Giving Research Directors Their Quarter Back, Without Taking Away Their Judgment
How Kodexo Labs helps a global agricultural research nonprofit turn scattered email, chats and meeting notes into a quarterly progress update that directors check, correct and submit. The AI drafts. The director decides.
6 Aug Teams chat over the 12 Jun donor report
One agreement covering two of five diseases.
Open decision: re-baselining, raised 19 Aug, not decided.
95–98%
Fields kept without a factual change
99%
Citations verified
~90%
Traffic light matches the director
20–40
Relevant items per objective
The client & the challenge
Every Quarter, Directors Report From Memory
Our client is a nonprofit research institute working across Africa and Asia on food security, livelihoods and climate resilience. It wanted ways of working its own teams adopt, run and keep improving after the partner leaves, not clever tools that nobody uses.
The real risk is not slowness. It is a report that sounds right but is quietly out of date.
Every quarter, directors report on the institutional objectives they own. Each update is a traffic light, a progress percentage and three short lines. Those updates feed the dashboard leadership uses to steer the institute.
The facts a director needs are already in their Microsoft 365 account: an email from a country lead, a Teams chat, a cancelled calendar invite, a meeting note in SharePoint. Pulling them together takes hours, so the quality of an update depends on how much time the director had that week.
A donor report written in June can still say "on track" in September, after the facts underneath it have changed.
Before & after
Same Evidence. A Very Different Update
Drag to compare how a quarterly update gets written today, and how it gets written with the workflow.
Today, the quality of an update depends on how much time the director had that week. With the workflow, every claim is one click from its source.
Illustrative scenario with invented people and figures. No client data is shown.
Reports written from memory
Every field drafted from the director's own evidence
Hours pulling facts together
A draft checked in minutes, every claim linked to its source
Older reports quietly repeated
Newer evidence flagged as overriding older
A tool that decides gets ignored, or trusted blindly
The AI proposes, the director decides
Why this is harder than it looks
A Summary Tool Would Not Solve This
Three problems sit on top of each other, and all three have to be handled at once.
Evidence is messy, directors are accountable for their own numbers, and updates must fit a fixed format leadership relies on.
Technical
Evidence is messy. Some is out of date, some conflicts, and nobody writes down what the traffic light should be.
A workflow that spots when newer evidence overrides older evidence and shows the director both.
Human
Directors are accountable for their own numbers. A tool that decides for them gets ignored, or trusted blindly.
A draft they can check in minutes, every claim linked to its source, nothing submitted without them.
Institutional
Updates feed leadership through a fixed format with closed vocabularies that must stay consistent.
Outputs that fit the format exactly, every time, with no silent changes to the data.
Our approach
Built for the People Who Have to Use It
We treated this as an adoption problem first and a technology problem second. Four principles shaped every decision.
The client wanted ways of working its own teams adopt, run and keep improving after the partner leaves, not clever tools that nobody uses.
The AI proposes, the director decides.
The workflow drafts every field, but it never submits and never writes to the system of record. The director sees the full update before anything is saved.
Every claim shows its evidence.
Each sentence links back to the email, chat or meeting note it came from, so checking takes minutes, not a second search.
Conflicts are surfaced, not hidden.
When an older report says "on track" and a newer meeting says otherwise, the director sees both, with dates, and the newer source is flagged.
Build the smallest thing people will actually use.
One workflow, one director, one quarter they can test against. Grow only when the first user is opening the draft and sending it.
We delivered it end to end.
Our team owned the whole build: the Microsoft 365 connections, the evidence assembly, the drafting and review experience, and the rules for learning from corrections, all inside the client's approved environment and security rules. One team, one point of accountability, and every part documented so the client's own staff can maintain it.
The solution
A Draft They Check, Correct and Submit
We built an AI-assisted workflow that reads the director's own Microsoft 365 material and prepares their quarterly update as a draft they check, correct and submit.
Built to last. It reruns on fresh evidence without anyone hand-editing the output, runs inside the client's approved Microsoft 365 environment, and ships with a plain runbook their staff can maintain.
What the director sees
In the exact format the reporting system expects: traffic light, progress figure, and three short narrative lines within their character limits.
Click a sentence and see the email or meeting note it came from, with its date.
When evidence disagrees, both versions show, and the newer one is marked as overriding.
Not answered. If a judgment call belongs to the director, the draft says so.
Including the suggested traffic light.
How it works
From Scattered Evidence to a Submitted Update
We built a workflow that reads the director's own Microsoft 365 material and prepares their quarterly update as a draft they check, correct and submit.
The director stays in the highlighted step. Everything before it saves them hours; nothing after it happens without them.
Read
Email, chats, calendar and notes.
Assemble
The facts behind each objective.
Flag conflicts
Newer evidence overrides older.
Draft
Every field, every claim sourced.
Check & submit
The director makes the call.
Saved
Only after the director approves.
Proof in action
The Update That Would Have Gone Out Wrong
A program director owns an objective to sign disease-surveillance data-sharing agreements with three national veterinary services, each covering five priority diseases. Last quarter she reported it Green, 45%. This quarter her evidence tells a different story, but no single document says so.
Illustrative scenario with invented people and figures. No client data is shown.
Year will close with one agreement covering two diseases. Re-baselining raised, not decided.
The signed agreement covers two of five diseases. Three were struck at signing.
Third-country scoping mission cancelled. Budget reallocated.
Second country stalled. Nobody assigned to the missing annex.
"On track" for three agreements covering five diseases.
First agreement cleared, covering all five diseases.
From the June donor report.
Quietly out of dateIt caught the change.
Newer evidence flagged as overriding older.
It suggested a status, with reasons.
The director can switch it with one click. The AI's suggestion is never the final word.
It left the hard call to her.
Re-baselining was flagged as an open decision for the director, not answered by the AI.
The result is a draft that tells leadership the truth, in the director's own hands, with every claim one click from its source.
Built for adoption
Built for Adoption, Not Just Delivery
A tool nobody opens is a failed project, however good the code. So the rollout is designed around how a busy director builds trust.
Trust is earned on a quarter they can check, not one they have to defend.
Start where the answer is already known.
The first director runs the workflow on a completed quarter and compares the draft with what they actually submitted. Trust is earned on a quarter they can check, not one they have to defend.
Learn from corrections, carefully.
With the director's consent, the workflow learns their style: wording, what they cut, which sources they trust. It never learns the traffic light, the percentage or a new fact on its own. Those stay human, and factual corrections go to a log a person reviews.
Measure use, not deliverables.
| What we track | Why it matters |
|---|---|
| Time from draft shown to submit | If it stays long, the draft is not saving work |
| How much the director edits | Heavy edits mean the draft is not trusted yet, and show which fields need work |
| Source thumbs up or down | Shows which evidence directors believe |
| Zero-edit submissions | A warning, not a win: either excellent or rubber-stamped, so early uses ask "did you review this?" |
Hand over as we go.
Every workflow ships with a plain runbook, a recorded walkthrough and a named owner on the client side who runs it while we watch, before sign-off.
Results
Measured Against a Completed Quarter
Measured in a back-test: the draft compared field by field with the update the director actually submitted, plus a director-reviewed sample of citations.
Time-saved and usage figures will be added once directors are using it in live quarters.
Citations verified against the source document
Checked in director review on a sample of citations.
Relevant items per objective per quarter
Traffic light matches the director
The suggestion is never the final word. The director switches it with one click.
Fields kept without a factual change
What this means for your organization
The Knowledge Exists. It Is Just Spread Out
Most research institutes, NGOs and international development organizations share this problem. The knowledge is spread across inboxes, chats and SharePoint folders in a dozen country offices. Reporting runs on memory, and a donor report can drift from the truth without anyone noticing.
The same pattern carries over to other work:
Donor and board reporting
Drafted from evidence with every claim cited.
Knowledge retrieval
Across past projects, so staff stop asking around for that one report.
Project status updates
That flag stale or conflicting information before it reaches leadership.
Proposal drafting
That reuses what the organization has already proven.
The same rules apply every time: your staff stay in control, nothing is written to your systems without a person approving it, and the capability stays with your team when we leave.
Running a Distributed Nonprofit or Research Program on Microsoft 365?
Let's find the one workflow your team would actually use first.
Related work
