Conversational AI for Customer Service: Costs, ROI and How to Budget

Vendors package their prices differently, which makes comparison hard. Break every quote into the same parts before you compare.

• Platform or seat fees: a monthly charge for the help desk or AI module, such as Zendesk, Intercom, Freshdesk or HubSpot Service Hub.
• Usage charges: per conversation, per message or per resolved ticket. Intercom’s Fin, for example, is priced per resolution at the time of writing.
• Model costs: if you build your own on OpenAI, Anthropic or Google APIs, you pay per token processed.
• Channel fees: WhatsApp Business Platform charges per conversation by category and country.
• Setup: writing the knowledge base, connecting order and CRM systems, testing.
• Ongoing tuning: reviewing conversations, updating content, adjusting rules each month.

The last two lines are often left out of business cases. They are where most of the real effort sits.

Take the total monthly cost of your support team, including salaries, benefits, software and a share of management time. Divide it by the number of tickets handled in that month. That gives a fully loaded cost per ticket, which is the number every AI option must beat.

Then sort tickets by type. Most help desks let you tag or categorise them. You will usually find that a small number of question types, such as order status, return policy and password resets, make up a large share of volume. Those are the tickets an AI assistant can realistically resolve.

Resolution rate

The share of conversations the AI closes without a human, where the customer does not come back with the same issue within a set window, often 48 or 72 hours. A conversation that ends because the customer gave up is not a resolution, so check repeat contacts.

Cost per resolved conversation

Add all AI costs for the month and divide by AI-resolved conversations. Compare this directly with your human cost per ticket for the same ticket types, not the overall average.

Customer satisfaction and handoff quality

Track CSAT for AI conversations separately. Also watch what happens after handoff: if agents must ask the customer to repeat everything, the AI is adding friction rather than saving time.

Revenue effects

For e-commerce, measure whether pre-sale questions answered instantly lead to orders. Tag conversations that start on product pages and follow them to checkout in your analytics.

Build a twelve-month budget with three phases, because costs and results are uneven at the start.

• Months 1-2, setup: higher spend on content, integrations and testing; low resolution rate, often in suggestion mode only.
• Months 3-6, ramp-up: usage costs rise as the AI handles more; resolution rate improves with weekly tuning.
• Months 7-12, steady state: usage tracks ticket volume; tuning drops to a few hours a month.
• Contingency: add a margin for seasonal peaks, since usage pricing grows with volume.
• Exit cost: estimate what it would take to move platforms, including exporting conversation logs.

For example, a mid-sized online store might model three scenarios for resolution rate, low, expected and high, and see at which point the AI pays for itself. If the business case only works in the high scenario, start smaller.

Poor knowledge content is the largest. If your help articles are outdated or contradict each other, the AI will repeat the errors, and fixing that content becomes an unplanned project. Integration work is next: answering “where is my order” needs live access to the order system, which may require developer time.

Compliance adds cost too. Under KVKK and GDPR you need to inform customers they are talking to an automated system, define retention periods for transcripts and review data processing agreements with each vendor.

Buying a help desk with built-in AI is faster and includes reporting, handoff and channel connections out of the box. Building on a model API gives more control over prompts, data and cost per conversation, but you also own hosting, monitoring, security and every integration.

For most small and mid-sized teams, buying is the sensible first step. Building makes sense when ticket volume is high enough that per-resolution pricing becomes expensive, or when you need deep integration with in-house systems no vendor supports. Even then, many companies start with a platform, learn what works and build later with real data in hand.

Whichever route you pick, ask for conversation logs in an export format you can read. Those logs are the raw material for the next decision, and losing them when you switch vendors means starting the analysis again.