A company adding AI to its support operation needs more than a model that can draft convincing replies. Someone must connect it to customer information, establish which answers require review, track operating costs and decide whether the system is improving service.
Those responsibilities offer tech professionals a route into AI work that builds on experience they already have in software development, business analysis and project delivery.
“Most of the work in using AI at a company has nothing to do with building models,” says Suresh Sigera, lead instructor at General Assembly, a technology training provider.
Developers can connect AI services to applications and handle errors, while analysts can evaluate whether outputs are accurate and useful. Project managers can organize pilots and coordinate security, legal and business stakeholders.
“From what I see, these are exactly the areas where companies are short on people,” he says.
Start With Workflow
Choosing where AI belongs requires understanding the work before choosing a tool. Sigera looks for tasks involving varied information, such as sorting support tickets, extracting details from inconsistent forms or answering questions from internal documents.
He then asks whether fixed rules could accomplish the same task. Conventional automation may offer a cheaper and more predictable solution when the inputs and required actions are straightforward.
“I’d only reach for AI when the inputs vary a lot, the task needs some judgment about language, and an occasional mistake can be caught and fixed,” he says.
For professionals exploring this specialization, workflow mapping provides a practical starting point: identify the inputs, decisions, exceptions and approvals involved, then establish where AI could help and where human judgment remains necessary.
Kandarp Desai, CTO at Xactly, says he sees opportunities for professionals who can work through those operational details.
“What's scarce is the ability to make a model work within a real business process—with its compliance rules, legacy systems, and the people whose jobs it affects,” he says.
Desai recommends recording handling time, error rates and transaction costs before making changes. That baseline gives professionals evidence to assess whether their implementation improves the process.
He also favors combining approaches: rules can handle predictable transactions, AI can address exceptions and people can review uncertain results.
“When the rules are clear, write them down and put them to work,” Desai says. “Deterministic logic is often cheaper, faster, easier to audit, and more stable.”
Build Integration, Evaluation Skills
Developers pursuing implementation work should strengthen their knowledge of APIs, authentication, retries and rate limits, Desai says. Enough Python or TypeScript to connect services, along with prompts managed as versioned and tested code, provides a useful technical foundation.
Working with company information also requires attention to access permissions. Desai highlighted permission-aware retrieval, which prevents a system from surfacing documents a user should not see, and understanding providers’ data retention and training policies.
Evaluation offers another specialization. He recommends creating test cases based on real work and rerunning them whenever the implementation changes. Logging model versions, outputs, costs and human decisions helps teams investigate failures and understand performance over time.
Sigera similarly emphasizes testing against examples with known good answers, alongside understanding how usage translates into costs.
“It helps to understand how these tools are priced, usually by how much text goes in and comes out, so you can estimate monthly costs before anyone gets a surprise bill,” he said.
Professionals also need to investigate what information leaves the organization and whether a provider stores it or uses it for training.
“Basic logging and monitoring matter too, so problems show up after launch instead of hiding,” Sigera said.
Build a Project Showing the Whole Process
A portfolio project can demonstrate these capabilities without requiring someone to train a model. Sigera suggests a tool that categorizes support requests and drafts responses, with a person approving or editing each reply before it goes out.
The project should include sample requests, expected categories and examples of acceptable responses. Candidates can then document accuracy, failures and the cost of processing each request.
A rollout plan should identify pilot participants, measurements, feedback channels and conditions for expanding or stopping the project.
“I’d include the failures in my write-up too,” Sigera says. “In my view, being honest about what didn’t work impresses employers more than a polished demo.”
Desai proposes a first-draft responder for security questionnaires or requests for proposals. It would retrieve information from an answer library, cite supporting documents and flag questions that lack a suitable match.
Reviewers would see each draft alongside its source. Testing would distinguish answers accepted without changes, edited responses and incorrect answers, with confidently wrong responses tracked separately.
“The artifacts that make it credible are the unglamorous ones: the evaluation set, the before-and-after numbers, the rollback plan, and the honest list of failure modes,” he says.
Make Career Case with Results
When applying for jobs or seeking promotion, professionals should explain the problem they addressed, their contribution and the results they measured. A case study can show how they balanced accuracy, processing time, cost and oversight.
Sigera recommends assessing skills against a change of provider: would they remain useful if the company switched AI vendors? Use-case selection, evaluation, data protection and rollout planning can transfer across platforms.
“When I talk about my work in interviews or reviews, I focus on results, not tools,” he says.
Desai likewise urges candidates to share failures and the controls they added, giving employers evidence of how they respond when a system falls short.
“Although tools expire, judgment compounds,” Desai says. “Lead with the number you moved, not the framework you used.”
