Budget for AI support add-ons

AI support costs are easy to understate because the visible product price may cover only one part of the workflow. A help desk might include reply drafting but charge separately for an autonomous agent. One product counts sessions, another counts credits, and another bills for a defined resolution. Those units are not comparable without a realistic sample of your own conversations.

Start with the customer job. If an AI assistant answers routine delivery questions accurately, that may save time. If it generates incorrect refund promises that agents later repair, a low per-conversation price is irrelevant. The AI software comparison covers product fit; this guide focuses on the budget and the conditions for using the feature responsibly.

Build a four-line AI budget

List separately: the base help desk plan, paid AI seats or add-ons, variable usage, and implementation/oversight. Usage should have a named unit and definition. Ask whether one customer issue can create several sessions, whether an escalated interaction is charged, and whether a customer reopening the issue triggers another billable event. Write the answer into the cost model.

Gorgias publishes billing rules for its help desk and AI services. Tidio separates customer-service, Lyro AI and Flows allowances on its pricing page. HelpDesk and Freshdesk publish their own AI entitlements on pricing and product pricing pages. These are reasons to compare units, not a claim that one provider has a permanently lower AI bill.

Model three scenarios

Use a normal month, a peak month and a poor-quality month. In the poor-quality case, more conversations escalate, repeat or need human correction. Calculate vendor charges and agent time in each scenario. A simple working formula is AI subscription + billable events × unit price + review and rework time. If the vendor sells bundles, apply bundle boundaries and unused credits rather than assuming perfectly smooth usage.

Record any cap, alert or hard stop. Some teams prefer to pause AI when a budget threshold is reached; others need predictable service during a peak. Decide the fallback before switching it on. A usage cap that leaves customers without a human route is not a safe cost control.

Keep the assumptions visible beside the calculation. Note what share of incoming questions is eligible for automation, the average billable units per eligible issue, the proportion that transfers to an agent and the time needed to check a suggested reply. If these figures come from a trial with a small sample, show a range rather than a precise savings promise. Recalculate when the provider changes a billing definition or the business adds a new channel.

Test whether the result is worth buying

Pilot one low-risk topic with known answer sources. Sample AI answers for accuracy, current policy, tone and privacy. Count how many customers finish the task, how many ask again and how many require agent intervention. Measure agent review time; copilot suggestions can save typing but still demand checking. Do not equate a tool-reported “resolution” with confirmed customer success.

Keep a human handoff visible and require an owner for content updates. If a product article changes, the AI knowledge source must change too. For sensitive questions or account actions, check identity and permission controls before allowing automation. The help desk security checklist and UK GDPR buyer questions apply to AI processing as well as ordinary tickets.

Put procurement questions in writing

Ask where prompts and outputs are processed, whether they are used for model training, which subprocessors are involved, how long data is retained, how the feature is disabled and how usage is exported. Ask for the current contract schedule and invoice example. A vague “AI included” answer is not enough to approve a budget. Record the pilot's quality threshold and revisit it after a product or policy change.

Build a usage-unit ledger before calculating savings

A useful ledger has one row for each separately billed unit. Public facts below were re-read in September 2026; operational rates in the worked example are fictional. Do not convert a session, credit or billed resolution into a successfully completed customer job without a separate definition of completion.

For example, 51 extra HelpDesk resolutions require two 50-resolution packages under a whole-package purchase assumption: 2 × $49.50 = $99, not 51 × $0.99 = $50.49. That leaves 49 purchased units unused within this calculation. Confirm package validity and carryover separately; this arithmetic does not establish either. For Freshdesk, 501 trial sessions do not imply another free 500 next month.

A fictional cost-per-accepted-outcome model

Assume 1,000 attempted automated contacts monthly, a £100 fixed AI fee, £0.50 per charged session with no allowance, and one charged session per attempt. Define an accepted outcome as an issue completed without a human handoff or repeat contact within a reader-selected seven-day window. The window is a modelling choice, not a provider billing rule.

The invoice numerator uses charged sessions, while the usefulness denominator uses accepted outcomes. Extra contacts can increase the numerator without increasing completion. Failed handoffs may also require agent repair even if the provider counts the preceding automated event as billable. For resolution-priced services, replace the invoice formula with that provider's definition; do not reuse the per-session assumption.

Keep human effort and cash separate

Suppose the central case needs 400 human handoffs at six minutes each plus ten hours of quality review. That is 400 × 6 ÷ 60 + 10 = 50 hours. At a fictional internal value of £25 per hour, effort is £1,250. Adding £600 AI cash gives £1,850, or £3.08 per accepted automated outcome, if the purpose is to allocate all those costs to that denominator. This allocation is not the cost per all supported customers; choose that different denominator explicitly when that is the question.

The lower-acceptance case leaves 600 handoffs: sixty handling hours plus ten review hours, valued at £1,750. With £600 AI cash, the comparable allocated total is £2,350, or £5.88 per accepted outcome. Neither scenario establishes savings without a consistently defined alternative workflow and its costs.

Freed staff capacity is not automatically payroll cash saved. Keep wages unchanged unless the budget genuinely changes. Enter only invoice amounts in the cost calculator's software fields and record internal hours separately. This makes it possible to judge affordability without claiming an invented automation success rate or promised reduction in headcount.

Sources

• Gorgias: billing and billable events
• Tidio pricing
• HelpDesk pricing
• Freshdesk pricing

The scenarios and quality threshold are Support Sorted's editorial recommendations.