Why is data missing, wrong or different between runs?

Last updated: October 12, 2026

This article explains why a report from Emma can come back with blank fields, a wrong figure, or different content from the same inputs, and what to change so the next run is right.

Emma reads only the documents you select when you generate, and writes what it finds in them. Most data problems trace back to the inputs: a document that was not selected, a figure that lives in no uploaded document, or a transcript error that was never corrected. The checks below start with the inputs because that is where most fixes are.

Before you begin

Have the report open, with the list of documents that were selected when it was generated.
Know which document the missing or wrong figure should have come from.
If the problem repeats across reports, keep one clear example with its inputs; the team will ask for it.

Check this first

Open the report and click the figure's citation. Every extracted data point in the snapshot and the finished report links back to the document it came from, so you can see in a moment whether Emma read the right source or none at all.
Check which documents were selected. Selected documents show green at generation. Emma does not read the client profile, earlier reports or anything outside that selection, so a document left unticked is invisible to it.
Check the figure is in a document at all. Fund charges, performance data and anything from a provider illustration must be uploaded to appear; Emma does not look them up.

The document was not selected, or the data is not in any document

This is the most common cause of a blank field or a half-empty table.

Click Regenerate on the affected section. Regenerate re-runs that one section, lets you pick which uploaded documents it reads, and takes a short instruction such as "take the ongoing charge from the illustration". You do not need to redo the whole report.
Upload the missing document if the figure was in none of them, then regenerate the section with that document selected.
Upload only the relevant pages of a long document. There is a limit of around 100 pages per document, and a 60-page illustration where three pages matter is easier for Emma to read when you upload just those three.

Please note: when a figure is missing from the inputs, Emma leaves the field blank or writes "information not provided" rather than guessing. The snapshot stage flags sections with insufficient information before the report is generated, so check it there and add the document before you go on.

A charge or figure was pulled from the wrong document

If you upload an illustration and a valuation that both show a platform charge, Emma can take the wrong one. Provider charge terminology (TER, OCF, AMC) is the weakest area of extraction.

Review every charge in the report against the source before it goes out. Where an illustration shows a standard platform charge but the client gets a lower family-linked rate, Emma pulls the illustration figure.
Remove or leave unticked any document that carries an out-of-date version of the figure. Fewer, more current documents give more accurate charges than a bulk upload of everything on file.
Ask for section-to-document mapping on your template if charges are wrong repeatedly. AdvisoryAI can set each section of a bespoke template to read specific document types, which is the main lever for stopping one document's figure overriding another's. Several firms' inconsistent runs traced back to this not being switched on.

A transcription error has carried into the report

Errors in the raw transcript, for example a misheard fund name or the wrong day, cascade into the letter if the transcript is an input.

Clean the transcript first on the past meeting, assigning each speaker to the right person.
Generate the meeting note and correct any errors in it.
Use the corrected note as the report input and untick the transcript. The corrected note is the version that should feed the report, not the recording.

Too many documents were uploaded

Over-uploading causes confusion and errors. A typical report needs the corrected meeting note, the fact find and the risk assessment, plus the documents specific to the advice such as illustrations, LOA summaries and performance data. It rarely needs more than about ten.

Untick documents that do not bear on this report rather than deleting them; they stay on the client for next time.
Regenerate the affected sections with the smaller set selected.

A joint client's second person is patchy

In a joint report the data for the second client can come out thin or fragmented even though both records are linked. This is a known limitation of how joint clients are handled rather than something you set up wrongly. The two records are linked with the joint client button, there is always a lead client, and a joint report pulls from both records with his and hers table columns and files under both.

Regenerate the affected section with an instruction that names both clients explicitly, for example "include costs and charges for both".
Report it if adding a second client changed the whole note. Adding a second client to a meeting and regenerating can change the note substantially; the team investigates these case by case, so send the meeting and both client IDs to support.

The relationship label you choose for the pair does not affect the output.

Two runs from the same inputs came out different

Generative output varies in wording from run to run. The facts should not vary: if one run picks up the charges and the next misses them, or an address is extracted on one run and not the other, that is treated as a bug.

Note the report IDs of both runs and email them to support@advisoryai.com with the figure that changed.
Ask whether the document classifier is enabled on the template. Several firms' inconsistent runs traced back to it being off.
For things like the number of options presented, ask for a template rule. AdvisoryAI can constrain the output, for example by tightening the paragraph options a section may choose from, so the report stops adding wording your template does not ask for.

Please note: each firm's templates and model are siloed. Content from another firm's template cannot appear in yours.

A calculation is wrong

Emma handles simple derived figures, such as a value multiplied by a charge, and charge comparison tables across ceding and receiving schemes. Anything beyond a couple of equations should be done outside Emma, because checking one wrong number takes longer than doing the sum yourself.

If a built-in calculation looks wrong, for example the tax treatment applied to a contribution, send support the report, the inputs and what you expected. The team will check the template logic and confirm what it should do.

Still not working

Email support@advisoryai.com with:

the report ID, or the report URL
the input documents, with the missing or wrong figure highlighted and the correct value
for a recurring miss, one complete example case, so the template can be trained on where that data lives

The AdvisoryAI team cannot see into your account, so the example is what lets them reproduce it.

Also asked

How do LOA pack summaries work, and how accurate are they?

Five standard LOA pack templates (pensions DC, pensions DB, investments, bonds, protection) are included at no extra cost; they are not visible until support enables them, so email support@advisoryai.com before assuming you need a custom build. The output is table-based by section: provider info, contributions, safeguarded rights, fund info, plan details, drawdown and crystallisation. Accuracy is typically above 90 percent across providers; policy number, name, safeguarded benefits and protected tax-free cash are highly reliable, and charge terminology is the weak spot to check first. Custom LOA versions can add firm-specific fields, missing-item checkers, field-by-field extraction instructions ("find the fund value, put it in this box" works better than a broad instruction about charges) and a repeated table per plan. A support licence gives admin staff LOA-pack-only access, so the packs can be run by the team that chases the letters.

How do I get performance and ongoing charges to pull through from the illustration, and keep unmanaged plans out of the review?

Performance and charges only appear if they are in an uploaded document, so upload the illustration and make sure it shows green at generation. If a section still misses them, click Regenerate on that section, select the illustration and give it the instruction "take the ongoing charge and performance figures from the illustration". To keep unmanaged plans out, control scope through the inputs: leave out the documents for those plans, or add a one-page note stating which plans are in scope for this review, and sections with nothing to recommend drop out. If the same section misreads the illustration on every run, email support@advisoryai.com with the template name and one example report, and the team maps that section to the illustration document type so it stops reading the wrong file.

The plan benefits table with ticks and crosses appeared once; how do I get it every time?

Tables are the weakest area of generated output and usually need one or two build iterations before they come out the same on every run. Send the raw output from the run that produced the table, plus the input documents, to support@advisoryai.com with the template name; the developers compare it with a run that did not produce it and change the template so the table is built the same way each time.

Can the charges be broken down per plan instead of one aggregated figure?

Yes. The template logic defines the charges table, and an instruction to repeat it for every plan produces a table per plan instead of one overall figure. Email support@advisoryai.com with the template name and an example of the per-plan layout you want; it is a straightforward template change.

Why did the report output get worse the last time we tried the paraplanning side?

Previously the model looked at every uploaded document for every section and had to guess which figure belonged where, so accuracy fell as firms uploaded more. Section-to-document mapping, set up by AdvisoryAI on the template, now points each section at a specific input document type, for example the illustration for charges, and that fixes most of the inaccuracy. Emma Express also adds inline citations so every figure links back to the document it came from. If your template predates this, ask support to switch mapping on.

Other AI tools make things up, so what stops Emma doing the same?

Every extracted data point in the snapshot and the finished report links back to its source document, so a checker can verify each figure inside the platform. When a figure is missing from the inputs, Emma leaves the field blank or writes "information not provided" rather than inventing one, and a per-template fail-safe can be set so it writes "not specified" with placeholders or highlighting instead. The recommendation itself has to come from an uploaded input; Emma writes the rationale but does not decide the advice. There is no confidence score on individual values, so the paraplanner check remains part of the workflow.

If performance is not on the transaction statement, can Emma work it out?

Yes. Given a valuation today and a valuation one year ago, Emma calculates the percentage, and simple derived figures such as a value multiplied by a charge are supported. Anything beyond a couple of equations should be done outside Emma, because checking one wrong number takes longer than doing the sum yourself.

Next steps

Give the report a thumbs up or down and say what was wrong. Feedback on the report, a marked-up example sent to AdvisoryAI, and the Dictionary for terminology are the routes that change future output; edits made inside the platform are not remembered.
For LOA pack summaries, treat charge terminology as the field to check first. Policy numbers, names, safeguarded benefits and protected tax-free cash are reliably extracted.

Ready to fix the section? Open AdvisoryAI, open the report and click Regenerate on the section that is wrong.