Insight
Where Should a Business
Start With AI?
You don’t need a complete AI strategy before taking a useful first step. You do need to choose the right one.
By FXNL Team8 min read
A business decides it needs to “do something with AI.”
That can mean very different things:
- Employees using ChatGPT or another assistant
- Training employees
- Establishing rules for AI use
- Improving one repetitive task
- Redesigning a workflow
- Connecting multiple systems
- Buying software
- Building something custom
- Evaluating opportunities first
Those are different problems.
There is no single first step every business must take. The research we reviewed doesn’t compare one standard starting sequence against another, and it’s hard to see how it could: businesses start from very different places.
The right starting point depends on:
- What employees are already doing
- Where the friction is
- What outcome matters
- What information is involved
- How much risk is acceptable
Start with the work, not the AI.
That principle has support. RAND’s interviews with AI practitioners found misunderstood problems and chasing the technology among the recurring reasons projects went wrong. Government implementation guidance from the UK and the U.S. General Services Administration both start with the work and its users before the technology.
It needs one qualification. Trying new AI capabilities can reveal opportunities nobody would have asked for in advance. Exploration is useful—as long as promising discoveries are eventually tied to a real business outcome before anyone scales them.
So the goal is modest: choose a proportionate first step, and put enough ownership, safeguards and measurement around it to learn something.
First, Find Out What Is Already Happening
Leadership’s AI plan may not reflect what employees are actually doing.
U.S. Census Bureau survey data put AI use at roughly 17–20% of U.S. employer businesses between December 2025 and May 2026. That measures businesses reporting use over a two-week period—not individual workers.
Employee surveys measure something different. In Gallup’s May 2026 data, 52% of employees said they use AI at work at least a few times a year, 30% a few times a week or more, and 15% daily. Occasional use is much broader than habitual use.
The two don’t always line up inside the same company. A 2026 Census working paper found worker tasks involving AI in firms that didn’t report adopting it, and firms that reported adoption without AI showing up in worker tasks. Adoption can start from the bottom up or the top down.
And formal policy doesn’t always keep pace. A 2025 University of Melbourne and KPMG study of more than 48,000 people reported employees using AI in ways that contravened policy, concealing their use, and gaps in training and guidance. Those are self-reports, not incident records, but the pattern is worth taking seriously.
None of these figures should be compared directly; they count different populations in different ways. Together, they make a simpler point: ask employees as well as leadership.
Not all unrecorded use is a problem. Some of it is perfectly reasonable, and some of it is where the best opportunities are hiding. Where it does create risk, we look at the pattern in What Is Shadow AI—and Why Should Businesses Care?
Before choosing a platform or a project, understand:
- What tools employees already use
- What they use them for
- What information they put into them
- Where AI is already helping
- Where people are creating workarounds
Four Places a Business Can Start
These aren’t stages in a maturity model. They overlap, and a business can pursue more than one at a time. Each fits a different situation.
AI Foundation
Fits when
Employees already use AI, or leadership wants general-purpose AI available safely and productively.
A useful first step may include
- Inventorying current use
- Choosing approved tools and accounts
- Defining what information is permitted
- Assigning an owner
- Teaching two or three real tasks
- Measuring actual use and quality
Buying licenses or writing a policy isn’t a foundation on its own. Access is an input. In a randomized NBER field experiment, workers given an AI assistant spent less time on email, but the researchers didn’t detect a change in the mix of tasks people did. Individual benefit and organizational change are different things; both need deliberate attention.
Different teams may also need different training. If you’re weighing which platform to give people, see ChatGPT, Claude or Microsoft Copilot: Which Makes Sense for Your Business? Our approach to this kind of work is described on AI for Business.
Quick Win
Fits when
One bounded piece of work consumes meaningful time, its inputs are accessible, and its result can be checked.
Examples
- Intake
- Recurring document preparation
- Classification
- Research preparation
- Drafting from approved information
“Quick” and “win” are both hypotheses. The purpose is to test a bounded opportunity: record how the work performs today, try a narrow improvement on representative cases, and see whether it holds.
A trivial task can be easy and still not matter. Sometimes a larger bottleneck deserves a careful first investigation instead.
Full Workflow
Fits when
The real problem spans multiple steps, systems, handoffs, approvals, queues or teams.
Speeding up one task may not move the actual bottleneck. A faster drafting step doesn’t help if the work then waits three days for approval.
So the first step is to map the process end to end: walk through real cases with the people who do the work, find the bottleneck and choose one manageable piece to change first. In McKinsey’s 2025 survey, workflow redesign had the strongest relationship with reported financial impact of the attributes it tested. That’s an association in survey data, not proof that redesign causes results, but it fits what implementation guidance recommends.
“Full Workflow” doesn’t mean automating every step. It means understanding the whole workflow before deciding where AI belongs. For deciding which steps are worth it, see What Business Processes Are Actually Worth Automating With AI? You can also see examples of workflow work in our Projects.
Assessment and Prioritization
Fits when
Leadership sees many possible opportunities but can’t tell which deserve attention, or the problem itself isn’t yet clear.
The first step may be to
- Inventory the opportunities
- Eliminate weak candidates
- Compare their value
- Identify prerequisites
- Understand the risks
- Decide which opportunity deserves testing
An assessment helps with uncertainty. It isn’t the answer to every barrier, and it should stop when the next decision is clear. A knowledgeable internal owner can often do it themselves. The AI Readiness Assessment is one way to get a first read on where you stand.
There Is a Fifth Valid Answer: Not Yet
Sometimes the right answer is to:
- Defer
- Simplify the process first
- Fix the information first
- Use a feature already available in existing software
- Use ordinary automation instead of AI
- Do nothing, because the value is too small
If there’s no valuable, feasible opportunity, no usable information or no way to control consequential errors, deferring a deployment is a sound decision. It’s still worth finding out whether employees are already using AI and what they’re sharing with it.
This isn’t a reluctant concession. When OECD surveyed small and medium-sized businesses, those not using generative AI commonly cited poor fit with their work, along with data, legal and skills concerns. Fit is a legitimate reason to wait.
Not recommending an AI project is a valid outcome.
What Every Starting Point Still Needs
Choosing a different route doesn’t remove the need for governance. Every real deployment needs a proportionate baseline:
- A defined purpose
- An accountable owner
- Agreement on what information may be used
- A way to evaluate whether it works
- A way to handle mistakes and exceptions
That doesn’t mean every small experiment needs a large governance program. NIST’s AI Risk Management Framework treats governance as a continuing function rather than a one-time gate, and the effort should scale with:
- Scope
- How sensitive the information is
- The consequences of an error
- How much the system acts on its own
- The number of people using it
A bounded experiment can have a departmental sponsor with enough authority to own its consequences. Company-wide sponsorship matters more as cost, coordination across teams and risk increase.
Does AI Foundation Have to Come First?
No.
Nothing we reviewed establishes that a company must finish a company-wide employee platform, training program and comprehensive policy before improving an isolated workflow.
A business can:
- Enable employees broadly
- Improve a specific workflow
- Or do both at the same time
Independent shouldn’t mean exempt. A workflow project still needs appropriate access, information, accountability and controls of its own.
There’s also a scale-up point to plan for. Local successes can produce duplicated tools, conflicting permissions and inconsistent records. As deployments multiply, shared standards for identity, information access, purchasing, ownership and incident handling become increasingly useful.
The routes share dependencies. They don’t have to share a timeline.
Four Questions to Ask
Before choosing a route, answer these:
What is already happening?
Which tools are people using, for what, and with what information?
Where is the friction?
Which work is slow, costly, error-prone or frustrating—or holding back a better outcome?
Can we test an improvement?
Is there a bounded version we can try on representative cases and compare with how the work is done today?
Who owns the result?
Who decides whether it works, keeps it running and answers for it when something goes wrong?
If you can’t answer them yet, that’s useful to know. The AI Readiness Assessment can help you see where you stand.
If you already have a likely starting point, the AI Investment Estimator gives a rough planning range for what it might involve.
The first step doesn’t have to be big. It has to be chosen on purpose.