Where to Start with AI, A Practical Guide for the Overwhelmed

“We know we need to do something with AI. We just don’t know what, or where to start.”

I hear this on nearly every call now. It’s become the opening line of 2026. Executives have seen the demos, read the press releases, watched competitors announce AI initiatives. The pressure to act is immense. But the path forward is unclear.

This isn’t a failure of imagination. It’s a symptom of a technology landscape changing at an unprecedented rate. When every vendor claims AI will transform everything, it’s genuinely hard to know where transformation makes sense for your specific organisation.

There is some good news though. Finding your starting point isn’t complicated. It requires looking at your organisation through a different lens, one that spots the patterns where AI agents deliver immediate, measurable value.

Follow the Frustration

The simplest question to ask is also the most revealing, “What’s the most annoying part of your job?”

Ask it widely. Ask the person who’s been there twenty years and the one who joined last month. The tasks people dread, delay, or do poorly because they’re tedious point directly to agent opportunities.

Frustration is a signal. It tells you where humans are doing work that doesn’t suit human strengths. Repetitive lookups. Manual data entry. Copying information between systems. Checking the same things over and over. These tasks feel annoying precisely because they’re mechanical, and mechanical work is what agents do best.

Don’t start with “where could AI add value?” Start with “where do people hate their jobs?” The answers will overlap more than you’d expect.

Find the Research Burden

Look for roles where people spend significant time just finding information before they can do their actual work.

The support agent who searches through six knowledge bases before answering a ticket. The purchaser who conducts hours of market research to predict supplier pricing. The sales rep who spends an hour researching a prospect before a call. The engineer who digs through documentation to find the right configuration.

This research burden is invisible in most productivity metrics. It doesn’t show up as a line item. But it’s often a significant proportion of someone’s day, and it’s exactly what agents excel at.

Given the right tools and access, an AI agent can search across systems, synthesise relevant information, and surface it at the moment of need. The human still makes the decision, still does the skilled work, but they’re no longer spending half their time on the treasure hunt that precedes it.

Ask yourself where people are searching before they can start working. That’s a starting point.

Spot the Swivel Chair

Watch for processes where someone is literally switching between applications. Tab to the CRM, copy a customer ID, tab to the billing system, paste it in, find the invoice, tab back, type in the total.

This is called “swivel chair” integration, and it’s everywhere. The human is acting as a slow, error-prone API between systems that don’t talk to each other. Every copy-paste is a chance for mistakes. Every context switch is a cognitive tax.

Agents don’t get tired. They don’t transpose digits. They don’t forget which tab they were in. If your process involves a human being the glue between software systems, that’s a process an agent can handle faster, more accurately, and at any hour.

Map a day in the life of your key roles. Count the application switches. Each one is a potential automation.

Look for Expert Bottlenecks

Every organisation has them, the person everyone goes to when they’re stuck. The one who knows how the legacy system actually works. The only person who can configure that particular module. The institutional memory who remembers why that process exists.

These experts are invaluable. They’re also a scaling problem.

When work queues up waiting for the one person who knows, you’ve found a bottleneck that limits your entire operation. The expert can only handle so many questions. They become a single point of failure. And when they leave, the knowledge often leaves with them.

Agents can help in two ways. First, by capturing and distributing that expertise, if you can document what the expert knows (or let an agent observe their patterns), you can make that knowledge available to everyone. Second, by handling the routine questions so the expert can focus on genuinely complex problems.

If someone in your organisation is constantly interrupted because they’re the only one who knows something, that’s a candidate for agent augmentation.

Audit Your Training Materials

How long does it take to onboard someone to your core systems? Two days? Two weeks? Two months?

The length of your training programme is roughly proportional to your agent opportunity.

Complex software requires extensive training because humans need to learn the interface, the terminology, the navigation patterns, the edge cases. A 200-page manual exists because there are 200 pages worth of things a user might need to know. A two-week course exists because the system is complicated enough to require two weeks of learning.

Agents don’t need onboarding. They can be given comprehensive documentation and use it immediately. More importantly, they can mediate between the user and the complexity, the user describes what they want, the agent navigates the system to achieve it.

Look at your onboarding materials. Look at your internal wikis and process documents. These represent accumulated complexity that agents can absorb and operationalise.

Target Decision Paralysis

Many organisations have invested heavily in data infrastructure. They have dashboards, data warehouses, reporting tools, business intelligence platforms. They are, technically speaking, data-driven.

But the data isn’t being used.

Dashboards that nobody opens. Reports that get skimmed and filed. Data warehouses that only three people in the company can query. The gap between “we have the data” and “we act on the data” is vast.

This is decision paralysis from data abundance. People don’t lack information, they lack the time and skill to turn information into insight. A question that could be answered in seconds with the right query goes unasked because formulating the query requires expertise most people don’t have.

Agents collapse this gap. “Which products have declining margins in the Northeast?” is a question anyone can ask. The agent handles the SQL, the joins, the data model navigation. The human gets an answer, asks a follow-up, explores further.

If you’ve invested in data and aren’t seeing returns, an agentic interface might be the missing piece.

Watch for Quality Variance

Some tasks have consistent outputs regardless of who does them. Following a checklist, running a standard report, processing a routine form.

Other tasks vary wildly based on the individual. The proposal written by your best salesperson looks nothing like the one from someone new. The code review from a senior engineer catches things a junior would miss. The customer response from someone with deep product knowledge is categorically different from someone still learning.

This quality variance is a sign that expertise matters, and expertise can be augmented.

Agents don’t replace the human judgment that makes experts valuable. But they can raise the floor. They can ensure everyone has access to the same information, the same templates, the same patterns that the best performers use. They can catch what juniors might miss, suggest what newcomers wouldn’t know to include.

If your outcomes depend heavily on which person handles a task, agents can help standardise without sacrificing quality.

Start Internal

A final piece of advice on where to begin. Start with internal tools, not customer-facing ones.

Internal deployments have lower stakes. Your employees can tolerate imperfection while you iterate. They can give feedback, report problems, help refine the system. If something goes wrong, the blast radius is contained.

Customer-facing agents carry higher risk. Mistakes are visible. Failures affect trust. The standard is higher, and rightly so.

Build organisational confidence internally first. Let your support team use an agent to find answers. Let your sales team use an agent to research prospects. Let your operations team use an agent to query data. Learn what works, what fails, what users actually need.

Then, with that experience, extend to customer-facing applications. You’ll deploy with confidence because you’ve already worked through the hard problems.

Picking Your First Project

If you’ve read this far with your own organisation in mind, you probably have several candidates. The question is which to pursue first.

Prioritise by three factors:

Visibility of impact. Choose something where success is obvious. Time saved, errors reduced, capacity increased. You need early wins to build momentum for broader adoption.

Contained scope. Start with a bounded problem, not a transformation of everything. One process, one team, one use case. Prove value before expanding.

Available data and access. Agents need to connect to your systems. Choose a starting point where the integration is feasible, where the data exists, where you can grant the necessary access without a six-month security review.

The intersection of high frustration, clear measurement, and practical feasibility is your starting point.

The First Step

Every organisation’s path will be different. But the analysis is the same.

Speak with your team members. Watch how work actually happens, not how process documents say it should happen. Look for the frustration, the research, the swivel chairs, the bottlenecks, the training, the data that goes unused, the variance that shouldn’t exist.

You’ll find more opportunities than you can pursue. That’s the right problem to have. It means the question shifts from “what can we do with AI?” to “what should we do first?”

That’s a much better question to be asking.

Where we can help

If the “decision paralysis from data abundance” section resonated, that’s where we’ve focused much of our work. inmydata studio lets anyone in your organisation ask questions of your data in plain English, no SQL, no dashboard navigation, no waiting for someone in IT to run a report. It’s the agentic interface that turns your existing data investment into something people actually use.

But whether it’s data, process automation, or something else entirely, we’re happy to talk through where agents might fit your organisation. Twenty years of helping companies make sense of their data has taught us that the best starting point is usually simpler than people expect.

Get in touch, we’ll help you find yours.