Fields and Values Drive Search Queries and Agentic AI
When people talk about getting better answers from AI, they talk about prompts: a clearer question, more context, the format they want. That matters. It is also only one pillar.
The other pillar is older than chatbots and older than Google: fields and values. Search has always been a matching problem. You type something. A system looks through labeled pieces of information and brings back the ones that line up. The labels are fields. What sits inside them are values. Together, they are the records of the parts of a product, a page, a spreadsheet, a ticket, a report.
Prompts decide how you ask and what you ask. Fields and values decide what can even be found. An agent can follow instructions perfectly and still come back empty-handed if the software, webpage, or spreadsheet has no field that holds the answer. If companies don't keep their knowledge and data stored in an online shared platform, then an agent can't even access it. It's not enough to have someone who "knows the stuff" in their head.
How Search Learned to Look Things Up
Before anyone typed a query into a box, people already knew this pattern. A library card catalog never stored the whole book. It stored parts under labels: title, author, subject, year. If there was no "subject" card, you could find a book by title — and you could not find it by topic.
The early web worked the same way. In the mid-1990s, Yahoo was a directory: websites filed into categories, like a phone book with headings. Those categories were fields. Then crawlers — AltaVista, Lycos, later Google — indexed the internet. They did not save each page as one unlabeled blob. They read the structure already on the page:
- the title in the browser tab
- the headings on the page
- the body — the actual words
- the URL
- the meta description and keywords the author declared
- later, the date, the author, the language, the links pointing in
Those were fields. "Cheap flights to Chicago" sitting in a title tag was a value. "Cheap flights to Chicago" buried in paragraph twenty was a different value in a different field — and early search engines treated them differently, because a match in the title was a stronger clue that the page was about that thing.
Those were fields. "Cheap flights to Chicago" in a title was a value — and a stronger clue than the same words buried in paragraph twenty.
When you typed keywords into Yahoo, the engine was not reading your mind. It matched what you typed against those labeled pieces so the results were relevant to what you searched. Advanced search made it obvious: intitle:, site:, filetype: were just you pointing at a specific field. No field in the index, no way to ask that question.
Search engines were built to index what was on the internet, using the fields and values on webpages, so that the keywords you typed came back as something relevant. Agentic AI did not invent this. It inherited it — and pointed it at your software, spreadsheets, and knowledge bases.
What a Field and a Value Actually Are
Strip the jargon out and the idea is almost too simple.
A field is a labeled slot. Name. Price. Date. Status. Author. Category. Quantity on hand. Last updated.
A value is what is recorded in that slot, for one specific thing. "Acme Corp." "$49." "March 12." "Open." "Jordan Lee." "Billing." "14." "January 2024."
A record is one whole thing made of those parts: one product, one customer, one ticket, one article, one row in a spreadsheet, one event on a calendar. The product is not a blob called "the product." The product is the set of its fields. The report is not a mysterious document the AI dreams up. The report is a gathering of values from records, arranged so a human can see a pattern.
That is what "fields and values are records of parts of the product or report" means. A shoe in an online store is not just a photo. It is a name, a brand, a size, a color, a price, a stock count, a SKU, a category. A weekly sales report is not a vibe about how the week went. It is units sold, revenue, region, product line, pulled from those same slots and totaled.
Software is built this way. So are websites, spreadsheets, CRMs, help desks, knowledge bases, calendars, and forms. The screen you look at is a layout. Underneath the layout is a set of named boxes with contents. Humans fill the boxes. Databases store the boxes. Search engines and agents query the boxes.
An agent looking at your world is doing what Yahoo was doing to webpages, only across more kinds of objects:
- a column header in a spreadsheet is a field; each cell under it is a value
- a form input on a website is a field; whatever a user typed is the value
- a property on a customer record in Salesforce or HubSpot is a field
- a column on a Jira or Zendesk ticket is a field
- a key in a JSON API response is a field
- a metadata tag on a knowledge-base article — audience, product, last reviewed — is a field
When you ask an agent to find, gather, and make something — a report, a slide deck, a list of improvements, a briefing for leadership — you are asking it to query those records, pull the values, and assemble them. The prompt is the shopping list. The fields are the aisles. If the aisle does not exist, the item cannot be picked.
The Hard Reality
If a software product, a webpage, a spreadsheet, or a report does not have a certain field — or the field exists but has no value — an AI agent cannot query and pull that type of information. If employees don't record what they know in an online shared platform, then not only can others not access that information, an agent can't either.
It can guess based on what it sees in the HTML of the page in your browser. It can look at general patterns and give an answer that sounds confident. What it cannot do is go get the fact, because the fact was never recorded in the UI (user interface), the source code (the company's code that makes the product), or the database (where all user and setting information and attachments are store).
People blame the prompt when this happens. "The AI is bad at reports." Sometimes the prompt is too vague. Often the agent did exactly what you asked, against a system that does not hold the part you wanted. Sometimes team members don't "fill in and record" everything they should so customer, client, ordering etc records don't all have the info that they each should. Sometimes, the info isn't recorded at all, there are no logs set up in the database to record those types of customer actions, feature uses, or type integration data specific to each integration.
Prompts without fields and values filled in lead to wishful prompting.
Three Examples Everybody Already Understands
1. Your email inbox
Every message in your inbox is a record. The fields are so familiar you stop seeing them: From, To, Subject, Date, Unread, Has attachment, sometimes Folder or Label. The values are "Amazon," "Your order has shipped," "October 3," "unread," "yes."
When you type from:amazon after:2026/10/01 receipt into Gmail, you are not asking a poet to interpret your feelings about shopping. You are querying fields. From must equal something like Amazon. Date must fall after October 1. Subject or body should involve a receipt. That is the same move as a 1990s web search, pointed at your mail.
Hand it to an agent: "Pull every Amazon receipt last month and make a spending summary." If From and Date exist, it can. If the inbox were an unlabeled pile of text, it could not run "last month" or "from Amazon." Ask for starred mail in October, and if there is no Starred field, there is no such list. The prompt was fine. The field was never there.
2. The spreadsheet of money
Column headers are fields: Date, Vendor, Amount, Category. Each row is a record. Each cell is a value.
"How much did we spend on software versus travel last quarter?" If Category is filled in, the agent can total and write the report. If Amount exists but Category does not, it can give a grand total and nothing you actually asked for. Guessing that Uber is travel is not a query.
Same for a sales sheet. "Which products sold best in the Midwest?" needs a Region column. No Region, no regional ranking — no matter how carefully you prompt. A spreadsheet with one column called Notes is a diary. Named columns are a database an agent can work.
3. The integration report a product team actually wants
Picture a SaaS company that has spent years bolting on integrations — Slack, Salesforce, Google Calendar, Zoom, QuickBooks, a dozen more. Leadership asks an agent a question that sounds like one request and is really a pile of queries:
Give me a detailed report of which customers are using which integration, which subscription tier they are on, when they started the integration, whether they are still using it, and which integrations the support team files development tickets against for integration bugs. I want to see which integrations are popular and cheap to maintain — those are worth keeping — and which sit on the other side of the quadrant: least popular, most issues, maybe not worth keeping.
That is not a summary. It is a query across product, billing, and the help desk: popularity on one axis, cost to maintain on the other. Every piece of that request has to exist as a field with a value.
| What the report needs | The field it requires |
|---|---|
| Which customers use which integration | Customer + Integration name |
| Subscription tier | Tier on the account |
| When they started | Start date / first connected |
| Whether they are still using it | Status or last-used date |
| Which integrations generate bug tickets | Integration + Type on the ticket |
If you don't track all the data points you need to regarding integrations, then your agent can't make a detail report about what is happening for the customers that use those integrations.
- Not tracking all customer communication in a support desk platform, leads to not being able to have an agent access how many times that customers has reach out about integrations and understand the tone/satisfaction using the integration over time.
- Not tracking the same data points for each integration, leads an agent to create a report that screws what's really going on and who uses which integration.
- Not tracking which tickets in the support desk are integration related means an agent can't realistically tell you who is using what and how often they reach about issues. It also means that you can't tell which integrations have the most issues.
The prompt can be perfect. If those fields are not on the software or in the right platform that you use at work, the agent cannot generate that report. It can write a memo about how a team might think about integrations. It cannot show your customers, your tickets, your keep-versus-cut list. Every column the requester wants has to already exist as a field with a value. No field, no slice of the answer.
What This Means When You Ask an Agent to Do Real Work
Once you see the pattern, a lot of frustrating AI moments stop looking like model failures. If you see what you think is an agent failure, go and validate the data and see if all the right fields and values are actually there where they should be. (Rule of thumb, you should always be validating what an agent outputs anyways)
"Make me a Q3 presentation of our pipeline by region" — needs region, amount, stage, and close date on the deals.
"Find customers who love the product but haven't upgraded" — needs some field that records sentiment or NPS, and a field that records plan tier. A notes box full of unstructured chatter is not the same thing.
"Suggest improvements to our help center" — needs, at minimum, which articles exist, what they cover, how old they are, and some signal of what people are actually asking. Tickets, search queries, and article metadata are those signals. If none of that is filed, the agent is guessing at a library it cannot catalog.
"Build a weekly digest of what shipped" — needs a field that means shipped: a release date, a status of Done, a version tag. A pile of Slack messages is not a shipped field.
So the practical habit is not "write longer prompts." It is: before you ask, look at the record. Or after the agent hands you the report and you think, "that can't be right based on what I know", go see what the agent pulled from and find out what might be missing. What are the fields? Which of them are filled in? Which question on your shopping list has no aisle?
If the field is missing, you have two honest options. Add the field and start recording the value — then the agent can query it forever after. Or change the question to one the current fields can actually answer. What you cannot do is prompt your way around a hole in the filing system. That hole is the whole story of why early search engines invested so much in titles, headings, and keywords, and why every serious product eventually becomes a debate about which properties to put on the object.
Good prompts still matter. They tell the agent which fields to use, how to combine them, what to ignore, and what the output should look like. A vague prompt against a rich set of fields wastes a well-labeled system. A precise prompt against empty fields is a well-written letter to an empty warehouse.
The work that makes agents look like magic is usually done years earlier, by people who named the parts.
Bottom Line
Search engines were never sources of truth or perfect in retruning the search results that you wanted. They were indexes. Yahoo and the crawlers that followed treated webpages as labeled pieces—titles, headings, keywords, dates, descriptions—so the words you typed had something relevant to land on. Fields were the drawers. Values were what was in them. Relevance was the match.
Agentic AI is the same idea, pointed at the rest of your work. It relies on the fields and values in software, websites, spreadsheets, CRMs, tickets, and knowledge bases to find what you want and turn it into reports, presentations, briefings, and suggestions. Those fields and values are the records of the product and the report. No parts on file, no report.
The limit is blunt. If the field does not exist, or the value was never entered, the agent cannot pull that information. It can still talk. Talking is not querying. Querying is how you get something true.
Two pillars only work as a pair. One is the prompt: ask clearly for what you want. The other is the filing system underneath: named fields filled with real values. The first is having a good conversation. The second is having a labeled world.
Build the labeled world—or at least notice when you haven’t—and the agent can find it, gather it, and make something you can use. Leave the world unlabeled, and you will keep buying better prompts for a question the records cannot answer.