Most AI features add another screen, another tab, or another place your team has to remember to check. Prompt columns do the opposite. They put the AI answer right inside the record your team already looks at every day. No plugin, no extra flow, no new habit to build. Just a column that thinks.
If you run Dynamics 365 Sales, Customer Insights, or Business Central, this is worth understanding, because it changes how you can use AI without changing how your team works. And if you are weighing where AI actually earns its keep in your Microsoft stack, this is exactly the kind of question our AI Enablement work at Reach is built around.
What a prompt column is
A prompt column is a Dataverse column that fills itself in using a natural language prompt instead of a formula or user input. You write the prompt, point it at other columns on the table, and an AI model generates the value. That value gets stored in the row like any normal field, so it is ready for forms, views, flows, and Power BI right away. Each table can hold up to five prompt columns, and they run on GPT 4.1 mini or another available model.
The shift here is subtle but real. Calling an AI model from Dataverse is not new. What was missing was a simple way to keep the answer sitting on the record, without building custom plumbing to carry it there.
A working example: scoring deal health
To see whether the feature holds up outside a demo, we built a prompt column called Deal Signal on the Opportunity table. It reads the stage, revenue, close date, probability, description, and linked notes, then scores the deal from 0 to 100 and explains why.
Two details in the prompt matter more than the scoring logic itself.
The first is a hard rule to use only the information supplied and to mark confidence as Low when a factor cannot be judged. Skip this instruction, and you get confident-sounding scores built on nothing, which is worse than having no score at all.
The second is asking for structured output. Instead of a paragraph, the prompt returns JSON, a format a screen can actually render rather than something a person has to read line by line. We bound a simple control to the column that turns that JSON into a gauge, a headline, and a short list of risks. One opportunity came back scored at 25, flagged as at risk, with no budget holder identified and no recent contact logged. The seller sees a gauge. The column still just holds text underneath.
Why structured output is the real unlock
A block of JSON on its own is invisible to the tools that make Dataverse useful. You cannot sort a view by it or chart it in Power BI. So the second piece of the pattern is a small flow that parses that JSON into proper typed columns, a whole number score, a choice field for the band, a text headline. Once those exist, a view showing every at risk deal over 100k closing this quarter takes minutes to build, and a health distribution dashboard follows right behind it. This is the part that actually changes how a sales manager runs their week.
When a prompt column runs, and when it does not
This is where a proof of concept meets production reality. A prompt column runs when a record is created, or when one of the columns named in the prompt definition changes. Nothing else sets it off.
Existing records are not backfilled automatically. There is no button to run it on demand out of the box. Editing the prompt itself does not recalculate anything already generated. And the one that catches most teams off guard: adding a note to the record does not touch a column on the opportunity, so the score does not refresh, even if three new notes just told you the deal is falling apart.
The workaround is a helper column. Add something like Signal Requested On to the record, include it in the prompt definition, and stamp it using a ribbon button. That counts as an input change, which gives you a working refresh action and a way to force a rerun after you tweak the prompt.
Making it usable outside a demo
Prompt columns run asynchronously, so they never hold up a save, but that also means the form has no built in way to show a value is on its way. Filter conditions become your cost control lever, since you can limit the prompt to open opportunities above a certain value so it does not run on every record. Every prompt column also comes with automatic Status and Details columns, tracking whether it is not started, in progress, completed, or failed, along with the error text. Put that Status column on a view before you show this to anyone.
Where to start
Three moves matter more than anything else. Pick one table where a person currently reads free text and forms a judgement call, since that is the exact shape of problem this feature fits. Design your first prompt to return structured JSON even if you only display it as plain text for now, because adding structure later means reprocessing every record you have already generated. And decide upfront how the column refreshes, because the trigger rules will shape your design far more than the wording of the prompt ever will.
The interesting part was never really the score. It is that the score now lives in the row, right where your team already looks.
Where Reach fits in
Prompt columns are a genuine step forward for Dataverse, but the distance between a working demo and a governed, production ready field is real, and it shows up fastest in the trigger rules most teams never read closely. This is the kind of gap our AI Enablement service exists to close, alongside the Project Assessments and Audits we run for teams who want an honest read on where their Dynamics 365 environment stands before adding anything new.
Whether you are scoping a fresh AI feature, recovering a project that has drifted off track through our ERP Rescue and Recovery work, or looking for ongoing support through Managed Services, Reach brings the same thing to every engagement, deep platform expertise, straight communication, and a team that treats your environment like it matters. If you are running Dynamics 365 Sales, Customer Insights, or Business Central and want to know what an AI enabled column could look like on your own tables, get in touch with Reach.
