65% AI-Powered Customer Service Automation for a Global Affiliate Network

MyLead operates a global affiliate marketing network connecting thousands of campaigns with publishers around the world. Every day, the customer-service team has to decide whether a given publisher should be accepted into a given campaign - matching the campaign's specifications against the publisher's marketing strategy, considering language, audience, and brand-safety constraints. It is high-volume work where the wrong call carries real risk - placing a children's product campaign on an adult site, for example, is the kind of mistake that damages both the advertiser relationship and the brand.

The pressure was structural: large daily volumes, sensitivity that demanded careful judgment, seasonal spikes, lost revenue every minute a publisher sat waiting in the queue, and a workforce that took weeks to onboard. Pure-human handling capped how fast the network could grow.

Softwise.AI was asked to figure out, quickly and concretely, how much of this work could responsibly be automated - and then to build the solution.

{Key Challenges}

Why human-only approvals couldn't keep up with a global affiliate network?

Discovery surfaced an interlocking set of operational pressures shaping the brief:

• High daily volume

  MyLead works with thousands of campaigns daily and must ensure each campaign's specification matches the marketing strategy described by the publisher.

  
  

• Sensitivity of campaign-to-site matching

  Customers and publishers come from very different segments worldwide - so it's critical not to place certain campaigns (e.g. children's products) on incompatible sites (e.g. adult content).

  
  

• Seasonality

  Volume varies sharply by time of day and time of year, making consistent staffing inefficient.

  
  

• Opportunity cost while publishers wait

  Every minute a publisher waits to be accepted into a campaign is a minute they cannot earn - and a minute MyLead does not capture revenue.

  
  

• Human errors and downstream complaints

  Manual decisions inevitably produce mistakes. Complaints mean less satisfied publishers, longer wait times, and more work resolving the very issues that automation could prevent.

  
  

• Slow, expensive agent onboarding

  New agents need weeks of training to be effective - driving up costs and creating a constant recruitment burden.

{Process}

How we built a hybrid statistical + GPT-4 automation in 4 stages?

Softwise.AI delivered the engagement in four clear stages, prioritising fast feedback to MyLead so the company could make informed decisions about scope and investment at every step.

1. Discovery - data analysis on a million historical inquiries

The Discovery focused on what MyLead already had: roughly 1,000,000 historical inquiries with accept/reject labels. The team worked with a deliberately bounded sample of 5,000 records - large enough to surface real correlations, small enough to analyze cleanly, and current enough to avoid issues with parameters not tracked in older data. Workshops with MyLead were used to add new data points as correlations emerged. Two feature groups dominated the analysis: application data (user scoring, description) and campaign data (provider sensitivity, restrictions).

2. Statistical solution - the cheap, accurate first pass

Before reaching for an LLM, the team tested whether a simpler statistical approach could carry part of the workload. Two clean correlations emerged: user scoring below 12 and campaigns categorized as "sensitive" together produced reliable auto-rejection on 23.4% of all applications with ~94% accuracy - comparable to human performance, at essentially zero LLM cost.

3. GPT solution - language-aware classification for the harder cases

For the remaining cases, the team evaluated multiple LLMs. GPT-3.5 would have been enough for English-only work, but MyLead's ~40 languages (across Europe and Asia) required GPT-4's stronger multilingual accuracy - despite being roughly 10x more expensive than GPT-3.5. A carefully engineered prompt instructed the model on MyLead's decision criteria and required structured outputs. The GPT-4 layer added auto-rejection on another ~23.5% of applications at ~75% accuracy - a level MyLead accepted given that historical data itself contains human error and genuinely borderline cases.

4. Proof-of-Concept, MVP, and post-launch iteration

After Discovery, the team validated the model on a held-out test set in a pre-production environment, then took the MVP into production. Within ~3 months, results exceeded original estimates: 65% combined automation - higher than the 43% projected from Discovery. Iteration after launch tightened the system further:

• Quality - prompt changes made the model less lenient, defaulting to rejection under uncertainty.

• Explainability - a diagnostic mode was added so every decision came with an explanation.

• Reliability - Temperature and Top_p parameters were tuned to give consistent classifications for identical inputs.