When production gets cheaper, the client relationship matters more

Published 15 April 20265 min read

Written by

Fahad IshtiaqFounder, Vela

A report can be delivered quickly and still leave the client uncertain about what to do. A customer can receive an immediate answer and still struggle to resolve the problem. Producing an output is one part of providing a service.

AI is changing the effort required for that production. A customer-support field study found that AI assistance increased issues resolved per hour by about 15% (Brynjolfsson et al., 2025). In the BCG consulting experiment, GPT-4 users completed more tasks, more quickly, within the model’s capabilities (Dell’Acqua et al., 2023).

These findings support a practical opportunity: use some of the capacity released by AI to improve the relationship around the work. That means understanding the client’s circumstances, explaining decisions and ensuring that someone remains responsible when the standard process is insufficient.

Retention matters, but slogans mislead

The economics of existing customers are often presented through a universal multiplier: acquiring a customer supposedly costs several times as much as keeping one. Keiningham et al. (2005) challenge such loyalty rules in Loyalty Myths. The practical lesson is to examine the business’s own costs and customer behaviour.

There is evidence that retention can be economically important. Reichheld’s financial-services example links a 5% increase in retention with more than a 25% increase in profit (Reichheld, 2001). The result reflects the economics of that example.

Gupta, Lehmann and Stuart demonstrate a valuation model using data from five firms. In their model, a 1% retention improvement increases estimated firm value by about 5%, compared with 0.1% for the same proportional improvement in acquisition cost (Gupta et al., 2004).

Reinartz and Kumar provide a necessary counterpoint. In their study of a catalogue retailer, longstanding customers were not necessarily the most profitable (Reinartz & Kumar, 2000). A useful relationship must therefore work for both parties. Retention without attention to service costs, fit and value can become an expensive habit.

For an adviser or service business, the practical question is which clients it can continue to serve well, and what would make that continuing relationship worthwhile.

What trust contributes to the service

Morgan and Hunt place commitment and trust at the centre of their relationship-marketing model, tested using automobile tyre retailers (Morgan & Hunt, 1994). Palmatier and colleagues’ meta-analysis finds relationship marketing particularly effective where relationships are important to customers, including services and business markets, and where the relationship involves an individual person (Palmatier et al., 2006).

The practical implication is to evaluate the people and processes around a deliverable alongside its technical quality. An accurate report still needs to address the client’s circumstances and support a useful decision.

Maister, Green and Galford offer a useful practitioner framework in The Trusted Advisor: credibility, reliability, intimacy and self-orientation (Maister et al., 2000). The framework asks a firm to look at both its competence and whose interests its advice serves.

Applied to AI-supported work, the framework suggests concrete questions. Is the advice accurate? Does the firm do what it promised? Can the client explain a difficult circumstance without being forced through an irrelevant process? Does the recommendation serve the client’s interests?

Technology can contribute to the answers. The firm remains responsible for organising a service that deserves confidence.

People respond to artificial intelligence differently in different settings

Research complicates the idea that customers simply distrust algorithms. Logg and colleagues found that lay participants sometimes followed advice more readily when they believed it came from an algorithm (Logg et al., 2019). Dietvorst and colleagues found that observing errors could reduce willingness to rely on an algorithm even when it outperformed a human forecaster (Dietvorst et al., 2015). Castelo and colleagues identify differences in acceptance depending on how subjective a task seems (Castelo et al., 2019).

The interaction matters. A routine calculation, an uncertain forecast and a difficult personal conversation ask different things of the service.

The perceived source can matter when emotional engagement is involved. Rubin and colleagues found that identical responses received different empathy ratings depending on whether people believed a human or AI had provided them (Rubin et al., 2025).

In a customer-sales field experiment, Luo and colleagues also found that disclosing chatbot identity affected purchases and perceptions of the bot’s knowledge and empathy (Luo et al., 2019). How the system is introduced shapes the interaction too.

For a business, these results suggest that a good interaction includes appropriate expectations about what the system does and a clear route to help when the customer needs something else.

Fast service still needs a route to resolution

Consumer surveys point to concerns a business should investigate. Gartner reported in 2024 that 64% of surveyed customers preferred companies not to use AI for customer service; difficulty reaching a person was a leading concern (Gartner, 2024). Qualtrics reported that nearly one in five respondents who had used AI customer service saw no benefit from it (Qualtrics, 2025).

PwC’s 2025 survey found that 86% considered human interaction moderately or very important to their brand experience (PwC, 2025). Salesforce reported that 45% would be more likely to use an AI agent with a clear escalation path (Salesforce, 2024).

Company experience is useful context too. Klarna reported that its assistant handled two-thirds of customer-service chats in its first month in 2024 (Klarna, 2024). Bloomberg subsequently reported renewed emphasis on access to human service (Daly, 2025). The company’s changing emphasis brings the service question into focus: how easily can a customer reach the help they need?

Gartner forecasts that half of organisations planning AI-related service workforce reductions will abandon those plans by 2027 (Gartner, 2025).

A business can examine a more immediate measure: does the customer reach a satisfactory resolution, including when the request falls outside the automated process?

Use the capacity to improve the relationship

The support-agent study offers a concrete example of assistance within a human-delivered service. Agents could use, edit or ignore suggestions, and the researchers observed improvements in customer interaction measures (Brynjolfsson et al., 2025).

Consider a hypothetical advisory firm preparing for a client review. AI assembles a draft summary of previous work and unresolved issues. The adviser checks it, identifies a change in the client’s circumstances and uses the meeting to reconsider priorities. The value comes from what the team can now understand and act on, with preparation supporting the conversation.

A practical service design would identify the routine work AI can assist, give clients clear expectations, and assign responsibility for exceptions. A handoff should carry the relevant history so the client can continue the conversation without reconstructing it. It would also evaluate repeat contact, unresolved cases, rework and the effort required to reach a person. Response speed belongs alongside those measures.

Use the capacity to arrive better prepared, remember what matters and resolve the problem with fewer exchanges. Clients should see the benefit in the work and know who will take responsibility when a decision needs judgement.

As production gets cheaper, the relationship gives a client a reason to keep choosing you.

Sources

Every figure in this article was checked against these sources.

All posts