A support assistant with fewer rewrites

Every draft took a second pass

A software consultancy maintained a reply assistant for a subscription software company. It retrieved product guidance and prepared drafts for support agents to approve. The consultancy paid the model and hosting bills under a fixed support arrangement, so unnecessary calls came out of its own operating budget.

The workflow handled 800 eligible requests per month, split between 560 routine questions and 240 complex ones. Billing disputes, account changes, and contractual questions went through separate manual routes.

The original route sent the full recent conversation to the model and rewrote every draft for tone. Even a reply that was already suitable went through another generation step.

Rewrite only when the draft needs it

The revised route selected the relevant context and checked the draft for format and tone. A rewrite ran only when that check failed. One targeted formatting repair was allowed before the task went to a person.

Product citations remained available to the approver. The revised route was introduced gradually, with the original route available for rollback.

Approval stayed with the support team

A successful draft needed no material factual correction and had to be ready within 30 seconds. Minor tone edits were permitted. Every reply still required a support agent's approval before sending.

The release criteria required at least 90% task success, no unsupported account-changing instructions, and a 95th-percentile end-to-end time below 20 seconds. Timing included queued requests, not only model generation.

The operating results

Two 30-day periods, with the same eligible volume, task mix, provider rates, and acceptance criteria. All amounts are in Indian rupees, excluding tax.

95th-percentile time is the time within which 95% of tasks completed. Technology cost per successful task includes the cost of unsuccessful work and retries.

Technology cost fell by ₹5,600 across the 30-day period. Model charges accounted for the full difference. Infrastructure and monitoring stayed at ₹9,000. Cost per successful draft fell 32.9%, reflecting both lower spend and eight more accepted drafts.

Across all attempts, input usage fell from 12 million to 4 million tokens and output usage from 1.8 million to 0.6 million. The model and rates stayed the same, at ₹400 per million input tokens and ₹2,000 per million output tokens. Those amounts reconcile to model charges of ₹8,400 before and ₹2,800 after.

The reduction came from the combined changes to context, rewriting, and repair. It does not establish a separate saving for each prompt edit. The starting workflow did a substantial amount of repeated generation, which explains the larger drop in model charges.

The consultancy received the operating benefit because it paid those bills. The client's retainer stayed the same. The useful change was a cheaper path to an approvable draft.

What this would mean over a year

At the same 800 drafts per month, the ₹5,600 period difference projects to ₹67,200 in gross annual technology savings. Annual technology spend would move from ₹2,08,800 to ₹1,41,600.