Want AI that actually works? Start by designing it around people, not tickets

Is AI really ready to replace all those time-tested, human-powered service processes in the world’s companies? Business headlines would certainly suggest so.

After all, an IDC-Microsoft global survey of 2,000 C-suite business leaders said that AI is already driving performance and saving time — realizing $3.50 to $8 in ROI for every $1 they invested in the tech. PwC’s 28th Annual Global CEO Survey backed up that sentiment, with 1 in 3 CEOs agreeing that generative AI increased revenue and profitability at their companies in 2025, and 50% of those surveyed said that they expect profit increases by 2026 because of their AI-driven projects.

With a hype cycle like this, IT leaders across the business spectrum are feeling the pressure to jump on the AI bandwagon. In fact, Gartner predicts that by 2027, 50% of business decisions will be augmented or automated by AI agents.

Why, then, is the decision to invest in AI going wrong for so many companies? For every news story touting AI’s success, you see another, like a 2025 analysis by MIT that showed only 5% of AI projects succeed.

The problem is clear. AI, at this moment, is failing to live up to its full potential. The reason isn’t because AI is removing human effort from work. It’s because we haven’t been using AI to make our processes more human-focused.

AI and the innovation paradox

AI and BI are supposed to supercharge productivity and the quality of output. But something happens when you get that shiny new dashboard, fresh cloud AI license, or that nimble chatbot. Initially, you might see some immediate gains, like customer service times decreasing, for instance. But soon, AI starts to multiply the work in ways you wouldn’t expect, setting your AI launch at a distinct disadvantage.

Some costs don’t show on a dashboard, and soon your organization may experience roadblocks such as these:

• Cognitive load migrating, with agents spending less time clicking and more time adjudicating edge cases — harder work that needs context and coaching.
• Apprenticeship moments shrinking, with bots unintentionally erasing opportunities for new hires to watch seasoned pros defuse tricky conversations.
• Customer belonging eroding when the first customer touch feels like a robot trying to contain their request, instead of a human listening and fully appreciating their problem.
• Employees becoming disenfranchised as they worry about losing their jobs to AI, or technology changes being forced that will make it impossible to do their job well.
• Errors piling up and processes lengthening as employees have to step up to correct miscommunications that arise from faulty results and data processes generated by AI.

These are some of the reasons why 88% of AI pilots never reach production, according to CIO.com’s own reporting. So, it’s no surprise that CIO Dive says that 42% of companies have now scrapped the majority of their AI initiatives.

The root causes aren’t technical. Neither are the solutions. All the data chaos, strategy misalignment and leadership inertia surrounding AI can be fixed by putting humans at the center of a company’s framework from day one.

How the most successful companies are thriving with AI

Human-powered use cases are the key to successful results for AI. Here are examples of how enterprise-level companies have gained market advantage by being early AI adopters—and putting both internal and external customer experience at the heart of their AI strategy.

Company/case study link	Results
General Motors	GM got interviews on calendars in 29 minutes using conversational scheduling, so recruiters could spend time with people — not logistics.
[7-Eleven]	7Eleven freed 40,000 hours/week for store leaders by automating screening, FAQs and scheduling with a conversational assistant.
Captain D’s	Captain D’s cut turnover 75% by pairing a conversational application tracking systems with personality assessments to hire for fit.
Tata Communications	Tata Communications increased women hires 19% with skills-based AI matching and masked profiles to curb bias.
Bank of America (Erica)	BofA’s Erica logged 3-plus billion interactions with a >98% find rate, shifting agents to higher-value, human conversations.

5 principles of human-first AI automation

The companies cited previously are mostly enterprise-level, but overall, AI affordability means that companies of all sizes can access AI, from SaaS tools like Microsoft Copilot to more customized bots and protocols. With so much competitive pressure to retool your company with AI, it’s tempting to research and purchase tools, then let your staff figure out how to use them. But if you want to create AI processes that have an impact on your business, it pays to think differently about how you’re implementing AI.

Here are my top five suggestions for how to achieve the right mindset for an AI-forward IT strategy.

Principle no. 1: Automate service for moments, not for tickets

Most AI service runbooks are ticket-centric: “if category = X, do Y.” The better question is, “What human moment am I in right now?” Anxiety after an outage is a different moment than curiosity about a new feature, for instance. This creates the perfect opportunity to build AI systems that respond to your customers’ feelings and needs, with strategies such as these:

• Adding intent and sentiment signals at event intake, so AI can handle simple inquiries, while more complex or high-emotion tasks are handled by your expert service staff immediately.
• Designing meaningful and specific first replies for first contact, acknowledging the exact request and telling customers what’s going to happen next.
• Close the loop with ticket notes that teach. When bots act, have them leave a short, human-readable note‑ in the service ticket: what was detected, what changed, how to reverse it. That single habit reduces repeat contacts and upskills the team.

Principle no. 2: Make ‘human-in-the-loop’ a feature, not a fallback

Customers don’t resent automation. They resent confusion, which happens when a system fights to finish a job it shouldn’t. Your best defense is a clear contract between humans and machines.

When putting together new processes for AI, there are three explicit guardrails to consider:

1.  Handoff rules. Define the thresholds for uncertainty, risk and emotion that trigger a human. (“If confidence < 0.8, or ‘urgent + billing’ intent, or negative sentiment twice → escalate.”)
2. Audit trail + explainability. Require bots to leave plain language‑ reasoning in the ticket: why a step happened, what changed. This protects customers, agents and compliance.
3. Two-way interrupts‑. Give agents a “pause/override” and customers an “opt‑out to human” that always works. Make these options visible in every channel.

Designing these from the start reframes automation as a service teammate with defined responsibilities, not a black box that occasionally misbehaves.

Principle no. 3: Lead with use cases, not tools

If you begin with tools, you’ll automate what’s easy. If you begin with use-cases, you’ll automate what matters. A simple discipline borrowed from business intelligence‑ programs helps here: force yourself to answer why, who and, what data before you build anything. (This mirrors the way data teams vet dashboards: define value, access and governance first, then invest)

Four questions to answer before writing a single rule:

1. Which outcome are we buying? Shorten time-to-reassurance? Reduce repeat contacts? Increase first-contact resolution for one class of issues?
2. Who gets access—and who doesn’t? Licenses, least privilege roles, sensitive data your automations may touch.
3. What are the regulatory side effects? If your automations surface or move protected data, how will you document consent, logging and retention?
4. How will we amortize ROI over time? Choose a use case that pays for itself within an acceptable amount of time, which then compounds as you reuse building blocks across adjacent workflows.

Tip: Write a one-page “use-case charter” for each automation. It should name the human moments involved, the experience metrics you’ll watch, the handoff rules you’ll enforce and the data boundaries you’ll respect. You’ll ship fewer experiments — but you’ll keep more of them.

Principle no. 4: Design data governance as your bedrock

Every helpful workflow you build is, at heart, a data flow with identities verified, privileges scoped, actions logged and outcomes auditable. If those foundations aren’t built safely, trust erodes quickly, inside and outside your company.

Make these non-negotiables‑ explicit:

• Identity and least privilege. Treat every bot, service account and integration as a first-class identity with the minimum scope required. Rotate secrets, expire tokens and separate duties so no single person can both decide and execute on sensitive changes.
• Zero trust by design. If your agents must reauthenticate for elevated actions, your automations should too. Think of continuous verification not as friction, but as a feature that preserves client trust and cleanly documents who (or what) did what, when and why.
• Data boundaries. Map which workflow touches protected data and encode red lines (no copying to external systems; redact personally identifiable information in notes; minimize retention for transient artifacts).
• Explainability plus audit. Require every automation to leave a plain language rationale in the record. This is how you pass audits and keep your humans in the loop with context that teaches.

If this sounds more like BI or security hygiene than “support,” that’s the point. The same discipline that makes analytics trustworthy makes automation trustworthy, because both are just governed movement of data.

Principle no. 5: Measure what matters to humans

Dashboards rarely show the dissatisfaction of your customer/audience. To catch and fix it, you need a paired scorecard that values experience alongside efficiency.

A balanced set we use at Integris:

• Operational key performance indicators: such as time‑to-reassurance, time-to‑first-response, first contact resolution by intent, repeat contact rate, or change failure rate for auto actions.
• Human experience: such as client sentiment deltas from first touch to resolution; deflection with‑ satisfaction (did the “self-serve” path still earn a high satisfaction narrative?)
• Apprenticeship minutes: such as how often automations leave notes that actually help a Level‑1 learn or improve an agent’s sense of autonomy pulse.

This is where the research is clear: poor automation design can degrade autonomy and informal learning. Make those risks measurable, or they’ll grow silently.

AI that lives up to its promise, now and in the future

If you want good front-end results with AI, the answer lies in having good systems and processes at the foundation, in addition to good data hygiene. That calls for company-wide planning before you design, purchase and implement new AI solutions. Identify the departments that will be affected by the new technology and execute departmental committees to conduct discovery sessions to decide what data sets your AI will access.  You can then brainstorm how their actions with AI should be stored and monitored. This will help sort out issues of process and document ownership, so you can build proper data governance guardrails.

Remember, your AI initiative is only as good as the preparation, governance and people behind it. Now’s the time to step up with the IT leadership that can take your organization’s productivity and customer satisfaction to the next level.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?