AI Buyer Trust and ROI: Enterprise Leaders Debate What It Really Takes to Win Confidence in AI Services

Diginomica analyst Jon Reed published a weekly enterprise roundup on 21 September 2026 arguing that AI vendors are focusing on the wrong problem when they ask how to get customers to trust their AI systems. According to Reed, the more productive question centres on transparency, explainability, and buyer empathy rather than return on investment alone.

The trust gap in AI services

Reed drew a distinction between trust in AI services and trust in AI itself, arguing that vendors routinely blur the two. He contended that obsessing over whether customers trust AI misses the mark: "You can get a very good result out of AI without fully trusting it. Arguably, you'll get a better result, via the judicious use of guardrails and approval steps."

Reed identified transparent pricing as one of the most significant trust factors, describing typical vendor AI pricing as "a gooey (and perpetually shifting mix) of embedded features, add-on credits, consumption pricing shifts, and outcome flirtations." He also pointed to data privacy and architecture visibility as critical components.

Observability as a selling point

An engineer identifying himself as Lukas Die Kunst responded to Reed's earlier piece on X, arguing that the willingness to instrument and monitor AI systems should be treated as a competitive advantage rather than an admission of weakness: "In aerospace I instrument everything precisely because failure is expected somewhere. Admitting your AI needs observation should be a selling point, not an embarrassment."

Limits of ROI as a measure of value

The roundup also highlighted external commentary questioning ROI as the primary lens for evaluating AI investments. Analyst Esteban Kolsky was cited as arguing that "ROI is therefore a poor primary measure of transformational business value when the objective extends beyond improving an existing activity. Legitimate returns from automation, lower costs, or faster processes show that an activity improved, but they do not establish whether the enterprise created a better operating model, changed how decisions are made, developed new sources of growth, or built capabilities competitors will find difficult to replicate."

Agentic failures and debugging complexity

On the challenge of diagnosing AI agent failures, the piece quoted a source in The New Stack: "'It's not enough to just look at the logs or the inputs and the outputs,' Hallak tells The New Stack. 'It is important to figure out how it got to the answer. What were the reasoning traces? What tools did it utilize? Where did it get stuck? Where did it decide to try a new approach?'"

Tokenomics and AI cost control

The roundup also referenced diginomica's ongoing tokenomics series, with coverage noting that cutting AI budgets may not resolve so-called "token shock." Citing Neo4j's Jim Webber and a paper from the NICD, diginomica wrote that "the more important way to look at the cost is how many round-trips the agent has to make to get a usable prompt."

Dreamforce 2026: Salesforce bets on interface agnosticism

Diginomica's coverage of Dreamforce 2026 featured Salesforce CEO Marc Benioff describing a deliberate strategy of avoiding lock-in to any single user interface. According to the piece, Benioff said at the event: "By the time we get back here [to Dreamforce 2027], or let's say by even as we get to January 1 of next year, I'm confident you'll have a lot of different interfaces. The interface customers will choose their religion here. We're agnostic."

Benioff added: "One big advantage we're going to have is the agnostic aspect of the interface. We're going to give customer choice. I think customers will have different religious preference here. There's no question."

The conference also featured an AI safety session in which Benioff interviewed Anthropic's Dario Amodei, OpenAI's Sam Altman, and Nvidia's Jensen Huang, according to diginomica's reporting.

MCP and AI skills

Reporting from MCPCon Europe quoted Soria Parra, a contributor to the Model Context Protocol, as saying that when MCP was being created two years ago, tool calling "barely worked. Models had to be constantly corrected." Parra and other maintainers were described as arguing that the skills needed to work with current protocol and platform layers were already present in the enterprise workforce.

Leadership development and AI mandates

The roundup cited a diginomica piece questioning whether leadership development programmes remain fit for purpose in an AI era, with Reed summarising the conclusion as: leaders need to guide employees through transitions rather than impose ill-considered AI usage mandates.