Field notes on AI, practically applied.

Every month, in companies of every size, the same ritual plays out. The finance analyst blocks out two or three days. The spreadsheet gets opened, pivot tables get built, charts get formatted, and a narrative gets assembled slide by slide for an audience that will spend ninety seconds per page. Nobody enjoys this work, and yet almost every leadership team depends on its output.

This is exactly the kind of workflow AI is genuinely good at. Not in the vague, keynote-slide sense. In the concrete sense: we built the Auralys AI Analyst to take a raw spreadsheet and return a board-ready PowerPoint in minutes, and it works because the task is pattern-rich and rule-bound. But after deploying it, we learned something worth writing down. The interesting part is not what the AI does. It is what changes for the human.

What the machine does well

A reporting workflow breaks down into steps, and AI handles some of them far better than others. It is excellent at ingestion: reading a messy spreadsheet, identifying the time series, the segments, the units. It is excellent at computation: variances, growth rates, outliers, the numbers that would take an analyst an afternoon of spreadsheet formulas. It is very good at first-draft structure: ordering the findings into a narrative arc, opening with the headline, then the evidence, then the detail.

What this means in practice: the hours of mechanical work at the start of every reporting cycle compress into minutes. The analyst does not disappear. They arrive at the interesting part of their job hours earlier.

The analyst becomes an editor

Here is the workflow change that matters. Before AI, the analyst was a builder: they constructed the deck from raw materials. After AI, the analyst becomes an editor and a reviewer. They check the machine's reading of the data, sharpen the narrative, and decide what deserves emphasis. The skill that becomes valuable is not spreadsheet speed. It is judgment about what the numbers mean and what the business should do about them.

This is a promotion disguised as automation. Analysts who make this shift spend their time on analysis instead of formatting, and their work gets better because the mechanical floor has been raised. The ones who resist it, who insist on building every chart by hand, are defending the least valuable part of their own job.

Where judgment still matters

Four things the AI cannot do, and should not be asked to do:

1. Decide what the numbers mean. A 12 percent revenue dip is a crisis in one context and a planned transition in another. The model sees the dip. Only someone inside the business knows which one it is.
2. Own the recommendation. A deck can suggest cost cuts or a pricing change, but someone has to stand behind that call in the boardroom. Accountability cannot be automated.
3. Challenge the input data. AI will happily build a beautiful deck on top of a broken spreadsheet. Catching the broken input is a human job, and it is the most important quality control in the whole workflow.
4. Write for this audience. A board, an investor, and a department head need different stories from the same numbers. Choosing the story is a judgment call about people, not data.

AI should remove busywork, not judgment. The moment a team lets the model make the call, they have automated the wrong thing.

The honest limits

Two warnings from real deployments. First, garbage in, gospel out: an AI-generated deck inherits every flaw in the source data, and it presents flawed conclusions with more confidence than a tired analyst ever would. Data hygiene has to come first. Second, over-trust: the better the draft looks, the less carefully people review it. Build a review step into the workflow deliberately, with a named owner, or quality will quietly decay.

How to pilot it

If you want to test this in your own team, do it the boring way. Pick one recurring report, the monthly pack or the weekly sales review. Run the AI-generated version in parallel with the manual one for a single cycle. Measure three things: hours saved, errors caught in review, and how much of the AI draft survived into the final deck. If the draft survives at 80 percent and the analyst got a day back, you have your answer. If it does not, you have learned something cheap.

The 5-minute board deck is real. What it buys you is not really five minutes. It buys back the analyst's attention for the part of the job that was always the point: thinking.

There is an unglamorous truth sitting underneath most failed AI projects, and it has nothing to do with models. The process was broken before the AI arrived. Automating it just made it broken faster, at higher volume, with a monthly software bill attached.

We see the pattern constantly in consulting engagements. A company wants an AI voice assistant for customer enquiries, or automated invoice processing, or an AI analyst for reporting. The demo looks great. Then we map the actual workflow and find the steps nobody wrote down: the approvals that happen over chat, the spreadsheet that only one person understands, the exception handling that exists entirely in someone's head. The AI cannot learn a process that the company itself has never defined.

The tech-first trap

The trap works like this. Leadership buys the tool first and asks the team to fit their work around it. The team, sensibly, keeps doing the work the old way and treats the tool as extra admin. Six months later the tool is blamed, the vendor is blamed, and AI gets a reputation inside the company as something that does not work. What actually failed was the sequencing. Process first, then AI. Never the reverse.

This is not an argument against AI. It is an argument for earning the right to automate. A workflow that is clear, owned, and measured is a workflow an AI system can actually execute. A workflow held together by heroics and memory is not automatable by anything, however advanced the model.

How to audit a workflow before adding AI

You do not need consultants or software for this. You need a whiteboard and honesty. Here is the audit we run with clients:

1. Map the as-is process, step by step. Not the documented version. The real one. Follow a single unit of work, one enquiry, one invoice, one report, from trigger to completion. Write down every step, every handoff, every waiting period. The gaps between steps are usually where the time goes.
2. Time each step. Rough numbers are fine. You are looking for the bottleneck, the one step that everything queues behind. In our experience it is almost never the step people complain about. It is the quiet approval, the data re-entry, the weekly batch that everyone has normalized.
3. Ask the redesign question. For each step, ask: if we were designing this workflow today, from scratch, would this step exist? Steps that survive this question are the process. Steps that do not are the debt.
4. Find the exceptions. Ask the team what breaks the process. If the answer is a long list of special cases handled differently each time, the process is not ready for automation. Standardize the common path first.

The three fixes to make first

Almost every audit ends with the same three homework items:

• Standardize the inputs. AI systems fail on inconsistent inputs more than on anything else. If enquiries arrive by phone, email, and messaging with different information each time, fix the intake before you automate the handling. One intake format, enforced.
• Assign clear ownership. Every step needs a named owner, including the AI's steps. When the system makes a mistake, and it will, someone specific needs to catch it and correct it. "The team" is not an owner.
• Measure a baseline. You cannot prove AI helped if you never measured before. Pick one metric that matters, handling time, error rate, cost per transaction, and record it for a month. This is also how you will justify the investment later.

A workflow that is clear, owned, and measured is a workflow an AI system can actually execute.

When you are ready for AI

The readiness checklist is short. The process is documented as it actually runs. Inputs are standardized. Each step has an owner. You have a baseline metric. The team understands what the AI will do and what stays human. If you can tick all five, automate with confidence. If you cannot, the AI project to start is the process project, and it will pay for itself before any model gets involved.

None of this is exciting. It will not make a keynote. But in every engagement where AI delivered real returns, this unglamorous work happened first. The companies that skip it do not save time. They just fail faster, at higher volume, with a monthly software bill attached.

Here is a pattern we keep seeing in our consulting work, and it runs against the usual narrative. The fastest AI adopters are not the enterprises with innovation labs and seven-figure budgets. They are the small businesses: the clinic, the plumbing firm, the boutique agency, the twenty-person manufacturer. They are moving faster, spending less, and seeing returns sooner. It is worth asking why, because the answer contains a lesson for everyone else.

Speed of decision

In a small business, the person who feels the pain is the person who signs the cheque. The owner misses three customer calls in a week, loses two jobs to a competitor who answered faster, and by Friday an AI voice assistant is handling after-hours enquiries. The decision cycle is days, sometimes hours. There is no innovation committee, no procurement process, no security review that takes a quarter. This is not recklessness. It is the absence of organizational drag, and it is a genuine structural advantage.

Contrast the enterprise path: a promising pilot gets identified in January, budgeted in March, reviewed by legal in June, and reaches production, if it survives, sometime next year. By then the use case has changed or the champion has left. Speed is not everything, but in AI adoption it is close to everything, because the technology and the best practices are both moving fast enough to punish delay.

Direct ROI visibility

The second advantage is that small business owners see the return directly. When the AI assistant books three appointments that would otherwise have gone to voicemail, the owner knows exactly what that is worth. The feedback loop is immediate and personal. There is no attribution debate, no need to build a business case with Monte Carlo simulations. The P and L tells the story.

Enterprises, by contrast, diffuse both the cost and the benefit across departments, which makes every AI investment a political exercise. The pilot that saves forty hours a month in operations looks small against a large budget, even though forty hours a month is a meaningful win. Small businesses measure in absolute terms. Enterprises measure in relative terms, and relative terms kill small wins.

Pain is personal

The third factor is motivation. For a small business, a missed call is lost revenue with a name on it. The pain of the problem is felt daily by the person deciding. For an enterprise manager, the same inefficiency is a line item, abstracted three levels up. AI adoption follows pain, and pain is sharpest closest to the customer. This is why customer-facing automation, voice assistants, enquiry handling, booking, is where small businesses are pulling ahead fastest. They feel every dropped enquiry. Enterprises feel the aggregate, months later, in a report.

What enterprises can learn

The lesson is not that enterprises should act like small businesses. They cannot, and in regulated industries they should not. The lesson is structural, and it comes in four parts:

• Give one team real authority. Pick a single team, give them a budget and a decision-maker, and let them ship a pilot in weeks. The committee model optimizes for avoiding blame. The team model optimizes for learning.
• Measure one metric. Small businesses win because the win is obvious. Enterprises should copy this: define the single number the pilot must move, and judge it on that number alone.
• Start where the pain is sharpest. Customer-facing workflows with clear failure costs, missed calls, slow responses, manual follow-up. These have the shortest path to visible return.
• Accept the 80 percent solution. Small businesses deploy tools that are good enough and improve them in production. Enterprises wait for perfect and deploy nothing. In AI, the learning comes from deployment, not from planning.

AI adoption follows pain, and pain is sharpest closest to the customer.

The honest caveat

None of this means small businesses have it easy. They lack data scale, technical talent, and the margin for error that enterprises enjoy. A bad AI deployment hurts a twenty-person company more than it hurts a ten-thousand-person one. Their advantage is speed and clarity, not resources. The winners on both sides will be the ones who combine the two: the small business that invests in getting the process right, and the enterprise that learns to move at the speed of a team with something to lose.

The race is not between big and small. It is between the companies that ship and learn, and the companies that plan and wait. Right now, the shippers are mostly small. That is a choice, not a destiny.