How We Used AI to Stop Losing Clients We Already Had

Acquiring a new client costs significantly more than retaining an existing one. Research consistently puts that ratio somewhere between five and seven times, depending on the industry — and in professional services, where trust and relationship depth are core to the value proposition, the gap is even wider. Despite knowing this, most companies (including Tomedes for a long time) invested far more in acquisition than in keeping the clients they already had.

The problem was not intent. It was capacity. Retention outreach is relational, which means it resists automation in the traditional sense. A generic newsletter does not re-engage a client who went quiet after a long-standing project ended. A scheduled follow-up email sent two weeks too early (or two months too late) signals that no one is actually paying attention. For a company serving 120,000+ businesses across 270+ languages, the scale of doing this well manually was simply not achievable.

So we built something different. Over the past several months, Tomedes developed an AI-powered retention system that uses behavioral signals from our CRM and email history to identify clients who are drifting — and to draft personalized, context-aware outreach before that drift becomes a departure. This article describes exactly what we built, why we built it, and what we have learned so far. We are sharing it publicly because we believe transparency about how translation companies actually operate serves the broader industry better than polished case studies written after the fact.

Table of Contents

• What problem were we actually trying to solve?
• How does the AI behavioral analysis work?
• What does the outreach workflow look like in practice?
• How do we keep humans in the loop?
• What results are we seeing so far?
• What did we get wrong the first time?
• What does this mean for how translation companies think about client relationships?
• FAQs

What problem were we actually trying to solve?

The core problem was silent churn — clients who do not cancel, complain, or send a goodbye email. They simply stop submitting projects, and by the time anyone notices, they have moved on.

Silent churn is particularly difficult to address in translation and localization because client engagement is naturally project-based. A client who placed ten orders last year and placed none this quarter could be in a lull between projects, could be testing a competitor, or could have shifted their strategy entirely. Without a system to distinguish between those states, the default response is to do nothing — and doing nothing is how you lose accounts that took years to build.

Manual retention efforts existed at Tomedes before this project, but they were inconsistent. Account managers would occasionally follow up with clients they personally remembered. Automated email sequences fired based on time elapsed since last order, but they lacked any context about who the client was, what they had ordered, or why they might have gone quiet. The emails were well-intentioned and largely ignored.

What we needed was a system that could read behavioral patterns at scale — not just "this client has not ordered in 60 days," but "this client ordered consistently for 18 months, their last three projects were urgent legal translations, and there has been no email contact since a delayed delivery in February." That level of context changes the outreach entirely.

How does the AI behavioral analysis work?

The system connects two primary data sources: the Tomedes CRM, which holds order history, project type, language pairs, and account notes, and the operational email inbox, which contains the actual correspondence (follow-ups, complaints, questions, and confirmations) that tells the real story of a client relationship.

Claude, Anthropic's AI model, analyzes these sources together to flag accounts that exhibit one or more risk signals. The primary signals the system is trained to recognize include:

• Silence after consistent activity — an account that ordered regularly and has gone quiet for 60 or more days without explanation
• Sudden drop in order frequency — a measurable reduction compared to the client's established pattern
• Complaint or delay in recent history — a negative interaction in the email record that was resolved operationally but never followed up on relationally
• Pricing sensitivity signals — language in email correspondence suggesting the client was comparing options or reacting to cost

For each flagged account, the AI generates a short rationale (two or three sentences explaining why this account was selected) and drafts a personalized outreach email grounded in the specific context of that client's history. The draft is not generic. It references the client's actual work, acknowledges any gaps in communication, and opens a conversation rather than making an ask.

Importantly, the AI rationale is never sent to the client. It is internal — shared with the account owner through a dedicated ClickUp channel so they understand the context before deciding whether to send, modify, or discard the draft. The email the client receives is always reviewed and approved by a human.

What does the outreach workflow look like in practice?

Once the system identifies at-risk accounts and generates draft emails, the workflow moves through three steps.

Step 1 — Notification. The account owner receives a ClickUp notification in a dedicated channel (separate channels exist for the Global team and the Israel team). The notification includes the client name, the AI's brief rationale for flagging the account, and a link to the draft email waiting in their inbox.

Step 2 — Review. The account owner reads the draft alongside the AI's reasoning. They can send it as written, edit it before sending, or discard it if the context is wrong — for example, if they already know the client is on a planned pause between projects.

Step 3 — Send. Approved emails go out from the account owner's address, not from a shared alias. This matters. The email looks and reads like a personal note from someone who knows the account, because it is shaped by someone who does.

The current volume for outreach emails increased per day across both teams. That figure is intentionally conservative while the system is still being calibrated. The goal is not maximum volume, it is meaningful contact at the right moment. Scaling too fast before the quality of the signals and drafts is validated would undermine the relational credibility the system is designed to protect.

How do we keep humans in the loop?

This is the question we take most seriously, and it is worth being direct about it: the AI does not send anything on its own. Every outreach email that leaves Tomedes is reviewed by a human being before it is sent. That is not a safeguard bolted on afterward, it is the design.

There are two reasons for this. The first is quality. AI behavioral analysis can identify patterns, but it cannot always distinguish between a client who is drifting and a client who is in a quiet period by choice. A human reviewer who knows the account (or who can check the notes quickly) catches those cases before a well-intentioned email becomes an awkward one.

The second reason is relational integrity. Client relationships at Tomedes are built on the understanding that a real person is accountable for the work. Sending AI-generated emails without human oversight would erode that accountability, even if the recipient never knew the difference. The value of human review is not just error prevention, it is the signal it sends internally about how the company treats its client relationships.

Tomedes holds ISO 9001:2015 certification for quality management, and the principles behind that standard (documented processes, accountability, continuous improvement) apply here just as they do on the translation floor. AI tools are integrated into workflows; they do not replace the judgment of the people running them.

What results are we seeing so far?

The system has been running at full capacity for a relatively short time, and it would be premature to draw strong conclusions. What the early data does show is a 15% conversion rate on outreach emails, meaning roughly one in seven flagged accounts responds positively and re-engages with a project inquiry or a conversation about upcoming work.

That number is meaningful in context. The previous approach (time-elapsed automated sequences) generated responses that were difficult to attribute to the emails themselves rather than to organic re-engagement. The new system's responses are clearly tied to the personalized outreach because clients frequently reference the specific context raised in the email.

What has surprised us most is not the conversion rate but the quality of the conversations the outreach opens. Several clients who responded did not immediately place an order — they explained what had changed on their end, which gave the account team information they would not otherwise have had. Understanding why a client paused is as valuable as getting them back, because it tells you something real about the relationship and how to serve them better going forward.

The daily volume will likely increase after the next review cycle. The question the team is weighing is not whether to scale, but how to scale without degrading the quality that makes the system work.

What did we get wrong the first time?

Several things, and transparency about those missteps is part of why we are publishing this.

The first mistake was operating at the account level rather than the contact level. In the early version, the system flagged accounts (meaning companies) rather than individual contacts within those companies. For clients with multiple stakeholders, this created confusion about who should receive the outreach and meant that some contacts received emails that did not match their specific relationship with Tomedes. The system was reconfigured to work at the contact level, which is more granular and more accurate.

The second mistake was not building in sufficient human review capacity before increasing volume. In the early weeks, the pace of flagged drafts outran the account team's ability to review them thoughtfully. Emails were approved quickly rather than carefully, and some of those emails were good enough but not as strong as they could have been. Volume was pulled back and review time was explicitly protected in the team's daily workflow before scaling resumed.

The third mistake was underestimating how much the quality of the AI rationale mattered to the human reviewers. The first version of the rationale was accurate but terse — it told the reviewer that an account had been quiet for 60 days, but did not provide enough context for the reviewer to quickly decide whether the outreach was appropriate. Expanding the rationale to include more of the behavioral history made the review process faster and the approval decisions more confident.

None of these are unusual problems for a new system. But they are worth naming because building in public means accounting for the version that came before the version that works.

What does this mean for how translation companies think about client relationships?

The translation industry has historically marketed itself on quality, speed, and linguistic range. Those remain the table stakes. But the companies that will build durable client bases over the next decade are the ones that treat relationship management as an operational discipline — not a sales function, and not something that happens informally between projects.

AI behavioral analysis makes it possible to operate at a level of relational attentiveness that was previously reserved for companies with very small client bases or very large account management teams. A mid-sized translation agency can now monitor relationship health across thousands of accounts, identify the ones that need attention before the client decides to look elsewhere, and reach out with enough context to make the conversation meaningful.

This does not reduce the importance of human judgment, it depends on it. The system produces signals and drafts. The account team produces relationships. Those are not the same thing, and conflating them would be a mistake that many companies will make as AI tools become more accessible. The value is in the combination: AI that scales attentiveness, humans who convert attentiveness into trust.

For buyers of translation services, this kind of operational transparency should raise the bar for what they expect from their language partners. A professional translation company in 2026 should be able to tell you not just what languages it covers and what certifications it holds, but how it manages the relationship between projects — and how it will know when something has gone wrong before you have to say so yourself.

FAQs

Q: Does Tomedes' AI retention system access or share client data with third parties?
A: No. The system analyzes data that already exists within Tomedes' internal CRM and email infrastructure. No client information is shared externally. All AI processing occurs within Tomedes' controlled environment, consistent with the company's information security policies and GDPR obligations. Client data is used solely to improve the quality of service that client receives.

Q: How is this different from a standard CRM automation sequence?
A: Standard CRM automation fires based on time triggers, "send an email 30 days after last order." Tomedes' behavioral analysis system fires based on pattern recognition — it looks at the full history of an account, including email correspondence, to determine whether the current silence is unusual relative to that account's baseline. The outreach is personalized to the specific context of each client, not templated. The difference in response rates reflects that distinction.

Q: Can smaller translation agencies implement something similar?
A: The principles are accessible to any agency that maintains structured CRM data and email history. The technical implementation requires connecting those data sources to an AI model capable of analyzing them, which has become significantly more achievable as AI tools have matured. The harder investment is the process design: defining what signals matter, building a human review workflow, and committing to the discipline of acting on what the system surfaces. Technology is the easier part.

Q: What happens if the AI flags a client who does not want outreach?
A: The human review step is specifically designed to catch this. Reviewers can see the client's full account notes, including any indications that a client prefers minimal contact or has requested reduced communication frequency. If a reviewer identifies that outreach is not appropriate, the draft is discarded. Clients can also unsubscribe from non-transactional communication through standard channels.

Q: Will Tomedes publish further updates on how this system evolves?
A: Yes. This article is part of a series in which Tomedes shares how the company is adapting its operations to incorporate AI workflows while maintaining the human accountability that defines professional translation services. Future updates will address what the data shows after a longer run period and how the system is being extended to other parts of the client relationship lifecycle.

Working with a translation partner who invests in your relationship

Understanding how a translation company manages client relationships (not just how it handles individual projects) is one of the most reliable ways to evaluate whether a partner will serve you well over time. Tomedes has spent nearly 20 years building the processes, certifications, and human expertise that make long-term partnerships possible across 270+ languages and more than 120,000 clients worldwide.

If you are looking for a professional translation services partner that combines that depth of human expertise with modern AI-assisted workflows, including AI translation capabilities, Tomedes is ready to discuss your specific needs.

Tomedes — Professional Translation in 270+ Languages
Human expertise. AI precision. Your content, handled right.
Get a free quote → · Try AI translation →

About the author

Rachelle Garcia
AI Lead of Tomedes
Connect on LinkedIn →