Summarize with AI
To explain metrics to non-technical users, put one plain sentence under every number. It should say what the number means in their words, whether it is good, and what to do next. Write that sentence before you pick a chart, because a nicer graph of a number nobody understands is still a number nobody understands.
This matters most in the first week. A new user opens your dashboard to find out if the product is working for them, and the dashboard answers in your units, not theirs. You know what every tile means because you built the pipeline behind it. Nobody else does, and they will not write in to ask. They log in twice and stop.
The office manager who reads six tiles and asks the front desk
Picture an AI phone receptionist sold to small dental practices. It answers calls, books cleanings into the practice calendar, and hands anything tricky to a human. The buyer is the office manager at a three-chair practice. She is not technical. She signed up because the front desk misses calls at lunch and after six.
After a week she logs in and lands on a page called Overview. Across the top sit six tiles. Containment 0.61. Intent match 0.82. AHT 2:41. Escalations 17. Calls answered 438. Sentiment +0.3. Under the tiles is a line chart called Volume by hour. On the right is a blue button that says Tune agent. It is greyed out, and its tooltip reads Upload at least 50 labelled transcripts to enable.
She came with one question. Did it book anyone, and did anyone hang up annoyed? That answer exists. It lives under Calls, then Filters, then Outcome, then Booked. Nothing on Overview points there. She does not know if 0.61 is a lot. She does not know what AHT stands for. Sentiment +0.3 could be fine or could be a warning. So she closes the tab, walks to the front desk, and asks the receptionist whether the robot booked anyone this week.
Why 0.61 looks finished to the person who built it
To the founder, every tile is complete. Containment is the share of calls the agent closed with no human, so 0.61 means it handled most of them. Intent match is how often the model put a caller's request in the right bucket. He wrote those definitions in a Notion doc in month two, and he reads the tiles with that doc open in his head.
Watch what he does with the dashboard. On sales calls he shares his screen and walks it tile by tile, and the practices that get that walkthrough tend to stay. The ones that sign up from the website log in on day one, sometimes day two, then go quiet. Support gets the same email in different words, some version of is 0.61 good. His fix so far is a Glossary link in the footer of Overview.
He reads the quiet trials as a targeting problem and narrows the ads to bigger practices. The Overview page has not changed since the first version shipped.
A chart will not explain the number, and neither will a glossary
Most advice on this question jumps straight to visuals. Pick the right chart type. Use green and red. Add a sparkline. Those help a reader who already knows what the number is. They do nothing for a reader who does not. A seven day line of Containment is seven numbers the office manager cannot read instead of one.
The glossary and the little info icon fail for a different reason. They define the word. Containment, the share of conversations resolved without human handoff. Now she knows what it is, and she still does not know if 0.61 is good or what she should do about it. A definition answers what. She was asking so what.
Here is the claim we would defend. A number without a sentence under it is not information, it is homework. On a first-run dashboard every tile needs a plain verdict before it needs a chart. If you only have time for one of the two this sprint, ship the sentence and delete the chart.
Is 0.61 good? Give every number something to stand next to
A verdict needs a baseline. Without one, the user has nothing to hold the number against, and she cannot tell a good week from a bad one. There are three honest places to get it.
- A published threshold. Google does this for page speed. It does not just print a load time. It says a Largest Contentful Paint of 2.5 seconds or less is good and anything over 4 seconds is poor, measured on the slower end of real page loads, as its LCP guidance on web.dev explains. LCP 3.2s alone means nothing to a marketer. Next to that line it means needs work.
- The user's own history. Last week, the first week, or the week before she changed a setting. Up from 0.48 last week is a verdict anyone can read, even before they know what the metric is.
- A target she set herself. At signup, ask one question: what would make this worth paying for? If she answers no missed calls after six, build a tile for exactly that and say how close she is.
What you should not do is make up an industry average because the tile looks bare without one. If you have no real data on other practices, her own last week is a better baseline than an invented one.
The same rule works far from software. Mi-VAD is building a heart pump that weighs about 12.7 grams, and that figure alone tells an investor very little. On the site we designed for Mi-VAD, the device sits beside the pump it competes with, which is the size of a fist, and beside a quarter dollar coin. The reader gets how small it is before any spec appears. The comparison does the explaining.
Turn each tile into a sentence that ends in a verb
The sentence under the number has three jobs, in this order. Say what happened, in the user's words. Say whether that is good. Say what to do next, or say plainly that nothing needs doing. Here is the Containment tile before and after.
Before: Containment 0.61.
After: Handled without your team, 61 of 100 calls. Good for a first week. Most of the other 39 were insurance questions. Add your insurance list and it can answer those too. Below that sits one button, Add insurance list.
The number is still there. It is just no longer alone. The unit changed as well. 61 of 100 calls is the same fact as 0.61, written the way a person counts. Counts of real things, like calls, invoices or posts, beat ratios and scores for this reader every time.
Escalations 17 gets the same treatment. After: 17 calls went to your front desk. 12 asked about insurance and 5 asked for a named dentist. Open these 17 calls. Now the tile is also the path to the screen she came for, the one that used to sit four levels down. The greyed Tune agent button goes too, or becomes the one step she can take today.
What to hide on day one
Not every metric deserves a sentence on the first screen. Some deserve to leave it. A good test is whether the user can act on the number this week. If she cannot, it belongs in a Details tab, not on Overview.
Intent match 0.82 is a number for the team tuning the model. It tells the founder whether the classifier is drifting, and the office manager can neither change it nor use it. AHT, average handle time, is a call center metric. For a dental practice, a longer call that ends in a booking beats a short one that does not, so a falling AHT might be bad news dressed as good. Sentiment +0.3 is a model's score of other conversations, and without its scale it reads as a coin toss.
Move those three to a tab called Advanced and put up the numbers she would have counted by hand: appointments booked, calls answered after six, and calls that needed her team. Six tiles become three, and all three answer her question. Our guide to SaaS dashboard design covers how to rank what stays on the default view over time. This post is only about the first week. If the practice has had no calls yet, the tiles need an empty state as well, and the advice on designing SaaS empty states applies: say what will show up here and what makes it show up.
A tile rewrite you can do this week
- Open the dashboard in a fresh account with a week of real or sample data. Screenshot the first screen exactly as a new user sees it.
- Under each tile, write the question your buyer would ask about it. It is almost always one of two: is this good, or what do I do now.
- Answer it in one sentence in the buyer's units. Turn ratios into counts of things she recognises.
- Add a baseline from a published threshold, her own last period, or a target she set at signup. If you have none yet, say so, and show the trend once there is one.
- End each sentence with an action, or with nothing needs doing. Link the action to the screen where it happens.
- Move every tile she cannot act on this week to a second tab.
- Show the new screen to one person who matches the buyer, for ten seconds. Ask whether the product is working. If she answers using your sentences, ship it. If she answers with a shrug, find the tile she skipped.
What an unread dashboard costs a self-serve AI product
For many products the dashboard is where a trial user decides whether it is working. If she cannot tell, she does not upgrade, and she does not complain either. The signup was already paid for, in ads or in the founder's evenings on LinkedIn, and that cost lands whether she reads the tiles or not.
Activation is where this shows. On product-led SaaS benchmarks, most products activate 20 to 40 percent of signups. The same benchmarks tie a ten point lift in activation to a 15 to 25 percent rise in free-to-paid. Tile copy is one of the cheapest places to find those ten points.
For an AI product the waste is sharper. Every one of those 438 calls cost inference, so a trial that ends quietly still ran up a compute bill. AI product builders average around 52 percent gross margin, against 70 to 80 percent for traditional software. You carry the same quiet churn with less room to absorb it.
At Studio Maydit we design the first week of AI products, and for many of them the dashboard is the screen where a customer decides to stay. Our product design work rewrites tiles like these alongside the flows around them. Fixed-scope projects take three to four weeks and end with a diagnosis of what is leaking in the product. If you would rather compare studios first, this ranking of design agencies for dashboard and data UI is a fair place to start. If your users keep asking whether a number is good, book a 30-minute call and we will rewrite your first screen with you.
