Agent

Deliverable fact-check

Deliverable fact-check is the pass a manager never has time to do properly before a steering committee. The agent takes the draft, follows each claim back to the source it cites, and flags anything unsupported, out of date or inconsistent with the rest of the pack. It produces a flag list, not a verdict; the manager decides what to fix.

How it runs

Step by step

01

Read the draft and its sources

The deliverable and the supporting files are read together from the engagement folder.

02

Trace each claim to source

Every factual statement is matched to the document that supports it, and the match is shown rather than asserted.

03

Flag the unsupported statements

Claims with no traceable source are listed first, because those are the ones that cause trouble in the room.

04

Check dates and versions

Statements resting on superseded data or an old version of a client document are flagged with the newer file alongside.

05

Compare across the pack

The same figure or statement appearing differently in two documents is surfaced as an inconsistency to resolve.

06

Manager reviews the flag list

The list goes to the engagement manager, who decides what to change. Nothing is posted to the team channel until that review is done.

The harness

Exactly what this agent can see, touch and change

The same five controls sit behind every ollo agent. These are this one's settings — visible before you build it, not buried in an admin screen afterwards.

Context

What reaches the model

Managers doing a final pass before a steering committee, and quality reviewers who need to know where to look rather than reading every page again.

Escalation

When it asks a person

The engagement manager reviews the flag list before it is posted to the team channel

Applications & Rights

Which tools it uses, and what it may do in each
Google DriveRead only
SharePointRead only
ConfluenceRead only
SlackWrites newCreates new records or documents. Never edits, moves or deletes anything that was already there.

Verification

How you know it’s right

Every claim links to the document it came from. A statement the agent cannot cite does not make it into the output — which is what makes the result reviewable in minutes rather than re-read end to end.

Who it’s for

Managers doing a final pass before a steering committee, and quality reviewers who need to know where to look rather than reading every page again.

What you’ll need

  • Draft and sources in one folder
  • Source-of-truth files identified
  • House citation convention agreed
  • Manager named as reviewer

What you get

  • A flag list of unsupported claims
  • A list of version and date conflicts
  • A cross-document inconsistency report

What it doesn’t do

It checks whether a claim matches its source. It cannot tell you whether the source was any good, and it does not sign the deliverable off or replace the review the manager owes the client.

How you get it

We build the first one with you

Not a template you configure alone. We sit with your team, build it on real data, and hand over the controls.

01

Scope

One session with the people who actually do the work. We agree what the agent reads, what it may write, and who approves.

02

Co-build

Built on your own data, not a sandbox. You watch it being made, so you know why it behaves the way it does.

03

Handover

You own the controls. Change the context, tighten the rights, move the approval gate — without coming back to us.

What it replaces

Parts of the work currently spread across the categories below. It does not replace any of those products outright.

Line-by-line manual checkingLate-night version reconciliationAd-hoc quality passes
Estimated savingNot setNo figure ships until someone at ollo owns it and the method behind it.

Frequently asked

Does it change the deliverable?

No. It produces a list of flags against a read-only copy of the draft. Editing stays with the team, because a fact-check that silently rewrites a document is worse than no fact-check at all.

What counts as unsupported?

Any statement the agent cannot trace to a connected source, including one where the cited file exists but does not actually say it. Both cases appear in the list, labelled differently, since they need different fixes.

Will it flag things that are fine?

Sometimes, and that is the intended bias. It is tuned to over-flag rather than to miss, on the basis that a manager can dismiss a flag in a moment but cannot recover from a wrong number in front of a client.

Can it check a slide deck?

Yes. It reads decks alongside documents and flags per slide, which matters because most claims in consulting reach the client as a slide rather than as prose.

Let's Build

AI is a capability you build. Let's build it together.

30 minutes with our team and you'll leave with a real plan — not a sales pitch.