything downstream depends on three numbers you should measure, not estimate. First, volume: interactions per month by type, from your case system, not from memory. Second, fully loaded cost per resolved interaction: team cost divided by resolutions, which for US support teams typically lands around $8 to $15 per interaction once salary, management, tools, and shrinkage are included (industry benchmarks commonly cite $11 and up; measure yours). Third, the shape of the queue: what share is routine and repetitive versus judgment work, because agents deflect the first category, not the second.

One honest hour with a RevOps analyst and your case reports produces all three. If you cannot produce them, that is the first finding: you are not ready to buy, you are ready to instrument.

Step 2: The Cost Side, With Real 2026 Pricing

Agentforce pricing is more knowable than the confusion suggests. The pieces, as published:

Pay per resolution: the Help Agent model charges $2 per resolved conversation, with no charge when the agent escalates to a human. This is the cleanest unit economics in the lineup, because you pay only for outcomes.
Flex Credits: other agent actions draw credits at $500 per 100,000, which works out to roughly $0.10 per action (about 20 credits), with voice actions costing more (about 30 credits). An agent averaging 10,000 actions a month runs about $1,000 a month in credits.
Editions: the Core, Advanced, and Max editions announced in September 2026 run $195, $395, and $550 per user per month and bundle Slack, Tableau Next, Agentforce Coworker, and 500,000 / 1 million / 2.75 million Flex Credits respectively. The standalone Agentforce add-on is $125 per user per month.
The gap to respect: Salesforce has published no official mapping of which job-ready agents come with which editions. Price a real quote; never price a screenshot of a pricing page.
Implementation: the line item teams forget. Scoping, data readiness, guardrails, and testing are real costs whether internal or external; we quote them fixed-price against a written scope so the model has a hard number instead of an hourly guess.

Step 3: The Value Side, Measured Not Quoted

Deflection is the headline variable, and the honest range is wide because case mix decides it. From our production builds: 48% of routine enquiries self-served for a senior-care network, and 92% autonomous resolution on a manufacturing warranty build handling more than 12,000 cases a month. Across organizations we see most land between 30 and 50% deflection within the first 90 days on a well-scoped queue. Model conservatively: take the share of your queue that is genuinely routine, assume the agent resolves half to two-thirds of it in year one, and let reality revise you upward.

Two other value lines deserve their own rows rather than padding the deflection number. Capacity: every deflected case returns handle-time to humans for the work that actually needs judgment, which is a staffing-plan number, not a soft benefit. And revenue-side agents (inbound qualification, outbound pipeline) run on different math entirely: pipeline contribution and conversion lift, which take longer to attribute honestly. Keep revenue agents out of a support ROI case; mixing them is how credibility dies in review.

Step 4: The Payback Math, Worked

An illustrative example with deliberately conservative inputs, every one of which you should replace with your own. A support team handles 10,000 cases a month at a measured $11 fully loaded cost per resolution. Scoped pilot queue: the 60% of volume that is routine. Assume 40% total deflection in steady state, which is 4,000 cases a month.

Gross saving: 4,000 cases at $11 is $44,000 a month. Platform cost against it: 4,000 resolutions at $2 is $8,000, plus licensing for the humans who remain (price your real edition mix), call it $3,000 to $5,000 a month attributable to the agent program. Net recurring benefit lands around $30,000 a month on these inputs. Against a fixed-price implementation in the low-to-mid five figures, payback arrives inside the first quarter after steady state. Run your own inputs through the ROI calculator, then stress them: at 25% deflection the case still clears; at 15% it does not, and knowing that threshold before you start is the entire point of the exercise.

Step 5: Design the Pilot to Prove It

A pilot designed to launch optimizes for going live. A pilot designed to prove ROI optimizes for a defensible number. The differences:

Scope one queue with honest routine volume, not the hardest cases and not a demo-friendly sliver.
Keep a control group. Route a comparable slice of volume to humans only, so the delta is attributable rather than seasonal.
Instrument before launch. Session tracing, containment rate, escalation reasons, and CSAT on both arms from day one. The governance stack you already own does this; turning it on is free.
Define the kill threshold in writing. The deflection rate below which you stop, agreed before anyone is emotionally invested.
Run 60 days minimum. Thirty days measures novelty; sixty measures a trend through at least one business cycle.

Step 6: Present It Like a CFO Thinks

Three habits make the difference in the room. Present a sensitivity table, not a point estimate: the case at 25%, 40%, and 55% deflection, so the conversation is about likelihood rather than belief. Separate vendor claims from your measurements, and label each; quoting Salesforce's own customer statistics as your projection is the fastest way to lose the room. And bring the risk register: data readiness, escalation quality, the agent-drift question, and who owns the number after go-live. A case that names its own risks reads as operated; a case that does not reads as sold.

If your org is new to the platform, start with our getting-started guide before this playbook; the readiness work there is what makes these numbers real. And if you want the model built against your actual queue data, that is literally what our scoping calls do: bring your volumes to a free 30-minute call at cal.com/cloudsheer-consulting/30min and we will hand back the spreadsheet, the threshold, and a fixed-price scope to test it. Our Agentforce practice runs the pilots too.

The Agentforce ROI Playbook: A Business Case a CFO Will Sign