Salesforce Agentforce Costs and ROI: How to Budget and Measure It

Platform and usage

The direct cost depends on how Salesforce bills your contract: by actions or conversations, by credits or through user licences that include agent features. Usage-based pricing means cost scales with volume. A service agent that handles thousands of conversations a month costs much more than a sales agent that drafts a few hundred emails.

Data preparation

Agents are only as good as the records and knowledge articles they read. Many organisations find duplicate accounts, outdated knowledge bases and inconsistent fields. Cleaning that up is often the largest one-off cost, and it is frequently left out of the first budget.

Build and configuration

Someone has to define topics, write instructions, connect actions to Flows or Apex and set guardrails. Agent Builder lowers the technical barrier, but complex processes still need an admin or developer who knows the org.

Supervision

After launch, people review transcripts, handle escalations and update instructions. This is a permanent cost, not a launch cost.

• Licences or usage charges, estimated from expected monthly volume
• Data Cloud costs if the use case needs unified customer data
• Knowledge base clean-up and rewriting
• Admin, developer or partner hours for build and testing
• Ongoing review time, usually a set number of hours per week
• Change management and training for the human team

Start with the process you plan to automate and pull real numbers from the last six months. For service, count cases by type and channel. For sales, count inbound leads by source. Then decide which share an agent could realistically handle.

Be conservative. Simple, repetitive requests such as order status or password resets are good candidates. Complaints, refunds above a threshold and complex technical issues should route to people. For example, if 40 percent of cases are order status questions, model a scenario where the agent resolves half of those at first, not all of them.

• Containment or resolution rate: share of conversations finished without a person
• Escalation quality: did the agent pass full context when it handed over?
• Average handling time for cases that still reach people
• Customer satisfaction on agent-handled conversations versus human-handled
• Lead response time and the share of agent-qualified leads accepted by sales
• Error rate: wrong answers, wrong record updates or policy breaches found in review
• Cost per resolved case, including usage and supervision

Resolution rate alone can mislead. An agent that closes conversations by frustrating customers into giving up looks efficient and damages the brand. Pair it with satisfaction scores and a sample of transcripts read by a person each week.

A simple model is enough for a first decision. Estimate the monthly volume the agent will handle, multiply by the cost of handling each item manually, and subtract agent costs and supervision. Then run the same model with pessimistic assumptions.

• Baseline: current cost per case or per lead, using loaded staff cost
• Scenario A: conservative resolution rate, full supervision hours
• Scenario B: expected resolution rate after three months of tuning
• Payback: months until one-off costs are covered under Scenario A
• Stop rule: the metric level below which the pilot ends

If the project only works under the optimistic scenario, it probably needs a different use case.

The financial model assumes the agent behaves as tested. In production, three risks can push costs above plan. Each needs a control written into the project from the start, not added after the first incident.

• Wrong answers: an agent that quotes an outdated policy creates refunds, complaints and rework
• Runaway usage: a loop or a spike in traffic multiplies usage charges within days
• Silent drift: knowledge articles change, but nobody updates the agent’s instructions

Controls are simple to describe and require discipline to keep. Set usage alerts in your Salesforce billing view, review a fixed sample of transcripts every week, and make knowledge article owners responsible for flagging changes that affect the agent. Add a clear route for customers to reach a person at any point in the conversation.

Include the cost of those controls in the ROI model. A business case that leaves out review hours will look better on paper and worse in the quarterly review, which is where trust in the project is won or lost.