Tharu Technologies places contractors and full-time hires with companies across the United States, so its phones ring for two very different reasons. Candidates call about their applications, their pay and their next role. Clients call about the people already working for them and the ones they still need. ConnectGen built Tharu an AI-powered contact center on Amazon Connect that answers both, and places outbound job-offer calls on its own. It started as a proof of concept and now runs in production in Tharu’s environment. This is how it was put together, and the design decisions that made it work.
The challenge
Most calls into a staffing firm are routine, but they are routine in many different ways. A candidate wants to know where an application stands, how to submit a timesheet, or whether anything is open in Dallas. A client wants to see which contractors are active on their account, open a request for three developers, or check an invoice. Each of these has a clear answer in a system somewhere, and each one handled by a person takes a recruiter or coordinator away from work only they can do.
The outbound side has its own cost. When a role opens, someone has to work through a list of suitable candidates and call each one, and many of those calls end with no answer, a request to call back, or a polite no. Tharu wanted routine conversations handled automatically in both directions, with its people brought in when a candidate says yes to a role or a caller needs judgment that a system cannot give.
Why two AI agents instead of one
The obvious design is a single AI agent that takes every call. ConnectGen built two. An inbound call starts with an open question about who is calling and what they need. An outbound job-offer call starts with Tharu already knowing both, because it placed the call about a specific role. One agent asked to switch between those modes tends to blend them, opening a campaign call with “thank you for calling” or treating a candidate it just dialed as a stranger. Giving inbound calls and outbound job offers their own agents removed that confusion and made each one easier to test and tune. Both introduce themselves as Sarah and speak with the same generative Amazon Polly voice, so callers hear one consistent assistant whichever way the call goes.
The solution: an AI contact center for candidates and clients
The platform is Amazon Connect with AI agents, using Amazon Lex for the conversation layer and AWS Lambda functions that the agents call as MCP tools. Amazon DynamoDB holds candidates, jobs, placements, staffing requests, billing and call outcomes. Amazon Connect Customer Profiles, Amazon SES and a knowledge base stored in Amazon S3 complete the picture. The diagram shows how inbound and outbound calls move through it.
Select any part of the diagram to see what it does in this setup.
Callers are recognized by their phone number
Callers rarely know their candidate ID, so Sarah never asks for one up front. As soon as a call connects, a lookup on the caller’s number tells the agent whether it is speaking with a known candidate, a client, or someone new, and a returning candidate who asks about an application hears the answer straight away. If the number isn’t on file, status can still be found by candidate ID or by name. When someone tries to register from a number that already exists, Sarah tells them they are already in the system and confirms their details rather than creating a duplicate record.
Candidates can register, search and update by phone
A new candidate can register entirely over the phone. Sarah collects their name, contact details, location, skills, preferred job type and expected pay, creates the record, and a confirmation email follows through Amazon SES. Existing candidates can search open roles by location, skill, pay or job type and hear the best matches described conversationally, then update their own profile when something changes. Questions about timesheets, pay schedules, benefits, 401(k), PTO, direct deposit and onboarding are answered from a knowledge base built on Tharu’s own policy documents. When Sarah isn’t confident in an answer, she offers a transfer to payroll, HR, benefits, immigration, recruiting or whichever department fits. Status answers are scoped so that internal notes and client feedback are never read out to a candidate.
Clients get the same self-service
Hiring managers at Tharu’s clients use the same line. They can check which contractors are active on their account with roles, start dates and status, submit a new staffing request and receive a request number before they hang up, check on that request later by its number or their phone number, and review invoices, amounts, due dates and payment status. Requests that are urgent or complicated go to a person, as described below.
Outbound job offers open with the facts
Tharu’s recruiters load candidate phone numbers into Amazon Connect Outbound Campaigns. For each number dialed, an AWS Lambda function finds the candidate and matches them to the best open role before anyone speaks. If there is no candidate on file or no suitable job, the call ends politely. When there is a match, the contact flow plays a greeting built from the candidate’s first name and the role’s title, company, location, pay rate and contract type, and then hands the conversation to the outbound agent, which can answer follow-up questions about pay or location from the same job details.
ConnectGen deliberately kept that opening line away from the language model. It is the one sentence every call depends on, and generating it from data in the contact flow means it is accurate on every call instead of most of them.
Every answer is handled, including no
A campaign proves its value in the calls that don’t go to plan. When a candidate declines, Sarah asks whether the issue is location, pay or the type of role, and offers alternatives that fit better. A candidate who is busy is asked for a better time to call back. A candidate who asks not to be contacted again is flagged as do-not-call and the call ends with an apology. Each outcome is logged against the candidate as interested, declined, callback or do-not-call, so recruiters can see what happened on every call without listening to any of them. Voicemail and unanswered calls are left to the campaign’s answering machine detection.
Handoffs arrive with context
When a person is needed, Sarah records the reason as contact attributes before transferring, so whoever picks up already knows why the caller is there. Four escalation paths cover the cases that matter most.
| Situation | Routed to |
|---|---|
| A candidate says yes to a job offer | Outbound recruiting |
| A client needs staff in less than a week | Client services |
| A client disputes an invoice | Finance |
| The caller is frustrated or the request is complex | General queue |
The first of these is where the business value sits. An interested candidate reaches a recruiter while the conversation about the role is still happening, rather than waiting for a call back.
Records stay in sync
Every profile Sarah creates or updates is synced to Amazon Connect Customer Profiles, so newly registered candidates are available to future outbound campaigns without anyone exporting or re-importing a list. Candidates receive an email confirming each change, which gives them a record of what they told Sarah and gives Tharu a clear trail of when it happened.
The outcome
The solution has moved from proof of concept into production and is running in Tharu’s own environment. Candidates can register, check on applications, find work and get answers to policy questions without waiting for a recruiter to be free. Clients can see their placements, raise and track staffing requests and look up invoices in the same way. Job offers go out automatically, every response is recorded, and recruiters spend their time on the phone with candidates who have already said they are interested. Tharu’s people still handle the conversations that need them, and those conversations now reach them with the context attached.
A word from Tharu Technologies
“Too much of our recruiters’ day was going to calls that had a simple answer. Sarah handles those now, for candidates and for our clients, and when a candidate says yes to a role the call reaches a recruiter with everything they need to pick it up. ConnectGen took us from proof of concept to production, and the system has been running smoothly in our environment.”
Anupama KatikaneniCEO · Tharu Technologies
When this approach works, and when it does not
This design suits businesses whose calls are frequent, similar in shape and backed by structured data, which describes most staffing firms well. It depends on that data being kept in good order. Phone-first lookup works because numbers are stored consistently, and job matching can only be as good as the job records behind it. It also does not replace people. Sarah answers questions and completes defined tasks, but she does not negotiate pay, settle billing disputes or make placement decisions, and the escalation paths are built around that boundary. Outbound calling to candidates carries compliance obligations too, so do-not-call handling and consent need the same care in operation as they received in the build.
Key takeaways
- Separate inbound and outbound agents. Each conversation is easier to get right than with one agent that switches modes.
- Keep must-be-exact lines out of the model. The opening of a job offer belongs in the contact flow, not in generated output.
- Identify callers by their number. It removes the most common source of friction before the conversation starts.
- Escalate with context. A handoff is only as good as what the person picking up knows when the call arrives.
How ConnectGen helps
ConnectGen is an AWS Advanced Tier Partner and Amazon Connect Delivery Partner. The work for Tharu draws on our Agentic AI services and our practice in AI-enabled contact centers, where we design AI agents around the conversations a business actually has and connect them to the systems that hold the answers. Once a solution is live, our managed Amazon Connect services team keeps it monitored and tuned as call patterns change, and for organizations starting from scratch, Contact Center Builder stands up a configured Amazon Connect contact center quickly.