Got data in silos, and no good way to see it all in one dashboard? Here is how to use AI to solve that safely, securely, and without breaking the bank.
This originally ran as an eight-part series between November 2025 and January 2026. It is collected here in one piece, because it was only ever split up to fit a social media post limit.
The gap
In the grand scheme of things, your problems are not unique. Countless other companies have hit them before. Books have been written, seminars have been held, and consultants have made big money telling everyone else how to do it.
But how do you apply all of that to your specific company, right here and right now, without accidentally shooting yourself in the foot – or the wallet?
It is all out there on the internet, with search engines at our fingertips. This should be a slam dunk. But have you tried to search for anything lately?
Meanwhile: generalist language models “know” the requisite business rules and KPIs. Specialist models “know” the tools to extract the data. And yes, Virginia, there are models for analysing and presenting things in readable formats. The tools to bridge the gap exist. The question is how to wield them for a specific problem – and, as it happens, a model can help with that too. Including the planning, the things you do not know, and the inevitable hurdles along the way.
What you are actually working with
I have written before about why language models are such a challenge for cybersecurity: they embody all of the weaknesses of computer security and all of the weaknesses of individual human beings. But they also embody some of the strengths – some of them close to super-powers – and that is where the real value is.
Plenty of people compare managing models to managing people. There are now hundreds of companies promising the best sales AI, the best admin AI, and yes, the best C-suite AI. The vast majority are thinking about it backwards.
Think of them as building blocks instead. They are limited by context, by their training data, and by the resources available to them. They are not human. They do not think, they do not actually “know” anything, and they cannot infer. All of that is an illusion – literally every output from a language model is a hallucination. We only notice when one goes off the rails.
What they are is incredibly malleable and versatile. In human terms they would be idiots savant on the subject of your choice, with very short attention spans. Every task description, down to the smallest detail, has to fit on a single sheet of paper – single-sided, standard fonts and spacing.
Step one: improve the prompt before you use it
Start with something simple:
“I have a business with multiple sets of related data that are isolated, and I would like a tool to provide a dashboard of standard KPIs for my industry.”
You will get back something very generic, plus a few generic clarifying questions – “what industry?” and so on.
Adding “interview me for what I’m actually needing and what you need to know to accomplish this” to the end gives you a series of leading questions that looks like a useful next step.
But this is not the way. Why not take it a step further and ask it to improve that prompt first?
This is one of the critical differences between a good outcome and a great one, for several underlying reasons:
- We cannot see what a model is “thinking”. The text shown as its thinking process is not that; it is curve-fitting against the outputs.
- Even when the output includes links to sources, we have no way of knowing what weighting was applied, or what else contributed without a handy link attached.
- Even after telling it to ask us questions, it builds those questions from the original request – which was, as established, simplistic and inadequate.
Asking the model to improve the prompt shifts the context into an advisory role, which lands you in territory concentrated on business best practice, metrics specific to your industry, related consulting and project management, and accounts of similar companies achieving similar goals.
Instead of throwing a ball into a ravine and hoping the slope and the rocks get it close enough to the right spot, you are having the system build you a laser-guided dart. Or: you are using a funnel rather than relying on perfect hand-eye coordination.
Try this with any prompt and compare the three results – the prompt alone; the prompt with a request for clarifying questions; and asking for help improving the prompt, then using that in a new session.
Step two: tune the prompt, then decide who executes
The improved prompt does two things for you. It fills in blanks you had not thought of – or did not know existed – and it validates that the model actually has the training to do the job. If the improved prompt does not knock your socks off, try a different service.
Then paste it into a text editor and tune it for the things it could not have known. If it references a CRM you do not use, take it out. If it lists reports, business functions or staff roles that are not yours, remove them or replace them with the right ones.
If you made a lot of changes, start a fresh session and ask for it to be improved again – the second round adds features the same way the first did. If you made a few, put them in the original session and ask for a re-do. If they were trivial, just keep your edits.
The decision point most people miss
At this stage the prompt is good enough to run. So: do you run it yourself, answer the questions, and ask the AI to break the result into prioritised phases with execution detail – or do you hand it to a trusted lieutenant to make it so?
Be aware what you are actually deciding. This is data, and reporting, and connecting to systems, and at the nitty-gritty level it gets technical. An AI can help, guide, even do the work – but it is limited by a narrow window of context, so the further into the weeds it goes, the less of the big picture it has to steer by.
If you work in an industry that routinely breaks large projects into work packages with defined deliverables, you already know what happens next, and that process carries through to completion. AI helps. It does not remove the need for human oversight at each stage, to catch what got lost in translation – and the technical detail only increases from here: accounts on services, where the scripts run, where the dashboard actually lives.
Step three: decide where to run it
With a prompt you are happy with, and your sample data gathered into one folder, the next question is where to run it. In broad strokes there are three options.
- Cloud – somebody else’s systems, somewhere else. Quick, easy and scalable; possibly poor security, with cost and complexity attached.
- Internal – your company’s existing infrastructure. Full control over everything. The downside is also full control over everything.
- Personal – your own PC or laptop. Quick, easy and private. Not scalable.
For simplicity, we will take door number three.
Step four: pick the model
You may have tried a few different services while working out your prompt, but actually building the application deserves some thought. What we are doing here is software development, even though you will not write a single line of code.
There are hundreds of options and they change fast. Limiting this to the big names – and I guarantee this summary will be out of date within six months – from worst to best:
- Microsoft Copilot – laughably bad in a number of ways, and insecure within the scope of your organisation despite their best efforts. You probably do not want interns discovering your KPIs through a helpful internal search result and posting your numbers on social media.
- ChatGPT, Perplexity and similar – fine for search, research and general work; not well suited to this task for various technical reasons.
- v0 by Vercel, Replit and similar – mentioned for completeness. Cloud options that scale well functionally but not conceptually; complex design needs tend to fail badly.
- Windsurf, Cursor – possible options, but they need some technical ability to set up and use. Adequate if you are comfortable troubleshooting for yourself.
- Claude, Gemini – setting aside systems built specifically for development, which are not household names and demand real technical chops, these are the two contenders for this project.
Both have straightforward web interfaces and both have nuances you may or may not find helpful. Try them side by side until you have a preference. Both need accounts, and you should start on the lowest paid tier at minimum – free tiers are not adequate for this, and even the highest tiers are well worth what you get.
Step five: set up, and take the first run
This next step is the hardest part: getting your system set up.
Open your browser to the service and ask it: “How do I install the command-line tool for Windows?” – or macOS, as appropriate. You will get back concise copy-and-paste instructions.
Here is the thing: you do not need to know what those instructions mean. If it looks like nothing you have ever seen and makes no sense, ask for help. “I don’t understand this, please explain step by step.”
The word “explain” is your friend. Every single step can be clarified by asking – “Explain ‘open your command prompt'” – and you will get more detailed instructions for doing exactly that.
Once setup is done, including launching the tools and authenticating, the interesting part starts. Your prompt:
“You are building a greenfield project with a local web interface, referencing example data here: [the folder with your sample files – you can drag and drop the folder into the text area and it will do the typing for you]. And instructions for processing and outputs as follows: [your big improved prompt]. Be sure to ask clarifying questions as needed.”
Then let it rip. There will be many questions – answer the ones relevant to your needs and accept the defaults for the rest. At the end you get a long explanation of what was done and a link to point your browser at. That is your dashboard.
And I guarantee it will not be what you want. That is fine. This is the first step, not the last one.
From here you are looking for three things: what it got right, what it got wrong, and what it is missing. Think carefully about the order you work through them in. If there is a source of time values – timecards, shipment dates, anything – and you notice days were read as hours, hours as minutes, minutes as seconds and seconds as microseconds, fix that before you go anywhere near utilisation or efficiency reporting.
Step six: sanity-check the numbers
Models have a lot of good training around KPIs, business process and code. What they do not have is any knowledge of your business. Two categories of error come up almost every time.
Calendar alignment
If your pay period is every two weeks, but there are calculations based on monthly or quarterly figures that do not account for pay-period boundaries, your numbers are wrong. The same applies to anything else that does not line up neatly with the calendar – per-week office supply expenses, quarterly bonuses actually paid in a different period, and so on.
Costing allocations
Say you are pulling data from your accounting package, including expenses, some paid by credit card. The card payment is itself an expense – so it is very easy for the code to count things twice, once for the original record and again inside the card payment. Many accounting systems also carry additional ledger entries that exist purely for reporting. An AI will not necessarily work that out, so the code can simply calculate the wrong thing out of the box.
Everything else
There will be other oddities. It helps to imagine you are handing the data to a precocious teenager: they will work out large parts of it and ask intelligent questions, but there will be corner cases where something is genuinely special about your environment – as there is in every company – and they will make a wrong assumption.
These are easy to fix. Tell the AI what is wrong and it rewrites the code. The hard part is working out the root cause behind an obviously wrong number – and again, the word “explain” is your friend: “Explain the calculation behind [whatever looks wrong].”
You can also go fishing, and ask what you have not thought of: “Based on the example data provided here, what additional KPIs should I consider? What other sources of information would help to validate these numbers, or could be used for more accurate reporting?”
Once the outputs look decent, widen the data range. If you had a month, try a quarter – enough to be confident there are no boundary errors lurking in the calculations.
Step seven: automate the data coming in
A brief digression first, because it will bother you otherwise. Without exception, every interface I have had an AI generate has had a terrible colour scheme. This fixes it:
“Look at [any website you like] and extract its color scheme. Then, excluding the highlights and other meaningful visual indicators, update this project to use a similar color palette. Focus on: primary/secondary brand colors, background colors, text colors, button and link colors. Keep the existing layout, just change colors.”
Now, automation. Your data sources fall into three categories.
Simple files – reports already being produced for other reasons. Easy: tell the AI to use that source. If you dragged the file in from a folder or a network share, tell it to go back there for updates.
Internal applications – anywhere from trivial to genuinely hard. Start by finding the person who knows that application best, or the vendor’s support, and asking one question: “Do you have a link for API documentation for this?” If they do, hand that link to the AI and it will ask for whatever else it needs.
If there is no API documentation, the options get worse:
- Keep providing the data manually.
- Ask the AI to work out how to extract data from that system. It will search the web to do it, which carries some risk – it should prompt before opening anything, so requests to read the vendor’s own site are probably fine, but watch for anything off-topic. It can also disappear down a rabbit hole of outdated documentation and go in circles. Be ready to abort if it is not making clear progress quickly.
- Explain what you do by hand and ask it to watch and write a script. There is no system that handles every application cleanly, and it may get lost here too.
External applications are the easy ones. Almost any SaaS platform has multiple ways to reach the data and decent documentation, and the AI will have no trouble. The catch is that many require hoops before your account can do anything, and some of those are complex and unfriendly to a ridiculous degree. The AI can usually walk you through it – sometimes you will need the vendor’s support or your own IT people.
And that is the job
If you got all the way to full automation, congratulations. If you hit stumbling blocks and are sitting on half a dashboard, you are not alone – and the point of the exercise was never only the dashboard. It was to show what this tooling can now do.
You can play with a back-hoe for half a day and come away knowing what it can do, even with nothing to show for it. That experience is what lets you have a meaningful conversation later about real work.
Four things to be clear about before you celebrate
- It is not a commercial application. Do not expect to sell it to others in your industry.
- It is not secure. You gave it curated, safe data in a controlled private environment and the output is for your eyes. This is not something to put on the internal network for everyone – unless you genuinely want everyone to see all of it.
- The architecture is exactly what you asked for, no more. It may or may not be scalable or flexible.
- It has not been properly tested. There may be subtle errors in the output you did not catch.
None of that is a showstopper, and none of it was wasted – including if you did not finish. Somebody with more knowledge in whatever area is incomplete can use AI the same way to carry it forward, and they do not need to be an AI expert to do it. That is the actual shift.
Itwerx Corp is a service-disabled veteran-owned small business providing IT services across Seattle, Bellevue, Everett and Snohomish County. This is the kind of thing our business process analysis work deals with – talk to us about yours.
