The Coordination Tax: Dan O'Connell on the Hidden Cost of AI Customer Operations

1. Guest Background

This episode of Eye on AI, hosted by Craig S. Smith, is framed around a problem that has become harder to ignore as companies move from AI pilots into production workflows: AI may answer more customer questions, but it can also create more coordination work for the people around it. The guest is Dan O'Connell, CEO of Front, and the conversation analyzes customer operations rather than AI in the abstract. O'Connell is speaking from the position of a company leader whose product sits directly inside customer-support, customer-success, account-management, and sales-adjacent work.

Front is described in the episode as a customer operations platform: a single workspace where human employees and AI agents work together to solve customer problems. O'Connell says the company is just north of a $100 million business and has about 9,000 customers. He also emphasizes that Front focuses on companies with some complexity in how they serve customers. In his account, the relevant users are not only support agents. They can include client services, customer operations, account management, customer success, and even sales teams, all of whom may need shared context around the same customer.

O'Connell's earlier experience gives his claims additional context. He previously built TalkIQ, a speech-recognition company acquired by Dialpad. He describes that earlier period as a time when his team had to build speech-recognition models, machine-learning systems, and AI tooling themselves because today's developer platforms, LLM ecosystem, and frontier-model labs were not available in the same way. That background matters because his argument in this episode is not simple boosterism. He is enthusiastic about agents, but he keeps returning to implementation realities: model cost, latency, quality, handoffs, context, governance, and the human consequences of asking employees to absorb new AI-mediated workflows.

2. What the Episode Covers

The episode first establishes what Front means by customer operations. O'Connell says Front competes with traditional help desks, ticketing systems, chatbot businesses, and account-management or client-services platforms. Its ambition is broader than replacing a call-center queue or adding a web chatbot. Front wants to be the place where customer conversations begin across web, voice, email, and SMS, and where those conversations can be handled by AI first, human first, or some combination of the two depending on the value and complexity of the relationship.

O'Connell describes the interface as familiar to people who use Gmail, Outlook, or Slack: an inbox or workspace where customer conversations from different channels appear. Inside that workspace, agentic workflows may already have tried to solve the problem before escalating it. The platform then helps people categorize, triage, and respond. This product framing is important because the episode is not mainly about chatbot accuracy. It is about what happens when customer work moves through multiple systems, multiple teams, and multiple AI-human handoff points.

The conversation then distinguishes Front's AI layers. Autopilot is the external-facing autonomous agent. It can be set up on the web or over email as the initiation point for a support conversation; if it cannot solve the issue, it escalates or routes the question to a human. Copilot is the internal assistant inside the workspace, intended to help customer-facing employees complete tasks faster and better. O'Connell also describes an upcoming direction he calls AI teammates: agents woven into the conversation itself, where a worker can mention an AI teammate and ask it to do something, rather than opening a separate sidebar tool.

The episode also covers the technical and economic side of AI productization. O'Connell rejects the assumption that every task should use the largest frontier model. He says Front uses deterministic rules, traditional machine learning, traditional NLP, and frontier models depending on the work being done. Front currently partners with Anthropic and OpenAI, and does not use open source models in production, though it experiments and tests them. He frames these choices around three constraints: cost, quality of experience, and latency. In his view, teams usually optimize for two of the three, not all three at once.

That foundation leads to the episode's central subject: the coordination tax. Front surveyed 700 leaders across customer support, customer service, and account management to understand how much work sits between handoffs, shared context, communication, and cross-team problem-solving. The episode cites a striking self-reported finding: roughly three hours of coordination for every hour of resolution. O'Connell confirms that this figure came from survey respondents, not direct system telemetry. The discussion also notes that a little over 40% of companies do not measure coordination. The episode therefore analyzes a specific organizational risk: AI may increase the volume of work entering the system while leaving humans to manage the unresolved edges.

3. Core Views: Reasoning, Examples, and Limits

O'Connell's first major view is that enterprise AI value should not be reduced to headcount reduction. He acknowledges that cost matters and that some businesses will need fewer people in certain functions. But he argues that the early narrative, in which AI automates away most support work and therefore allows companies to cut support teams, is incomplete. In his account, humans do not disappear. AI elevates the type of work humans handle by automating remedial tasks and moving people toward higher-value work. He supports that claim with a small survey he was conducting among founder CEOs and support leaders; at the time of the interview, he says he had about 20 replies, with many respondents focused more on improving net revenue retention, expansion, and customer experience than on reducing cost. That is useful evidence of operator sentiment, but it should be treated carefully. It is not a universal market measurement. It is an early, speaker-cited signal about how some buyers are framing the value of AI.

His second and more distinctive view is that AI agents can increase throughput while also expanding the amount of coordination humans must perform. The coordination tax is not merely meetings or administrative chatter. It is the hidden labor required when a customer problem moves from an AI agent to a human, from one human to another, or from one team to another. O'Connell gives a concrete example: a customer request enters an inbox; ideally, AI resolves it. If a human must intervene and that first person does not know the answer, they need to find someone else inside the business. That second person needs the request context, customer history, account size, prior interactions, and possibly renewal status. Each step requires explanation, judgment, and context transfer. In that sense, AI does not simply remove work. It changes where work appears.

This explains the episode's apparent contradiction about AI-first organizations. The host notes that such organizations report high technology satisfaction, around 4.5 or 4.6 out of five, while also reporting more coordination issues. O'Connell reconciles those facts by arguing that agentic experiences may genuinely solve more things, which drives satisfaction, while also creating more handoffs, more context sharing, and more human follow-up at the boundaries. That reasoning is persuasive because it fits how workflow automation often behaves in practice: a system that handles more intake also exposes more exceptions. But the limitation matters. The satisfaction and coordination data in the episode are survey-based, and the three-hours-to-one-hour claim is explicitly self-reported. It should guide investigation and design, not be repeated as a measured law of enterprise AI.

A third view concerns model choice. O'Connell argues that companies should not treat the biggest model as the default answer for every job. Front combines deterministic rules, traditional machine learning, traditional NLP, and frontier models. The reasoning is straightforward: the right model depends on the task and on the tradeoff between cost, quality, and latency. Simple classification, routing, tagging, or rules-based escalation may be better served by deterministic logic or smaller models. Rich summarization, complicated reasoning, or agentic task execution may justify stronger models. The limitation is that O'Connell does not disclose specific thresholds, benchmark results, or internal cost curves, so the episode supports a principle rather than a technical prescription.

A fourth view is that customer context is the real bottleneck. O'Connell is skeptical of the idea that companies can buy a chatbot, ticketing system, Slack workflow, account-management tool, sales system, LLM layer, MCP servers, and APIs, then reliably stitch all of it into one coherent customer-operation brain. His argument is that customer problems live across current conversations, past conversations, account history, renewal data, product usage, and multiple people inside the customer organization. If those signals are fragmented, an AI agent may be acting on partial reality. This view aligns with Front's platform thesis, so it should be read as both analysis and product positioning. Still, the operational point is valuable: agent quality depends not only on model intelligence but also on whether the system can see the right customer context at the moment of action.

Finally, O'Connell's long-term view is optimistic but bounded. He believes that agents will eventually contact other agents to solve problems and may even deploy code to fix issues. But he does not think this becomes the fully realized enterprise default in the next couple of years. He also rejects the idea that every company becomes a one-person business surrounded by agents. His reason is not nostalgia; it is that high-value business still involves trust, relationship-building, buying journeys, and the human experience of building things with other people. The likely future, in his framing, is not the disappearance of customer-facing workers. It is the blurring of support, success, account management, and sales roles as AI handles remedial work and people move into more judgment-heavy forms of customer engagement.

4. Learning and Application

The first practical lesson is that AI programs need coordination metrics, not just automation metrics. A company can report that an agent handled a large share of incoming questions and still make employees more exhausted if every unresolved case requires a messy handoff. Teams should measure how often AI escalates, how many people touch the escalated case, how much context has to be reconstructed, how long humans spend checking AI output, and whether customer issues are actually resolved faster and better. O'Connell notes that customer support is already a highly measured environment, with metrics around ticket cost and response speed. That makes it a strong place to demand clearer AI ROI rather than accepting vendor promises.

The second lesson is to define handoff boundaries before scaling agents. If a customer is high value, renewal is near, account history is complicated, or the issue carries relationship risk, human involvement may be the right design rather than a failure of automation. The operational question is whether the human receives enough context to act well. A good escalation should include the customer request, prior conversation history, account data, product usage when available, and the reason the AI could not or should not complete the task. Without that context, the company has not automated the work; it has redistributed the burden to the employee.

The third lesson is to separate model strategy by task. Enterprises can classify customer-operation work into deterministic actions, routing, tagging, summarization, prediction, generation, and agentic execution. Deterministic rules or traditional models may be enough for simple categorization and workflow triggers. More advanced models may be worth the cost for ambiguous customer requests, multi-step reasoning, or high-quality summarization. This approach helps manage cost and latency while preserving quality where it matters. The tradeoff is operational complexity: teams must maintain multiple layers of intelligence, monitor their performance, and revisit decisions as frontier pricing, open source capabilities, and internal data needs change.

The fourth lesson is to treat customer conversation data as a source of directional signals, not automatic truth. O'Connell argues that customer conversations can reveal sentiment, purchase intent, churn risk, product frustrations, marketing-message resonance, and roadmap input. But he is careful about prediction. Models should be directionally right rather than presumed perfect. For churn or upsell prediction, a team should combine what customers say in support interactions with product usage, account-level context, and patterns across several people at the same customer. Used well, these signals can help prioritize outreach and investigation. Used poorly, they can turn one negative interaction into an overconfident account judgment.

The fifth lesson concerns people and organization design. O'Connell describes a future in which support, customer success, account management, and sales boundaries blur. If AI handles remedial work, companies can either reduce headcount or redeploy people who already understand customers and the industry. The more constructive path is to move those employees into higher-value customer work: risk-account engagement, upsell support, relationship management, and supervision or construction of agents. But that shift only works if expectations are explicit. Otherwise, employees inherit old duties, new tools, more requests, and the emotional burden of proving that AI has made them more productive.

The final application is procurement discipline. Buyers should ask AI vendors to show observability, governance, measurable impact, and cost clarity. Why did the AI make a recommendation? What did it cost to run the task? What happened when it failed? How was context passed to a human? Can the vendor connect the feature to customer experience, retention, expansion, or bottom-line impact? O'Connell says buyer expectations have evolved from accepting the dream of AI to demanding proof. That does not eliminate the coordination tax, but it turns it into something visible enough to manage.

Source

• Original episode: The Coordination Tax: Why AI Is Burning Out Your Best People | Dan O'Connell, Front