An account research agent that knew when to stop

The agent kept repeating completed work

A consultancy built an account research agent for a small business-development team. Given a company name and domain, it searched public sources and prepared a brief covering products, locations, recent announcements, and source links. A salesperson reviewed the brief before using it.

The consultancy paid model, search-tool, and infrastructure charges. Each period covered 480 briefs with the same mix of straightforward and harder-to-disambiguate companies. The agent did not send outreach or update customer records automatically.

The original agent used an open-ended plan, search, draft, and critique loop. It revisited domains with similar searches, passed growing search histories into model calls, and regenerated a full brief when one field lacked a citation. A downstream timeout could restart a partially completed job from the beginning.

Give each stage a finish point

The revised agent followed defined stages to resolve the company, gather evidence, extract supported facts, and write the brief. It stored completed stage outputs within the job so a timeout could resume from the failed stage. Identical searches within that job reused the existing result. Separate jobs and different companies did not share research results.

The final check reviewed required fields and source links. Missing evidence triggered one focused retrieval and repair attempt for the affected section. If the evidence remained unavailable, the brief marked the field as unavailable where the acceptance criteria allowed it.

The model stayed the same. Fewer calls and less repeated context reduced the cost.

Fewer searches still had to produce a usable brief

A successful brief identified the correct company, supported material claims with accessible sources, marked unavailable information explicitly, and completed within five minutes. A wrong company identity or an unsupported claim made the brief unsuccessful.

The revised route had to maintain at least 90% task success with the same research scope. Review covered ambiguous company names, conflicting sources, and missing information. A salesperson continued to review the output before use.

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 ₹11,040 across the 30-day period, a 36.8% reduction. Cost per successful brief fell 38.2%, from ₹69.44 to ₹42.90, with ten more briefs meeting the acceptance criteria.

Model calls fell 40.0%. Average model cost per call also moved from ₹4.00 to about ₹3.67 as later stages received selected evidence instead of the full accumulated history. These are blended usage amounts, not flat call prices. Together, the changes account for the ₹8,640 reduction in model charges.

Search calls fell from 2,400 to 1,200 at the same ₹2 rate, saving another ₹2,400. Infrastructure and monitoring stayed at ₹6,000. The consultancy received the benefit because it paid these bills.

The starting design matters. An agent that already reuses results, resumes failed stages, and limits repeated searches would have less room to improve. Here, those controls cut repeated work while preserving the required research scope.

What this would mean over a year

At 480 briefs per month, the ₹11,040 period difference projects to ₹1,32,480 in gross annual technology savings, split between ₹1,03,680 in model charges and ₹28,800 in search charges. Annual technology spend would move from ₹3,60,000 to ₹2,27,520.