AI COST OPTIMIZATION

Reduce AI Costs for Support Teams: 7 Changes That Lower Cost per Investigation

Most support teams overpay for AI because they run the wrong model on every ticket, skip caching, and buy pricing models that charge for volume they don't control. Seven operational changes fix most of it.

7 changes
Operations-focused
Cost per investigation metric
Quick answer

To reduce AI costs for support teams: segment tickets before routing to AI, enable prompt caching on shared context, audit and trim token footprint, use the Batch API for async work, match pricing model to ticket mix, measure cost per investigation, and review AI escalation rates quarterly.

Where AI support costs come from

Support teams usually look at AI cost as a single line item, but it comes from two different places. First, there is LLM API cost: token volume multiplied by model price. That is where model choice, prompt length, cached reads, and batch discounts matter. Second, there is vendor pricing. Per-seat contracts charge a flat amount whether ticket volume is low or high. Per-resolution pricing moves in a straight line with handled volume, which can look cheap at first and get expensive fast as automation expands.

The more important driver is often not price per token. It is investigation time. If an agent still spends 30 minutes opening four systems, checking account history, confirming billing state, and reviewing logs, then lower token prices will not change the unit economics in a meaningful way. The queue is still expensive because the workflow is slow.

That is why support cost work needs two lenses at once: API efficiency and operational design. Optimizing only prompts can help, but it will not fix a queue where too many tickets are treated as full investigations. Good teams lower cost by routing simple work away from expensive models, caching shared context, and measuring what the hard tickets actually cost to finish.

7 changes that lower cost per investigation

Segment tickets before routing to AI

A two-tier routing layer classifies intent first, then assigns the model. Simple intent such as FAQ, status check, and password reset can go to GPT-4o mini at $0.15 per million input tokens. Complex investigation such as system query or root-cause review should go to GPT-4o or Claude Sonnet. Teams doing this often cut expensive model calls by 40 to 60 percent on mixed queues.

Enable prompt caching on shared context

If your system prompt contains product docs, support playbooks, or policy text, turn on prompt caching. OpenAI charges $1.25 per million tokens for cached GPT-4o input instead of $2.50. Anthropic charges 0.1× base price on cache reads. A 10,000-token system prompt can trim about 90 percent of that prompt's input cost on every request that shares the cached prefix.

Audit and trim your token footprint

Run a token count audit on 100 real tickets. Most teams find that 30 to 40 percent of tokens are formatting rules, repeated context, or fields the model never uses. Trim those first. Token reduction usually produces the best return per hour because it improves every request before you change the model mix.

Use the Batch API for non-urgent AI work

Post-resolution tagging, quality scoring, and escalation review do not need real-time processing. OpenAI and Anthropic both publish 50 percent discounts on batch jobs. If even 30 percent of ticket volume moves to batch, that slice of API spend drops in half without touching live support flows.

Match pricing model to ticket mix

Per-seat pricing at roughly $40 to $100 per agent per month is predictable. Per-resolution pricing at roughly $0.99 to $3 scales with handled volume and rewards containment volume. Per-investigation pricing from Altor charges only when investigation happens, which is often 20 to 40 percent of B2B support volume. The pricing structure can change total spend more than model selection does.

Measure cost per investigation, not cost per ticket

Cost per ticket blends simple deflections with hard cases. Cost per investigation tells you what the difficult work actually costs. Use this formula: API spend on investigated tickets divided by number of investigations. That is the number to optimize when your queue includes technical or account-specific support.

Review AI escalation rates quarterly

An AI workflow that escalates 70 percent of tickets back to humans is adding a step more than removing work. Review escalation triggers and decision trees every quarter. High escalation rates usually point to missing playbooks, unclear routing, or bad data access rather than weak model capability.

Cost math for a 10-agent support team

For a 10-agent team, contract structure can outweigh API math. A per-seat tool stays flat as volume changes. A per-resolution tool looks simple but rises in direct proportion to handled volume. A per-investigation model narrows spend to the subset of tickets that need actual work, which is often the better fit for US B2B teams with complex accounts and lower ticket counts.

Pricing model 1,000 tickets/mo (60% AI resolution) 5,000 tickets/mo Key risk
Per seat ($50/agent × 10) $500 flat $500 flat Doesn't flex with volume
Per resolution ($0.99 × resolutions) $594 $2,970 Scales linearly; expensive at high volume
Per investigation (20-30% investigation rate) Based on 200-300 investigation events Scales with investigation count Right-sized for investigation-heavy queues

Note: Per-investigation cost is not published. Contact Altor for a quote based on your investigation rate.

What good looks like after 6 months

Six months into an optimization program, the goal is not a low headline API bill. The goal is a lower cost per investigation while the support experience stays steady. Teams that make progress usually track a small set of operating numbers and review them the same way they would review backlog, SLA, or staffing metrics.

Cost per investigation

(API + vendor spend) ÷ investigations completed

Escalation rate

Percent of tickets routed back to humans

Resolution rate

Percent of tickets closed without human review

Agent time on investigation

Minutes per investigated ticket

When these four numbers move together in the right direction, the economics improve for the right reason. You are not just squeezing prompt length. You are routing better, containing more tier-1 work, and reducing the amount of human time needed on the tickets that remain.

Next step

Model your queue before you change contracts

Book a 30-minute demo and compare per-seat, per-resolution, and per-investigation costs against your actual support mix. You can also start with the calculator.

FAQ

What is a typical AI support investigation cost?

Using Claude Haiku 4.5 at about 3,700 tokens per conversation, API cost is about $0.0037 per investigation. GPT-4o at 2,500 total tokens costs about $0.011. Vendor pricing such as per-resolution or per-seat often adds $0.50 to $3 per handled ticket on top.

How much can support teams realistically save on AI costs?

Teams applying model routing, prompt caching, and Batch API for async work can cut API spend by 50 to 80 percent. The larger variable is pricing model choice: per-seat, per-resolution, and per-investigation can produce cost gaps that are much larger than model selection savings on a given ticket.

Is per-resolution or per-seat cheaper for a 10-person team?

At under 500 resolutions per month, per-seat at $500 for 10 agents at $50 each is cheaper than per-resolution at $0.99 per event. Above 500 resolutions per month, per-seat usually wins unless you have more agents. Model your own queue in the AI Support ROI Calculator.

Does switching to GPT-4o mini hurt support quality?

On tier-1 tickets, GPT-4o mini often performs comparably. On complex multi-step investigation that needs billing checks, logs, and account state, quality drops. Route by ticket intent instead of replacing GPT-4o entirely.

Related pages