The company and the original problem

Wonolo connects people looking for work with local businesses that need workers. Its support team serves two audiences: employers, for example with posting jobs, and workers, for example with paperwork and preparing for a shift. Some customer data and requests had previously been handled in a custom database; finding the right information and drafting a reply took time, while answer quality depended heavily on each specialist’s experience.

In its published Salesforce case study, Wonolo describes generating suggested replies for chat agents. This is not a fully autonomous bot that independently resolves a request and closes the case. AI proposes wording while an employee remains part of the service. The suggestions were especially useful for newer agents and for employees who felt less confident writing in English, even though they spoke it fluently.

How the described workflow works

The case names Salesforce Service Cloud, Einstein AI, case routing, and internal knowledge articles. The support team was split into two groups according to audience and skill set; queues and case types helped direct requests to the appropriate staff. Knowledge articles included, among other things, trust and safety policies. Against this background, an agent received several suggested reply options: according to Salesforce’s story, generation took two to three seconds, compared with five to ten seconds for other tools the company evaluated.

The public material does not disclose the exact document-retrieval design, system prompt, model, dataset size, or quality-assurance procedure. It therefore does not support claims that Wonolo used a particular RAG pipeline, a locally hosted model, or an autonomous agent. The practical lesson is already useful: combine the request context, the right queue, current rules, and a fast suggestion for the human agent. This is a platform vendor’s customer story, not detailed independent technical documentation.

What the results show—and what they do not

Salesforce reports a 20% decrease in average chat handling time for newer agents and a 12% decrease for more experienced agents. A separate summary says that a team of 11 agents saved more than five hours in one week. The company also associates the suggestions with greater confidence among newer employees and more consistent answers; these are useful observations, but the published material does not provide the sample size, confidence intervals, or an independent verification.

These figures cannot be transferred directly to a Russian company or turned into a promise of the same savings. Results depend on request types, starting response times, the share of routine questions, knowledge-base quality, languages, and how employees use the suggestions. Saved time is initially released team capacity, not automatically a reduction in expenses: value appears if that capacity is redirected to a backlog, more complex requests, or service quality.

There is wider context. In an OECD survey of more than 5,000 SMEs across seven countries, 37.3% of the businesses already using generative AI reported using it in customer service. However, the survey was not conducted in Russia, and the Salesforce case is not a representative market average. Treat it as a testable example, not a forecast.

When this scenario fits a small business

It is reasonable to test reply suggestions when several conditions hold:

  • the team regularly handles written requests and spends noticeable time searching for answers or typing repeated explanations;
  • requests can be grouped into clear topics such as order status, documents, service terms, or initial instructions;
  • approved answers, a knowledge base, or an accountable expert are available to confirm the rules quickly;
  • an employee can review the draft before sending, and errors can be noticed and corrected;
  • requests have histories, categories, and timestamps that make it possible to measure a baseline.

If only a few requests arrive per day, every response requires negotiation or legal judgment, and the relevant knowledge exists only in the owner’s or a key specialist’s head, a model will not remove that organizational debt. Document the rules and boundaries first, then automate wording.

A minimal architecture without unnecessary autonomy

For a first version, it is often enough to connect a model to the agent’s workspace through a CRM or helpdesk:

1. The system passes the request text and only the necessary details about the customer, order, or service to the model.
2. Approved articles, templates, and rules are selected based on the request category. If there are many documents and they change frequently, add retrieval over an approved knowledge base; a small, stable collection can be handled with controlled templates.
3. The model proposes a draft and, where useful, briefly identifies the rule it relies on.
4. The employee edits, accepts, or rejects the text and sends the reply. Cases that must be escalated to a supervisor are defined separately.
5. The draft version, final text, handling time, topic, and employee decision are logged, subject to the personal-data retention policy.

Do not immediately give the model access to payments, order changes, refunds, or uncontrolled correspondence. The first-stage goal is to speed up search and drafting, not to give a machine the authority to make commitments on behalf of the business. If replies involve personal data, financial terms, or documents, decide in advance which fields can be sent to an external provider, which must be masked, and where logs are stored.

How to run a pilot and calculate the economics

Start with one channel and a few recurring topics. Before enabling suggestions, measure average handling time, time to first reply, repeat-contact rate, escalations to a senior specialist, and customer ratings over a comparable period. Then enable drafts for a limited group and compare the same request categories, separating new and experienced agents. Do not mix complex exceptions with simple questions; the resulting average will be hard to interpret.

For each draft, record whether it was accepted unchanged, substantially edited, or rejected; also sample-check factual accuracy. A reduction in time has value only alongside acceptable quality. If employees spend the saved minutes checking errors or customers contact the business again, a better average handling time does not mean the process improved.

A simple calculation is:

`hours released per month = eligible requests × baseline handling minutes × measured time reduction / 60`.

`net effect = value of released time or added throughput − licenses − integration − training − quality assurance`.

Use your own volumes and actual tool costs. Do not record released time as payroll savings unless it produces a measurable reduction in overtime, avoids external spend, or increases completed volume without lowering quality. If data is still limited, a pilot is useful for checking the share of eligible requests and the frequency of errors—but it is not yet proof of payback.

A practical next step

Choose one recurring topic, such as service-status or document requests. Assemble permitted examples and approved replies, record baseline metrics, and test drafts with a limited agent group. Expand automation only if replies remain accurate, agents genuinely spend less time, and escalation rules for complex cases are followed.